---
okf_version: "0.1"
---

# Radar — Rota Nacional

* [Autonomous science limited by physical validation, according to new arXiv paper](https://ai.exponencialadm.net/okf/en/radar/ciencia-autonoma-limitada-pela-validacao-fisica-b1d23a.md) - A recent arXiv paper argues that autonomous science depends on physical validation capacity and proposes markets and institutions to allocate that scarce work.
* [Intel GPU D-Series: inference card with 480GB LPDDR5X already sampling](https://ai.exponencialadm.net/okf/en/radar/intel-gpu-d-series-placa-de-504953.md) - Intel presents the GPU D-Series (formerly Crescent Island), with up to 480GB of LPDDR5X at 1.5TB/s, already sampling. Radar briefing with the specifications stated in the source.
* [μLattice: memory-scaling alternative to attention, according to its author](https://ai.exponencialadm.net/okf/en/radar/lattice-alternativa-de-escalonamento-de-memoria-d52015.md) - Open-source project and paper describing μLattice, a memory design that grows with input without quadratic cost. The author claims fewer operations than attention and perfect recall of thousands of associations.
* [QwenGyre: elastic RL framework for training xLong-horizon agents](https://ai.exponencialadm.net/okf/en/radar/qwengyre-framework-de-rl-elastico-para-04ec16.md) - Radar: the QwenGyre framework, described by the Qwen team at Alibaba, reallocates GPUs between rollout and training and groups branched agent runs into trajectory trees, with a reported speedup of up to 1.85x over Colocate and about 44% less training time.
* [Evaluate agents with less context and explicit criteria](https://ai.exponencialadm.net/okf/en/radar/avalie-o-agente-com-menos-contexto-cd0404.md) - A more elaborate workflow does not always improve outcomes. Compare components by task quality, consumption and data exposure.
* [HERMES: component-level harness with LLMs for software engineering tasks](https://ai.exponencialadm.net/okf/en/radar/hermes-harness-em-nivel-de-componente-ed1773.md) - An arXiv paper proposes HERMES, a code-agent harness in which each repository component has a resident LLM. The authors report gains in migration and on four software engineering benchmarks against comparable baseline harnesses.
* [HeyGen Voice tops the Artificial Analysis Controlled Voice TTS Arena ranking](https://ai.exponencialadm.net/okf/en/radar/heygen-voice-lidera-ranking-de-sintese-15c25d.md) - According to Artificial Analysis, a commercial voice service ranked first in the Controlled Voice TTS Arena, with an Elo of 1,201 across eight cloned voices in US and UK English.
* [Open-source Qwen3-TTS 1.7B CustomVoice voice cloning model](https://ai.exponencialadm.net/okf/en/radar/modelo-open-source-de-clonagem-de-3d31d6.md) - A post presents an open-source 1.7B text-to-speech model for voice cloning, with controls for emotion, age, tone, accent and multilingual support. Its realism and latency claims remain unverified.
* [Whistle: 16.9 MB on-device speech-to-text model on a public model platform](https://ai.exponencialadm.net/okf/en/radar/whistle-modelo-de-speech-to-text-869c9e.md) - Cactus-Compute published Whistle, a 16.9 MB speech-to-text model for on-device use. According to the author, it rivals Whisper base for that kind of use.
* [DeLM: decentralized code agents with shared context](https://ai.exponencialadm.net/okf/en/radar/delm-agentes-de-codigo-descentralizados-com-e1bdac.md) - Stanford article on DeLM, which replaces a central coordinator with shared context and a task queue for long-horizon code agents. According to the author, it is up to 2.49× faster and more accurate than the cited baselines.
* [LayerRoPE: normalization weights with encoded depth for Transformers](https://ai.exponencialadm.net/okf/en/radar/layerrope-pesos-de-normalizacao-com-profundidade-28aed1.md) - A research item proposes LayerRoPE, which encodes depth in normalization weights to handle growth of hidden-state norm across layers in very deep Transformers.
* [LittleBit, Samsung's method to compress 13B LLMs below 1 GB](https://ai.exponencialadm.net/okf/en/radar/littlebit-metodo-da-samsung-para-comprimir-e16e43.md) - According to the report, Samsung open-sourced LittleBit, a method that uses latent factorization to shrink weights of 13 billion parameter models to under 1 GB, with sub-1-bit levels and XOR in place of multiplication.
* [RoboJEPA: Scaling JEPA World Models for Robotics](https://ai.exponencialadm.net/okf/en/radar/robojepa-escalando-modelos-de-mundo-jepa-e641b8.md) - Meta presents RoboJEPA, a JEPA-based latent world model scaled from 22M to 8B parameters and trained on data from 12 robotic embodiments. The paper studies whether latent prediction error predicts later planning performance.
* [RoboJEPA: scaling laws for robotic world models, second post](https://ai.exponencialadm.net/okf/en/radar/robojepa-leis-de-escala-para-modelos-61969e.md) - A social media post describes RoboJEPA, an 8B-parameter JEPA world model trained on 15,022 hours of robotic video, and reports scaling-law adjustments that predict results for its 4B and 8B variants.
* [Multilingual speech-to-text benchmark for transcribe.cpp models on a laptop](https://ai.exponencialadm.net/okf/en/radar/benchmark-multilingue-de-speech-to-text-344d73.md) - Site comparing accuracy and speed of speech-to-text models supported by transcribe.cpp, measured per language on FLEURS, with data from tests on a 4750U laptop.
* [Voice clones of non-native English speakers rated more human](https://ai.exponencialadm.net/okf/en/radar/clones-de-voz-de-falantes-nao-a612f7.md) - A NeurIPS 2026 paper cloned the voices of 86 non-native English speakers. In 4,000 human ratings, the clones were judged warmer, more authoritative and more human.
* [H-JEPA: hierarchical JEPA world models for long-horizon visual planning](https://ai.exponencialadm.net/okf/en/radar/h-jepa-modelos-de-mundo-hierarquicos-b23a6e.md) - Radar: a paper led by Yann LeCun proposes H-JEPA, which stacks JEPA world models across multiple time scales for long-horizon visual planning.
* [LiFT: Loop Flow Transformers for image generation with flow matching](https://ai.exponencialadm.net/okf/en/radar/lift-transformers-de-loop-para-geracao-b3010f.md) - Radar: LiFT trains an image generator with flow matching and a shared transformer core, where each loop progressively refines the initial velocity estimate. According to the author, a model trained with 2 loops improves with up to 16 loops at inference.
* [Single-voice TTS model distilled from Kokoro-82M at 9MB](https://ai.exponencialadm.net/okf/en/radar/modelo-tts-de-voz-unica-destilado-f087a0.md) - A post reports a single-voice, single-language text-to-speech model of about 9MB of weights, distilled from Kokoro-82M, which has 8M parameters. The author claims it runs anywhere and is very fast, without independent verification.
* [Memory-Efficient Expert Routing for Distributed MoE Training](https://ai.exponencialadm.net/okf/en/radar/roteamento-de-experts-com-eficiencia-de-6b779e.md) - An arXiv paper, dated October 7, 2026, proposing methods to reduce peak memory in distributed Mixture-of-Experts training, focusing on the all-to-all dispatcher.
* [On-policy multi-teacher distillation to combine specialist models](https://ai.exponencialadm.net/okf/en/radar/destilacao-on-policy-com-multiplos-professores-a2c112.md) - Radar: the post describes MOPD, a method in which a student model learns from its own rollouts and receives token-level supervision from domain-specialist teachers, aiming to combine the abilities of separately trained specialists.
* [EmbeddingGemma 2 and multimodal retrieval in one shared vector space](https://ai.exponencialadm.net/okf/en/radar/embeddinggemma-2-e-retrieval-multimodal-em-29008d.md) - Radar: a post claims EmbeddingGemma 2, from Google, maps text, code, images, video and audio into a single vector space and uses an Apache 2.0 license. The source gives no metrics or technical documentation.
* [JEPA-Anything applies predictive learning to seven domains](https://ai.exponencialadm.net/okf/en/radar/jepa-anything-aplica-aprendizado-preditivo-a-7a3e0b.md) - Radar: the post presents JEPA-Anything, a domain-agnostic framework for world modeling based on orthogonal predictive factorization, which explores predictive representation learning across seven domains, from biology to climate.
* [LiFT: shared DiT core for iterative image inference](https://ai.exponencialadm.net/okf/en/radar/lift-nucleo-dit-compartilhado-para-inferencia-e5484d.md) - Radar item: LiFT repeatedly applies a shared Diffusion Transformer core and, according to the publication, outperforms a dense DiT on ImageNet 256×256 with about 60% fewer parameters.
* [Quantizing 10M EmbeddingGemma vectors to 1 bit with MRL cuts RAM](https://ai.exponencialadm.net/okf/en/radar/quantizacao-de-10m-de-vetores-embeddinggemma-76a60a.md) - Radar: a report that 10 million EmbeddingGemma 2 vectors, quantized to 1 bit and reduced to 256 dimensions with MRL, dropped from about 30 GB to 0.4 GB of RAM, with a declared quality loss of about 5%.
* [Paper presents Native Hybrid Attention (NHA) for long-context attention](https://ai.exponencialadm.net/okf/en/radar/artigo-apresenta-native-hybrid-attention-nha-8a32d4.md) - Briefing on a paper presenting Native Hybrid Attention, an approach that seeks to balance the speed of Transformers in long contexts with the retention of precise details from linear attention.
* [Astral GPU Indexes add support for PyTorch 2.13–2.14 and Python 3.15](https://ai.exponencialadm.net/okf/en/radar/astral-gpu-indexes-adicionam-suporte-ao-9d4dfe.md) - Radar notes that Astral GPU Indexes now support PyTorch 2.13 and 2.14 and Python 3.15, with precompiled GPU-enabled wheels for FlashAttention, FlashAttention 3, causal-conv1d and DeepSpeed.
* [How attention sinks and residual sinks rescale transformer components](https://ai.exponencialadm.net/okf/en/radar/como-attention-sinks-e-residual-sinks-4a2fc6.md) - Summary of a paper that investigates attention sinks and residual sinks in large language models and proposes that these outliers, along with softmax attention and RMSNorm, rescale other components of the architecture.
* [Dust: training transformers without gradients through activation perturbations](https://ai.exponencialadm.net/okf/en/radar/dust-treino-de-transformers-sem-gradientes-2555c8.md) - Radar: the publication describes Dust, a zeroth-order optimization method that uses parallel perturbations of activations as a virtual population to pretrain transformers, with scale experiments and comparisons with backpropagation and evolution strategies.
* [Shared Memory Reduces Memory Use in Recurrent Transformers](https://ai.exponencialadm.net/okf/en/radar/memoria-compartilhada-reduz-o-uso-de-4b1438.md) - Radar: an article on sharing memory across recursions in recurrent Transformers, reporting a 76–79% reduction in context memory for models from 150M to 1B parameters, compared with standard Transformers.
* [Pre-training a 30B-A3B MoE model scaled from 16 to 512 B200 GPUs](https://ai.exponencialadm.net/okf/en/radar/pre-treinamento-de-um-moe-30b-00eb13.md) - Report of scaling with nearly linear scalability and 35.3% MFU, using small-scale tests to narrow the search space before adding more GPUs.
* [Switching Linear Attention alternates between multiple linear maps](https://ai.exponencialadm.net/okf/en/radar/switching-linear-attention-alterna-entre-multiplos-814659.md) - Radar: the SwiLA paper, from COLM 2026, proposes switching between multiple linear maps and discusses the trade-off between attention expressiveness and KV cache memory cost.
* [Symmetric memory for NCCL communication](https://ai.exponencialadm.net/okf/en/radar/symmetric-memory-para-comunicacao-nccl-em-8c9d61.md) - Briefing on a section of the Machine Learning Engineering Open Book that describes symmetric memory in NCCL and PyTorch for small and medium payloads and for overlapping communication and computation.
* [Google releases 4B Gemma model for speaker diarization](https://ai.exponencialadm.net/okf/en/radar/google-lanca-modelo-gemma-de-4b-580379.md) - Radar: according to a note dated October 4, 2026, Google released a 4B Gemma model, fine-tuned to correct speaker labels and reduce word-level diarization errors on four benchmarks.
* [BF16 gradients in FlashAttention-3 may grow late in training](https://ai.exponencialadm.net/okf/en/radar/gradientes-bf16-no-flashattention-3-podem-25485c.md) - A reported 450-million-parameter transformer trained with BF16 FlashAttention-3 saw its gradient norm rise 1,000× after 25 billion tokens, ending with higher loss than FP32 attention and no NaNs.
* [Projection sampling for rewriting SFT training data](https://ai.exponencialadm.net/okf/en/radar/projection-sampling-para-reescrever-dados-de-f75426.md) - Paper proposes rewriting expert demonstrations as trajectories that are correct for the task and likely under the target model, before using them in standard SFT.
* [RL setup for calibrated probabilities in decisions](https://ai.exponencialadm.net/okf/en/radar/setup-de-rl-para-probabilidades-calibradas-a1d0dd.md) - Radar briefing on a reinforcement learning setup for decision models that return calibrated probabilities instead of generated text. Publication date: October 4, 2026.
* [VeriHarness verifies agent rollouts in long-horizon tasks](https://ai.exponencialadm.net/okf/en/radar/veriharness-verifica-rollouts-de-agentes-em-90a604.md) - The paper presents VeriHarness, which checks agent claims in long tasks using a fixed base model, without reference answers or rubrics during testing.
* [Distillation study: per-token KL direction matters more than rollout policy](https://ai.exponencialadm.net/okf/en/radar/estudo-sobre-distilacao-a-direcao-do-3a2a09.md) - A systematic study reports that the direction of per-token KL matters more than the rollout policy in on-policy versus off-policy distillation, questioning the idea that on-policy rollouts are inherently better.
* [GOAT adds trainable priors to attention](https://ai.exponencialadm.net/okf/en/radar/goat-adiciona-priors-treinaveis-a-attention-4eff20.md) - Radar: a paper that reads standard attention as Entropic Optimal Transport with an implicit uniform prior and proposes GOAT, with a continuous trainable prior that stays compatible with optimized kernels such as FlashAttention.
* [KohakuFA targets backward-pass precision problems](https://ai.exponencialadm.net/okf/en/radar/kohakufa-mira-problemas-de-precisao-no-c162eb.md) - The publication claims cuDNN, FA4 and FlexAttention have precision problems in backward passes and presents KohakuFA as an alternative kernel, with a public GitHub repository for inspection.
* [Looped Diffusion Transformer reuses blocks during denoising for image generation](https://ai.exponencialadm.net/okf/en/radar/looped-diffusion-transformer-reutiliza-blocos-durante-017aac.md) - Radar briefing dated 3 October 2026 on a text-to-image model that reuses Transformer blocks at each denoising step, with the performance claim stated by the original post.
* [LoopCD: training-free contrastive decoding for Looped Transformers](https://ai.exponencialadm.net/okf/en/radar/loopcd-contrastive-decoding-sem-treinamento-para-d54367.md) - Radar: the paper presents LoopCD, a training-free contrastive decoding method for recurrent Transformers that uses representations from earlier loops as weaker predictions to select tokens.
* [LoopCD improves decoding in looped Transformers](https://ai.exponencialadm.net/okf/en/radar/loopcd-melhora-a-decodificacao-em-transformers-3d015a.md) - Radar: LoopCD refines decoding in looped Transformers by using the gap between early and late predictions of recurrent states, with no extra training, and reports accuracy gains on AIME 2024 with the Ouro-2.6B-Thinking model.
* [SFT with sampling for post-training: claims on generalization and forgetting](https://ai.exponencialadm.net/okf/en/radar/sft-com-sampling-para-post-training-c736eb.md) - A post describes adding sampling to the post-training stack and claims SFT can compete with current methods, with better generalization and less forgetting in some cases.
* [Training language models across multiple agent harnesses](https://ai.exponencialadm.net/okf/en/radar/treinamento-de-modelos-de-linguagem-em-23cd6f.md) - A guide describes a proxy that records token IDs and sampled logprobs to train models with RL across four agent harnesses, with reported score gains and a comparison with imitation learning.
* [When Transformers Need Chain of Thought](https://ai.exponencialadm.net/okf/en/radar/chain-of-thought-e-tarefas-seriais-c6bc13.md) - The article examines whether chains of thought do computational work in serial tasks; according to the post, removing them lowers accuracy to the shallow-circuit limit.
* [MOPD combines capabilities from specialist models](https://ai.exponencialadm.net/okf/en/radar/destilacao-on-policy-com-varios-professores-63e19b.md) - Published on October 1, 2026, the text presents multi-teacher on-policy distillation (MOPD): a method that trains a student model on its own completions, using teachers’ per-token log probabilities and a reverse KL objective.
* [Local inference: evaluate performance and exposure in the same test](https://ai.exponencialadm.net/okf/en/radar/inferencia-local-avalie-desempenho-e-exposicao-736e5e.md) - Running a model close to data can change the architecture. It does not replace input processing, access control or real evaluation.
* [Jagged Flash Attention with TLX on NVIDIA Blackwell](https://ai.exponencialadm.net/okf/en/radar/jagged-flash-attention-com-tlx-para-9f70ad.md) - A PyTorch blog post describes Jagged Flash Attention, an attention kernel behind Meta's Generative Ads Model, developed with TLX for NVIDIA Blackwell. The post reports better performance than FlashAttention-4 on the GEM jagged shapes.
* [Looped-DiT reuses Transformer blocks for image generation](https://ai.exponencialadm.net/okf/en/radar/looped-dit-testa-reutilizacao-de-blocos-cf8b9d.md) - Published on October 1, 2026, the post says Looped-DiT reuses Transformer blocks during denoising and that a 260-million-parameter model beat one 6.5 times larger on text-to-image benchmarks, with 4.9 times less inference computation.
* [Radar: cache and attention techniques in LLMs](https://ai.exponencialadm.net/okf/en/radar/otimizacao-de-kv-cache-e-atencao-5aa8eb.md) - A post dated October 1, 2026 reports 890 bytes of global KV cache per token and a persistent cache eight times smaller in a language model.
* [Radar: attention factorization and gradient flow](https://ai.exponencialadm.net/okf/en/radar/attention-fatorada-e-fluxo-de-gradientes-465204.md) - A September 30, 2026 publication argues that factoring S = WWᵀ changes gradient flow in attention and removes an information-exponent bottleneck that can stall unfactorized models.
* [DeepEP-Ascend: communication for Huawei NPUs](https://ai.exponencialadm.net/okf/en/radar/deepep-ascend-e-comunicacao-para-npus-702758.md) - Published on September 30, 2026, the item describes DeepEP-Ascend as a high-performance communication library for machine learning training and inference on Huawei Ascend NPUs.
* [Efficient Transformer inference at scale](https://ai.exponencialadm.net/okf/en/radar/eficiencia-na-inferencia-de-transformers-d269cb.md) - An article by Reiner Pope and coauthors discusses scaling Transformer inference efficiently, a relevant topic for engineers evaluating ways to serve these models.
* [ProRL Agent separates rollouts from LLM training](https://ai.exponencialadm.net/okf/en/radar/prorl-agent-separa-rollouts-do-treino-7fff19.md) - Published on September 30, 2026, ProRL Agent is open-source infrastructure for multi-turn LLM agent rollouts. Its rollout-as-a-service model separates trajectory generation, environment orchestration, and evaluation from policy optimization.
* [BF16 rounding may destabilize FlashAttention gradients](https://ai.exponencialadm.net/okf/en/radar/bf16-no-flashattention-arredondamento-e-gradientes-fdeb02.md) - A September 29, 2026 report says BF16 rounding in FlashAttention may break a softmax-gradient symmetry and cause gradients to explode near the end of training. The authors propose GProj as a correction.
* [Study organizes positional encoding alternatives](https://ai.exponencialadm.net/okf/en/radar/como-comparar-positional-encodings-em-transformers-576fc4.md) - The article examines positional encodings for Transformers, notes RoPE's bias toward nearby tokens, and organizes evaluations of alternatives across distinct configurations.
* [DepthBench examines computational depth in Transformers](https://ai.exponencialadm.net/okf/en/radar/depthbench-compara-metodos-para-transformers-mais-d8a634.md) - A September 29, 2026 preview presents a benchmark for comparing information-flow methods and assessing whether deeper Transformers use their layers effectively.
* [Radar: distilling between Transformers and Mamba](https://ai.exponencialadm.net/okf/en/radar/destilacao-transfere-conhecimento-de-transformers-para-78af0c.md) - Published on September 29, 2026, the summary describes a study on transferring knowledge from pretrained Transformers to Mamba models and notes memory and throughput advantages for SSMs.
* [KITE expands Transformers with invariant KV computation](https://ai.exponencialadm.net/okf/en/radar/kite-amplia-transformers-com-calculo-de-d828c4.md) - A September 29, 2026 entry describes KITE, a scaling paradigm combining scaling with upcycling and expanding decoder components that reuse a KV cache generated by a smaller prefiller.
* [Agent memory: preserve context without accumulating personal data](https://ai.exponencialadm.net/okf/en/radar/memoria-de-agentes-guarde-contexto-sem-576d18.md) - Long context can extend exposure. Separate what an agent needs to remember from the identity of people mentioned in a conversation.
* [MiMo RL Environment Explorer](https://ai.exponencialadm.net/okf/en/radar/mimo-rl-environment-explorer-explore-ambientes-69584b.md) - Published on September 29, 2026, the MiMo RL Environment Explorer is described as an interactive space to inspect MiMo-V2.6 reinforcement-learning environments and run rollouts.
* [A 7B model learns continuously from prediction markets](https://ai.exponencialadm.net/okf/en/radar/modelo-7b-aprende-continuamente-com-mercados-95ed8a.md) - A publication dated September 29, 2026 describes a 7B model that continuously updated its weights using prediction-market data and ran continuous backtests over 390 days. It says the model outperformed several larger models in the same test environment.
* [OPAL trains linear attention with its own outputs](https://ai.exponencialadm.net/okf/en/radar/opal-treina-linear-attention-com-as-2e0d7b.md) - The approach distills linear attention models in long contexts and, according to the report, fully recovers full-attention retrieval while reaching 83–93% of mathematical reasoning performance with 3 billion tokens.
* [OPD tested with a single training example](https://ai.exponencialadm.net/okf/en/radar/opd-em-um-exemplo-destilacao-com-9e3c59.md) - A paper published on September 29, 2026 tests on-policy distillation with one query across mathematics, coding, instruction following, and agentic tool use.
* [PTXBench measures AI-written GPU kernels](https://ai.exponencialadm.net/okf/en/radar/ptxbench-mede-kernels-de-gpu-escritos-723413.md) - Published on September 29, 2026, PTXBench evaluates model-generated matrix multiplication and attention kernels in architecture-specific PTX for H100 and B200 GPUs.
* [RESA aims to recover context lost in sparse attention](https://ai.exponencialadm.net/okf/en/radar/resa-recupera-contexto-ignorado-pela-atencao-f65358.md) - The September 29, 2026 entry describes RESA, a residual-estimation proposal for recovering context missed when sparse attention selects only part of the KV cache. The post claims the technique adds no computational cost.
* [TIRx Harness explores GPU kernel optimization](https://ai.exponencialadm.net/okf/en/radar/tirx-harness-kernels-gpu-otimizados-por-f77183.md) - Announced on September 29, 2026, TIRx Harness is an open compiler harness for agents to explore, debug, and optimize GPU kernels, with compiler analysis, a catalog, hardware specifications, and remote evaluation.
* [Vx uses inline MLIR for leading-zero operations](https://ai.exponencialadm.net/okf/en/radar/vx-faz-lowering-inline-de-mlir-4eac70.md) - Published on September 29, 2026, the bookmark describes Vx's use of the `mlir!` macro to lower MLIR directly into leading-zero count instructions, avoiding manually maintained routines for different hardware.
* [Autonomous agents: constrain tools and process incoming data](https://ai.exponencialadm.net/okf/en/radar/agentes-autonomos-limite-ferramentas-e-trate-d8b689.md) - An isolated runtime controls what an agent can do. Privacy also requires controlling the content it sends to AI.
* [Agents optimize kernels for Kimi Delta Attention](https://ai.exponencialadm.net/okf/en/radar/agentes-otimizam-kernels-para-kimi-delta-d44ba5.md) - Published on September 28, 2026, the post reports that kernel design agents generated and optimized Kimi Delta Attention kernels, reaching up to 2.96× the performance of FlashKDA on B300 with one-tenth the state error.
* [Log-linear block selection for sparse attention](https://ai.exponencialadm.net/okf/en/radar/atencao-esparsa-reduz-selecao-de-blocos-2437d8.md) - Published on September 28, 2026, the article describes a hierarchical block-selection method for sparse attention in long contexts, reducing selection cost from O(N²) to O(N log N).
* [Audio and AI: transcripts need a data policy too](https://ai.exponencialadm.net/okf/en/radar/audio-e-ia-a-transcricao-tambem-768193.md) - Turning speech into text changes the format, not the sensitivity. Evaluate every stage before introducing audio into a workflow.
* [FrontierSmith releases 200 open RL problems](https://ai.exponencialadm.net/okf/en/radar/frontiersmith-reune-200-problemas-abertos-de-ee313e.md) - Published on September 28, 2026, the collection presents 200 environments for studying reinforcement learning in challenging, open-ended, long-horizon tasks.
* [IFM-AI opens tools for LLM training and serving](https://ai.exponencialadm.net/okf/en/radar/ifm-ai-abre-ferramentas-para-treino-3aefda.md) - A September 28, 2026 announcement presents open-source tools for pretraining, attention, checkpoint conversion, and language-model serving.
* [Local inference: evaluate performance and exposure in the same test](https://ai.exponencialadm.net/okf/en/radar/inferencia-local-avalie-desempenho-e-exposicao-1670ee.md) - Running a model close to data can change the architecture. It does not replace input processing, access control or real evaluation.
* [Matryoshka Attribution ranks components](https://ai.exponencialadm.net/okf/en/radar/matryoshka-attribution-mapeia-influencias-no-comportamento-9847c4.md) - Published on September 28, 2026, the post describes a technique that ranks neural representations or fine-tuning weight changes by their influence on behavior preserved under soft interventions.
* [Vx models heterogeneous memory for compiler checks](https://ai.exponencialadm.net/okf/en/radar/memoria-heterogenea-verificacoes-de-compilador-com-cd0632.md) - Published on September 28, 2026, the post presents Vx, which describes memory capacity, connections, and visibility to check programs. It reports two capacity errors found on a laptop without a GPU.
* [On-policy self-distillation for reasoning in LLMs](https://ai.exponencialadm.net/okf/en/radar/self-distillation-on-policy-para-raciocinio-b6f808.md) - Published on September 28, 2026, the material presents Self-Distilled Reasoner, an on-policy self-distillation approach for language models. Its preview describes student-sampled trajectories and dense teacher supervision at token level.
* [SIGReg and repulsion in JEPA training](https://ai.exponencialadm.net/okf/en/radar/sigreg-e-a-repulsao-entre-pares-abb5f9.md) - A September 28, 2026 post describes how SIGReg, used by LeJEPA and LeWM, approximates a sliced MMD whose expansion reveals repulsion between pairs of samples in a batch.
* [HySparse2 proposes two-level KV sharing](https://ai.exponencialadm.net/okf/en/radar/hysparse2-combina-compartilhamento-de-kv-em-a2b8f5.md) - Published on September 27, 2026, HySparse2 is described as a hybrid sparse-attention method with two-level KV sharing, aimed at efficient prefill and compact KV-cache storage for agents with long contexts and multiple interactions.
* [Efficient inference for probabilistic Transformers](https://ai.exponencialadm.net/okf/en/radar/inferencia-autorregressiva-em-transformers-probabilisticos-b10a16.md) - An ICLR 2026 paper addresses efficient autoregressive inference for probabilistic Transformer models and, according to the post, a decoding bottleneck.
* [LENS combines reasoning and text-guided image segmentation](https://ai.exponencialadm.net/okf/en/radar/lens-combina-raciocinio-e-segmentacao-de-ecab88.md) - Published on September 27, 2026, the item describes LENS, a reinforcement-learning framework that jointly optimizes reasoning and text-guided image segmentation.
* [Memory Attention reuses keys in V=K+M](https://ai.exponencialadm.net/okf/en/radar/memory-attention-testa-reutilizar-keys-em-1a1980.md) - A September 27, 2026 publication reports preliminary ablations suggesting V=K+M is an effective architectural choice; the result still needs validation at larger scales.
* [Post claims IPA decryption without a physical iPhone](https://ai.exponencialadm.net/okf/en/radar/post-alega-descriptografia-de-ipa-sem-7d6bfc.md) - A report claims that FairPlay-protected iOS IPAs can be decrypted without a physical iPhone or jailbreak. The claim may interest people studying app protection, but the briefing does not verify it.
* [Xtrace: intra-kernel GPU tracing](https://ai.exponencialadm.net/okf/en/radar/xtrace-tracing-de-kernels-gpu-por-a91528.md) - A paper on intra-kernel GPU tracing through binary-level instruction splicing; its preview describes probes inserted into kernels to record execution data.
* [Native sparse attention for long contexts](https://ai.exponencialadm.net/okf/en/radar/atencao-esparsa-nativa-para-contextos-longos-84a8d3.md) - An article presents NSA, a trainable sparse-attention mechanism that combines algorithmic choices with hardware-aligned optimizations for efficient long-context modeling.
* [Evaluate agents with less context and explicit criteria](https://ai.exponencialadm.net/okf/en/radar/avalie-o-agente-com-menos-contexto-a12bae.md) - A more elaborate workflow does not always improve outcomes. Compare components by task quality, consumption and data exposure.
* [Sparse layers and looping in language models](https://ai.exponencialadm.net/okf/en/radar/camadas-esparsas-e-looping-na-escala-93c18c.md) - Published on September 26, 2026, the article compares standard Transformers and Mixture-of-Experts, with and without looping. The authors report that Looped-MoE scales better than the standard baseline, while dense-looping models do not.
* [How vLLM schedules requests and manages the KV cache](https://ai.exponencialadm.net/okf/en/radar/como-o-vllm-agenda-requisicoes-e-370456.md) - An article published on September 26, 2026 follows a vLLM request from queue to decoding and explains per-iteration scheduling and KV-cache block allocation.
* [Sharing model context also shares data risk](https://ai.exponencialadm.net/okf/en/radar/compartilhar-contexto-entre-modelos-tambem-compartilha-648b8e.md) - Internal representations do not make information anonymous. Define data boundaries before connecting multiple models.
* [Unsloth's GRPO guide for reinforcement learning](https://ai.exponencialadm.net/okf/en/radar/guia-de-grpo-da-unsloth-do-5e4fcd.md) - Published on September 26, 2026, Unsloth's guide ranges from beginner to advanced and covers using GRPO to train reasoning models.
* [JAZ explores minimalist recursive agents](https://ai.exponencialadm.net/okf/en/radar/jaz-explora-loops-recursivos-para-agentes-df21a8.md) - Published on September 26, 2026, the note presents JAZ, a framework using a single recursive code primitive for agent loops, and reports a claim of lower cost and better performance than two alternatives.
* [Agent loops: simplify workflows without expanding permissions](https://ai.exponencialadm.net/okf/en/radar/loops-de-agentes-simplifique-o-fluxo-f98b08.md) - Fewer components can ease maintenance. Data and action boundaries must remain explicit at every step.
* [PolaFormer++ proposes polarity-aware features](https://ai.exponencialadm.net/okf/en/radar/polaformer-propoe-features-com-consciencia-de-0867b4.md) - Published on September 26, 2026, the work presents a theoretical criterion for feature-map spikiness and a channel-wise polarity-aware feature map. The authors evaluate the approach across six families of visual tasks.
* [Self-play explores pretraining without natural data](https://ai.exponencialadm.net/okf/en/radar/self-play-testa-pre-treinamento-sem-776002.md) - A study uses program-generated byte sequences and an adaptive generator to investigate whether systems can learn transferable structures without training on natural data.
* [SVD-LLM proposes model compression that accounts for truncation](https://ai.exponencialadm.net/okf/en/radar/svd-llm-propoe-compressao-de-modelos-b49995.md) - Published on September 26, 2026, the paper presents an LLM compression method using singular value decomposition (SVD) that updates compressed weights after truncation.
* [Reducing Transformer loop redundancy](https://ai.exponencialadm.net/okf/en/radar/transformers-com-loops-menos-redundancia-menos-756f0f.md) - An article reports three kinds of computational redundancy across loops in looped Transformers. The authors say exploiting them reduces latency by 1.65× and memory use by 6×.
* [TWT cuts compute in Vision Transformers](https://ai.exponencialadm.net/okf/en/radar/twt-reduz-o-custo-de-vision-66553a.md) - Published on September 26, 2026, the bookmark says Transformer-Within-Transformer replaces contiguous groups of similar layers with substitute attention layers. In DINOv2, it reports about half the compute, with a small drop in accuracy.
* [Xiaomi MiMo releases RL environments and training code](https://ai.exponencialadm.net/okf/en/radar/xiaomi-mimo-publica-recursos-para-pesquisa-8e1aaa.md) - On September 26, 2026, a Xiaomi MiMo post pointed to reinforcement learning environments, training code, and a MiMo-V2.6 RL dataset for exploring language-model workflows.
* [Memory Attention proposes token memory](https://ai.exponencialadm.net/okf/en/radar/memory-attention-propoe-memoria-de-tokens-3d202d.md) - Published on September 25, 2026, the post describes an alternative to attention value projection: token-indexed memory specific to each layer, combined with contextual keys.
* [Memory Attention adds token-indexed memory](https://ai.exponencialadm.net/okf/en/radar/memory-attention-propoe-memoria-indexada-por-29b569.md) - A post published on September 25, 2026 describes a change to Transformer attention: replacing the learned value projection with a sum of contextual keys and token-indexed memory vectors specific to each layer. The memory can reside on the CPU.
* [NVIDIA Model Optimizer gathers model compression techniques](https://ai.exponencialadm.net/okf/en/radar/nvidia-model-optimizer-reune-tecnicas-de-9e9041.md) - NVIDIA library with quantization, distillation, pruning, neural architecture search and speculative decoding, with checkpoints exportable to inference runtimes.
* [Self-play explores pretraining without real data](https://ai.exponencialadm.net/okf/en/radar/self-play-explora-pre-treinamento-sem-8db569.md) - A study investigates scalable pretraining from random initialization using a program generator and a learner, without training on real data.
* [SmolDataEnvs collects 5,000 verifiable RL tasks](https://ai.exponencialadm.net/okf/en/radar/smoldataenvs-reune-5-000-tarefas-verificaveis-bdc1c1.md) - Launched on September 25, 2026, the open-source collection includes tasks, environments, evaluations, and training to optimize small models for programming and data science.
* [Inference caching: evaluate performance and isolation](https://ai.exponencialadm.net/okf/en/radar/cache-de-inferencia-desempenho-e-isolamento-9d30e4.md) - Efficient memory can improve service capacity. The design must also prevent one session's context from appearing in another session.
* [Test-time training layers for sequence modeling](https://ai.exponencialadm.net/okf/en/radar/camadas-de-test-time-training-para-e52f0f.md) - An article published on September 24, 2026 presents linear-complexity layers with expressive hidden states, updated through self-supervised learning.
* [Backing up high-value secret keys](https://ai.exponencialadm.net/okf/en/radar/como-fazer-backup-de-secret-keys-32b77b.md) - A September 24, 2026 bookmark points to an article about backing up valuable secret keys. The record notes the security concern but gives no specific methods or results.
* [Differential Transformer subtracts attention maps](https://ai.exponencialadm.net/okf/en/radar/differential-transformer-subtrai-mapas-de-atencao-e7736b.md) - A September 24, 2026 publication describes an attention variant that subtracts two maps to reduce noise and highlight signal. It notes potential interest for engineers exploring alternatives to standard attention.
* [From REINFORCE to GRPO: policy-gradient methods for LLMs](https://ai.exponencialadm.net/okf/en/radar/do-reinforce-ao-grpo-metodos-de-4da424.md) - A publication dated September 24, 2026, outlines policy-gradient methods used to train LLMs, from vanilla policy gradient and REINFORCE to PPO, GRPO, and GRPO variants.
* [Weight spectra signal memorization in LLMs](https://ai.exponencialadm.net/okf/en/radar/espectros-de-pesos-podem-sinalizar-memorizacao-90e49c.md) - A preliminary analysis reports that the attention matrix K exponent α tracks memorization of random labels during training, with variation across layers.
* [FlashAttention-3 aims to speed up attention on Hopper GPUs](https://ai.exponencialadm.net/okf/en/radar/flashattention-3-busca-acelerar-attention-em-fff924.md) - The paper presents techniques for speeding up attention on Hopper GPUs, exploring asynchronous execution, low-precision computation, memory traffic, and hardware utilization.
* [LoRA+ proposes distinct learning rates for adapter matrices](https://ai.exponencialadm.net/okf/en/radar/lora-propoe-taxas-distintas-para-matrizes-1f2171.md) - Dated September 24, 2026, the summary describes an analysis of equal learning rates for LoRA matrices A and B and a proposal to use different rates in wide models.
* [Mamba and Transformers through state-space duality](https://ai.exponencialadm.net/okf/en/radar/mamba-e-transformers-a-conexao-pela-664b03.md) - Published on September 24, 2026, the article explores theoretical links between state-space models such as Mamba and attention variants, using structured semiseparable matrix decompositions.
* [Megalodon explores sequence modeling with unlimited context](https://ai.exponencialadm.net/okf/en/radar/megalodon-explora-sequencias-com-contexto-ilimitado-3a3bfe.md) - An architecture designed for efficient pretraining and inference on long sequences aims to address Transformer limitations.
* [Memory Attention proposes an alternative to value projections](https://ai.exponencialadm.net/okf/en/radar/memory-attention-troca-projecoes-de-valores-eb2117.md) - Published on September 24, 2026, the summary introduces Memory Attention, a proposal replacing the Transformer's learned value projection and examined across several attention variants.
* [Document RAG: process data before composing the answer](https://ai.exponencialadm.net/okf/en/radar/rag-com-documentos-trate-os-dados-25a084.md) - Retrieval can combine information from several documents. Privacy must follow retrieval, context assembly and the answer.
* [RetNet combines parallel training and recurrent inference](https://ai.exponencialadm.net/okf/en/radar/retnet-une-treinamento-paralelo-e-inferencia-d7550d.md) - Published on September 24, 2026, the article presents the Retentive Network, a language-model architecture proposing parallel training alongside recurrent and chunkwise recurrent inference paradigms.
* [Encryption and privacy: secure transit and reduce content](https://ai.exponencialadm.net/okf/en/radar/criptografia-e-privacidade-proteja-o-transporte-69314f.md) - An encrypted connection protects transit. The recipient still receives what you send, so policy must act before delivery.
* [HySparse2 explores KV sharing for long contexts](https://ai.exponencialadm.net/okf/en/radar/hysparse2-reduz-custos-em-inferencia-com-fff083.md) - A post published on September 23, 2026 describes two levels of KV sharing intended to reduce prefill cost and cache size, while improving retrieval in long-context workloads.
* [NVIDIA launches Nemotron 3 Diarization](https://ai.exponencialadm.net/okf/en/radar/nemotron-3-diarization-identifica-falas-sobrepostas-24ddfb.md) - Announced on September 23, 2026, Nemotron 3 Diarization identifies who spoke and when, including in overlapping speech, and supports up to eight speakers.
* [DSec: sandboxes for training agents at scale](https://ai.exponencialadm.net/okf/en/radar/sandboxes-elasticos-para-treinar-agentes-em-bfcffa.md) - On September 23, 2026, an article on DSec described elastic, stateful environments for agent training and evaluation, as well as defenses against agents that exploit training environments.
* [Streaming speech: protect personal data before transcription](https://ai.exponencialadm.net/okf/en/radar/streaming-de-fala-proteja-dados-pessoais-c9ffbd.md) - Real-time ASR and translation models can process speech containing personal data; Rota Nacional’s privacy barrier applies the organization’s policy before a model runs.
* [Zlaya: Laya CPU inference in Zig](https://ai.exponencialadm.net/okf/en/radar/zlaya-inferencia-do-laya-em-cpu-59a46d.md) - Zlaya is a CPU inference engine for Laya, implemented in Zig and described in its repository as fast and compact.
* [iSDFT explores continual learning in LLMs](https://ai.exponencialadm.net/okf/en/radar/isdft-propoe-self-distillation-para-continual-cae1e2.md) - Published on September 22, 2026, iSDFT describes a continual-learning method for LLMs based on on-policy self-distillation and information proximity. Its repository provides training code, checkpoints, and evaluation results.
* [mini-AGI explores continual learning with 8 GB of VRAM](https://ai.exponencialadm.net/okf/en/radar/mini-agi-explora-aprendizado-continuo-com-9f6881.md) - Published on September 22, 2026, the account describes a model trained from scratch on a notebook with 8 GB of VRAM, using batch size 1, disk-paged weights, and learning-rate adjustments to reduce forgetting.
* [Mixture-of-Depths allocates compute per token](https://ai.exponencialadm.net/okf/en/radar/mixture-of-depths-distribui-compute-por-2d8121.md) - An article published September 22, 2026 presents a method for Transformers to allocate compute to sequence positions while respecting a per-layer budget.
* [onPanda: model annotation and inspection](https://ai.exponencialadm.net/okf/en/radar/onpanda-anotacao-por-token-e-inspecao-628905.md) - Published on September 22, 2026, this entry describes an open-source tool for annotating LLM data and inspecting models, including token-level edits and inspection of token probabilities and decoding.
* [Quiet-STaR explores reasoning before an answer](https://ai.exponencialadm.net/okf/en/radar/quiet-star-explora-raciocinio-antes-da-d22cb0.md) - Published on September 22, 2026, the brief describes a training approach that teaches language models to generate internal reasoning before visible text.
* [Production sandboxes: reported scale](https://ai.exponencialadm.net/okf/en/radar/sandboxes-em-producao-escala-relatada-em-fe763f.md) - A September 22, 2026 publication cites a paper on sandbox infrastructure and reports production-scale operating figures.
* [Xeno-interpretability and native representations](https://ai.exponencialadm.net/okf/en/radar/xeno-interpretabilidade-limites-das-representacoes-8ec464.md) - A paper investigates whether language models use internal distinctions for which humans lack adequate concepts. It calls these structures xeno-representations and examines the limits of interpreting them through familiar human concepts.
* [CodeMidas creates RL tasks from code](https://ai.exponencialadm.net/okf/en/radar/codemidas-cria-tarefas-de-rl-a-f7b248.md) - Published on September 21, 2026, the report introduces Xiaomi’s CodeMidas: agents extract executable tasks from open-source codebases. The system generated 5,545 tasks across 23 languages.
* [CodeMidas proposes RL tasks for coding](https://ai.exponencialadm.net/okf/en/radar/codemidas-transforma-codigo-fonte-em-ambientes-cc05c8.md) - Published on September 21, 2026, the announcement describes a pipeline that turns features from open-source projects into executable reinforcement-learning environments.
* [MiMo Pro raises its DeepSWE benchmark score](https://ai.exponencialadm.net/okf/en/radar/rl-eleva-pontuacao-em-benchmark-de-a5bf5f.md) - A post reports that MiMo Pro's DeepSWE score rose from 58.41 to 72.57 after reinforcement-learning runs, nearing comparative results cited in the post.
* [TR-DPO adds a KL penalty to the DPO loss](https://ai.exponencialadm.net/okf/en/radar/tr-dpo-acrescenta-penalidade-kl-a-e55a42.md) - A September 21, 2026 publication describes Trust Region DPO (TR-DPO), which incorporates a KL divergence penalty into the DPO loss to limit deviation from the reference model.
* [Voice of Reason answers spoken math problems](https://ai.exponencialadm.net/okf/en/radar/voz-nativa-para-resolver-problemas-matematicos-f31e33.md) - In a September 21, 2026 post, Kyutai Labs says its model reasons about spoken math problems and answers by voice, without an intermediate transcript or text LLM. The post reports 77.1% on GSM8K, versus 27.3% for GLM-4-Voice.
* [BITCOS reduces ternary weights to 1.485 bits per weight](https://ai.exponencialadm.net/okf/en/radar/bitcos-reduz-pesos-de-llms-ternarios-99d03b.md) - A September 20, 2026 report on Intel’s BITCOS method says it stores exact ternary weights at 1.485 bits per weight and reports decode-throughput gains of up to 18% on CPUs and 27% on GPUs.
* [Gated Linear Attention combines gates and linear attention](https://ai.exponencialadm.net/okf/en/radar/gated-linear-attention-une-atencao-linear-9a5ec7.md) - Published on September 20, 2026, the item describes a transformer formulation with linear attention, data-dependent gates, and an equivalent representation as an RNN with matrix-valued hidden states.
* [Jamba combines Transformer and Mamba layers](https://ai.exponencialadm.net/okf/en/radar/jamba-combina-transformer-mamba-e-especialistas-011a4e.md) - Published on September 20, 2026, the item describes Jamba, a base language model with a hybrid Transformer-Mamba architecture and mixture-of-experts.
* [On-Demand Attention selects when to use full context](https://ai.exponencialadm.net/okf/en/radar/on-demand-attention-escolhe-quando-usar-f16c22.md) - Published on September 20, 2026, the post describes a technique that uses a lightweight head to trigger full context when needed, aiming to speed decoding while recovering performance lost with local attention.
* [QLoRA reports fine-tuning a 65-billion-parameter model on a 48 GB GPU](https://ai.exponencialadm.net/okf/en/radar/qlora-reduz-memoria-para-ajustar-modelos-ccc7c6.md) - Published on September 20, 2026, the article summary reports fine-tuning a 65-billion-parameter model on one 48 GB GPU, using 4-bit quantization and LoRA adapters.
* [RWKV combines parallel training with recurrent inference](https://ai.exponencialadm.net/okf/en/radar/rwkv-une-treino-paralelo-e-inferencia-4d10fc.md) - Published on September 20, 2026, the article summary presents an architecture combining Transformer-style parallel training with RNN-style recurrent inference.
* [Speculative decoding may speed up inference](https://ai.exponencialadm.net/okf/en/radar/speculative-decoding-pode-acelerar-a-inferencia-6cca8e.md) - The technique pairs proposals from an efficient model with parallel checks by a larger model to reduce serial decoding steps without changing the output distribution.
* [GEPA adapts a model with prompts and labeled examples](https://ai.exponencialadm.net/okf/en/radar/como-adaptar-decisoes-de-um-modelo-3b4315.md) - A publication dated September 19, 2026 describes a GEPA demonstration adapting a commercially available model to specific decision criteria by changing its prompt and labeling examples.
* [CUDA C++: GPU execution, memory, and correctness](https://ai.exponencialadm.net/okf/en/radar/cuda-c-execucao-memoria-e-correcao-073ca9.md) - A concise guide to core CUDA programming concepts: kernel launches, thread organization, memory, synchronization, runtime initialization, errors, and validating GPU results against a CPU implementation.
* [Radar: language models that self-evaluate](https://ai.exponencialadm.net/okf/en/radar/feedback-de-llm-as-a-judge-63a1a4.md) - Published on September 19, 2026, the paper studies language-model feedback prompted as LLM-as-a-Judge during training, compared with reward models based on human preferences that remain frozen.
* [GuppyLM: a language model with 9 million parameters](https://ai.exponencialadm.net/okf/en/radar/guppylm-experimento-de-linguagem-com-9-eb9a0d.md) - A post introduces GuppyLM, a language model with about 9 million parameters, and claims it can be trained from scratch in five minutes. The project is presented as a small-scale way for engineers to explore language-model training.
* [ORPO combines SFT and preference optimization](https://ai.exponencialadm.net/okf/en/radar/orpo-reune-ajuste-supervisionado-e-preferencias-ed6f8e.md) - Published on September 19, 2026, the paper introduces ORPO, a method combining preference alignment and supervised fine-tuning without a reference model.
* [Open-source tax-page classifier reports 100% accuracy](https://ai.exponencialadm.net/okf/en/radar/classificador-open-source-de-paginas-fiscais-1365d4.md) - A GitHub project describes classifying pages of tax documents and reports results on 261 IRS forms, at an estimated cost of about US$0.001 per page.
* [JEPA-Anything explores orthogonal factors](https://ai.exponencialadm.net/okf/en/radar/jepa-anything-explora-fatores-ortogonais-5b8cfb.md) - The domain-agnostic framework uses orthogonal predictive factorization to decompose latent targets into complementary factors; its authors say the code, checkpoints, and report are open.
* [Never Give Up steers RL toward harder problems](https://ai.exponencialadm.net/okf/en/radar/never-give-up-direciona-compute-de-3812ca.md) - A note dated September 18, 2026, reports that reinforcement learning (RL) tends to improve LLM performance more on easy problems than on hard ones. The Never Give Up method uses adaptive sampling to direct compute toward harder cases.
* [New OCR models and benchmark result](https://ai.exponencialadm.net/okf/en/radar/novos-modelos-de-ocr-ampliam-opcoes-e22f00.md) - A September 18, 2026 publication presents WeVisDoc 2B and 4B and jina-ocr-v1, with different licenses, and reports WeVisDoc-4B's OmniDocBench v1.6 score.
* [Qualcomm releases IDL FastRPC compiler for Hexagon](https://ai.exponencialadm.net/okf/en/radar/qualcomm-abre-o-compilador-idl-fastrpc-f937f4.md) - On September 18, 2026, Qualcomm announced the release of the IDL FastRPC compiler source code for Hexagon. Engineers can consult the code in the QAIC repository on GitHub.
* [Flow Models as recurrent models for reasoning](https://ai.exponencialadm.net/okf/en/radar/flow-models-refinam-raciocinio-em-etapas-fc86bf.md) - A paper described on September 17, 2026, turns a flow model into a recurrent model for tasks such as Sudoku, refining predictions iteratively without backpropagation through time.
* [Agora records shared agent memory in Git](https://ai.exponencialadm.net/okf/en/radar/git-como-memoria-compartilhada-para-agentes-310468.md) - In a September 17, 2026 publication, Agora describes an append-only Git DAG for recording research-agent contributions and making results and verification status searchable.
* [GrainSpeech proposes speech synthesis with less context](https://ai.exponencialadm.net/okf/en/radar/grainspeech-explora-sintese-de-fala-com-722676.md) - Published on September 17, 2026, the GrainSpeech project is described as generating more detail in speech synthesis with less context. Its project page points to an official repository and a technical paper.
* [RULER turns agent evaluations into GRPO rewards](https://ai.exponencialadm.net/okf/en/radar/ruler-transforma-comparacoes-de-agentes-em-0b6268.md) - Published on September 17, 2026, the report describes RULER in OpenPipe ART: an LLM ranks trajectories using plain-English criteria, and the comparisons are converted into rewards for GRPO training.
* [Supersonic TTS: CPU and support for 31 languages](https://ai.exponencialadm.net/okf/en/radar/supersonic-tts-roda-em-cpu-e-a42696.md) - Published on September 17, 2026: the author says Supersonic TTS runs on CPU and supports 31 languages. The author plans to use it to generate synthetic data.
* [TinyLoRA tests RL fine-tuning with 13 parameters](https://ai.exponencialadm.net/okf/en/radar/tinylora-testa-fine-tuning-com-rl-cf2520.md) - A publication dated September 17, 2026 describes TinyLoRA and reports gains on three benchmarks using RL training with 13 parameters.
* [Transformers in a loop as programmable computers](https://ai.exponencialadm.net/okf/en/radar/transformers-em-loop-como-computadores-programaveis-283446.md) - Published on September 17, 2026, the article describes a framework using specific weights and recurrent Transformer layers to emulate computational operations, including conditional branches and program counters.
* [Block-recurrent Transformers process sequences with linear complexity](https://ai.exponencialadm.net/okf/en/radar/transformers-recorrentes-em-blocos-processam-sequencias-2acad7.md) - An architecture applies a Transformer layer recurrently to blocks of tokens, using self-attention and cross-attention to process sequences with complexity linear in their length.
* [Universal Transformers reuse one layer recurrently](https://ai.exponencialadm.net/okf/en/radar/universal-transformers-reutilizam-uma-camada-recorrentemente-22cfd2.md) - A September 17, 2026 post describes an architecture that repeatedly applies the same Transformer layer instead of stacking distinct layers, and highlights sequential versus parallel computation.
* [CocktailASR-1 transcribes a selected speaker in mixed audio](https://ai.exponencialadm.net/okf/en/radar/cocktailasr-1-transcreve-um-falante-em-b06eac.md) - Xiaomi's model uses a reference voice to transcribe a selected speaker in overlapping audio; the publication reports support for English and Chinese.
* [Crusoe reports inference performance on MI355X](https://ai.exponencialadm.net/okf/en/radar/crusoe-relata-desempenho-de-inferencia-em-7dfbbe.md) - In MLPerf Inference v6.1 results, Crusoe reported 5.75 million tokens per second for gpt-oss-120b using 512 AMD Instinct MI355X GPUs.
* [GOAT proposes trainable priors for attention](https://ai.exponencialadm.net/okf/en/radar/goat-propoe-priors-treinaveis-para-attention-52074b.md) - Published on September 16, 2026, the article presents attention as Entropic Optimal Transport and proposes replacing the implicit uniform prior with a continuous, trainable prior.
* [KATA explores recall in linear attention](https://ai.exponencialadm.net/okf/en/radar/kata-explora-spherical-packing-para-recall-e7a349.md) - A September 16, 2026 article frames associative recall in linear attention as a spherical packing problem and presents KATA activations derived from a homogeneous self-dual cone.
* [SAS proposes a context selector inside Softmax](https://ai.exponencialadm.net/okf/en/radar/sas-filtra-contexto-dentro-do-softmax-a50ebd.md) - In a September 16, 2026 publication, SAS describes a selector trained inside Softmax to filter context in attention designed for hardware with limited computational resources.
* [Softmax Linear Attention aims for efficient global competition](https://ai.exponencialadm.net/okf/en/radar/softmax-linear-attention-busca-eficiencia-sem-a07dc8.md) - Softmax Linear Attention proposes applying softmax at the head level to restore global competition in linear attention while retaining the goal of linear complexity. The approach can be compared with standard softmax normalization.
* [Switch Attention varies attention patterns](https://ai.exponencialadm.net/okf/en/radar/switch-attention-combina-padroes-de-atencao-f20f90.md) - A paper presents a dynamic, granular approach combining full and sliding-window attention in hybrid transformers, addressing the former's quadratic cost and the latter's narrower receptive field.
* [Thinking with Looped Flows uses local denoising](https://ai.exponencialadm.net/okf/en/radar/thinking-with-looped-flows-testa-denoising-bb40dc.md) - A post published on September 16, 2026, reports that local denoising objectives helped recurrent Looped Flows models with multi-step computation. It says they outperformed earlier looped models on six reasoning benchmarks, including 58.8% accuracy on ARC-AGI-1.
* [Tucker Attention unifies attention variants](https://ai.exponencialadm.net/okf/en/radar/tucker-attention-unifica-mha-gqa-e-19ed85.md) - The article presents Tucker Attention as a tensor-factorization approach that frames MHA, GQA, and MLA within a shared perspective on low-rank approximate attention.
* [Attention in Long-Context Transformers](https://ai.exponencialadm.net/okf/en/radar/atencao-em-transformers-de-contexto-longo-960766.md) - Published on September 15, 2026, the article examines rank collapse as context grows and analyzes rescaling attention scores.
* [auto-gpu-kernel 1.0 generates GPU kernels and harnesses](https://ai.exponencialadm.net/okf/en/radar/auto-gpu-kernel-1-0-gera-2dafcb.md) - Version 1.0 introduces an autonomous meta-harness that evolves both GPU kernels and the harness layer; the project won the DSA category in an MLSys 2026 FlashInfer contest.
* [Harness adds browser and computer options](https://ai.exponencialadm.net/okf/en/radar/harness-amplia-opcoes-experimentais-para-agentes-9deabb.md) - Version v0.1.6-alpha.1 adds a Web sidebar with persistent terminals, multiple tabs, and MCP resource support. Experimental options include browser tools and computer drivers.
* [Octop presents self-hosted, multi-user AI agents](https://ai.exponencialadm.net/okf/en/radar/octop-leva-agentes-de-ia-multiusuario-ef6bfc.md) - Published on September 15, 2026, the post describes Octop as an open-source, self-hosted AI assistant platform supporting multiple users and agents.
* [RoLA proposes low-rank linear attention for Diffusion Transformers](https://ai.exponencialadm.net/okf/en/radar/rola-propoe-atencao-linear-de-baixo-a82eab.md) - Published on September 15, 2026, the paper presents RoLA, a low-rank linear attention method with rotary position for Diffusion Transformers. It addresses quadratic scaling in dense spatiotemporal self-attention and a compatibility problem between RoPE and the global branch
* [RVC WebUI for training voice conversion models](https://ai.exponencialadm.net/okf/en/radar/rvc-webui-treinamento-de-modelos-de-63a99f.md) - Published on September 15, 2026, the entry describes an open-source project with a web interface for training voice conversion models using 10 minutes or less of voice data.
* [ASR tests isolating one speaker in overlapping speech](https://ai.exponencialadm.net/okf/en/radar/transcricao-de-um-locutor-em-fala-669961.md) - Published on September 15, 2026, Xiaomi-CocktailASR-1 demonstrates speaker-conditioned speech recognition on overlapping audio.
* [Radar: scaling byte and token models](https://ai.exponencialadm.net/okf/en/radar/escala-de-modelos-bytes-versus-tokens-dc63f8.md) - A paper compares distilled byte and token models trained at scales up to 1 trillion bytes and reports a higher performance ceiling for byte models as compute increases.
* [PC-ALM trains deep residual MLPs with local dynamics](https://ai.exponencialadm.net/okf/en/radar/pc-alm-explora-aprendizado-local-em-3f39ad.md) - Published on September 14, 2026, the summary presents PC-ALM as a local alternative to backpropagation for training residual MLPs up to 1,000 layers, with performance described as nearly equivalent.
* [Method trains sparse selection of KV blocks](https://ai.exponencialadm.net/okf/en/radar/simple-attention-sparsification-otimiza-blocos-kv-8cd0e4.md) - Published on September 14, 2026, Simple Attention Sparsification describes how language modeling loss can optimize which KV blocks each query attends to.
* [Tencent presents WeKnora, an LLM knowledge framework](https://ai.exponencialadm.net/okf/en/radar/weknora-propoe-uma-plataforma-de-conhecimento-005b43.md) - Published on September 14, 2026, the post introduces WeKnora, an open-source framework for document understanding, semantic retrieval, and autonomous reasoning.
* [GAD proposes on-policy distillation for black-box LLMs](https://ai.exponencialadm.net/okf/en/radar/gad-destilacao-de-modelos-sem-acesso-656aaf.md) - An article dated September 13, 2026 presents a method that learns from a proprietary teacher model’s text outputs, without access to logits or parameters.
* [Open book surveys AI infrastructure in 12 chapters](https://ai.exponencialadm.net/okf/en/radar/livro-aberto-reune-fundamentos-de-infraestrutura-51ffe1.md) - Published on September 13, 2026, the book covers model architectures, accelerators, inference optimization and distribution, training systems, and resource scheduling.
* [Post reports alleged KV cache reduction](https://ai.exponencialadm.net/okf/en/radar/post-relata-queda-alegada-no-uso-da8a4c.md) - A post claims KV cache use fell from 389,000 bytes to 890 bytes per token in three years. It says a linked article explains the approach with diagrams.
* [Token retention may reduce KV cache use](https://ai.exponencialadm.net/okf/en/radar/retencao-de-tokens-pode-reduzir-uso-509980.md) - A publication reports that keeping four initial tokens and a 64-token window, without training, matched or outperformed distilled linear-attention models on almost all benchmarks.
* [RLT carries recurrent state between tokens](https://ai.exponencialadm.net/okf/en/radar/rlt-mantem-estado-recorrente-entre-tokens-cfb85a.md) - Published on September 13, 2026, the brief describes a Transformer with a recurrent decoder that carries state between tokens during the prompt and generation.
* [A framework for recursive AI self-improvement](https://ai.exponencialadm.net/okf/en/radar/uma-estrutura-para-pensar-o-autoaperfeicoamento-f7581b.md) - An article published on September 13, 2026 presents a review and conceptual framework for cycles in which AI systems identify limitations, validate changes, and retain improvements.
* [The Deadlock Empire teaches concurrency through a game](https://ai.exponencialadm.net/okf/en/radar/aprenda-concorrencia-jogando-com-escalonamentos-691282.md) - Published on September 12, 2026, the listing describes The Deadlock Empire, a game where players act as schedulers, explore programs with bugs, and learn about concurrency.
* [Looped flows train recurrent updates with local denoising](https://ai.exponencialadm.net/okf/en/radar/looped-flows-exploram-denoising-local-no-ad4536.md) - Published on September 12, 2026, the article proposes training recurrent hidden-state updates with local denoising objectives, seeking to address a difficulty associated with backpropagation across only a few updates.
* [Large model shifts across GPU, RAM and NVMe](https://ai.exponencialadm.net/okf/en/radar/modelo-grande-entre-gpu-ram-e-32e5b3.md) - A report published on September 12, 2026 describes running a model on an RTX 5090 with 125.7 GiB of RAM and part of a 502 GB GGUF file on NVMe.
* [A visual introduction to contraction operators in RL](https://ai.exponencialadm.net/okf/en/radar/operadores-de-contracao-em-aprendizagem-por-4f240c.md) - An article in an intuitive reinforcement learning series presents contraction operators, a fundamental concept, through a concise visual approach.
* [Spinlock optimization and its trade-offs](https://ai.exponencialadm.net/okf/en/radar/otimizacao-de-spinlocks-e-seus-trade-1d43fc.md) - The article discusses spinlocks, which keep threads busy on the CPU instead of yielding to the scheduler, and notes that this can avoid system calls and context switches.
* [Document classifier with agents and SetFit](https://ai.exponencialadm.net/okf/en/radar/classificador-de-documentos-com-agents-e-b532e2.md) - Published on September 10, 2026, the account describes a document-purpose classifier trained on 200 examples labeled by an agent. The model classified 191,724 documents for about US$0.70 in inference compute.
* [Colibri streams MoE experts from disk in C](https://ai.exponencialadm.net/okf/en/radar/colibri-transmite-experts-de-moe-do-4db931.md) - Published on September 10, 2026, Colibri is described as a dependency-free, pure-C inference engine that streams MoE experts from disk to run models on local hardware.
* [PRISM and binary translation on Windows on Arm](https://ai.exponencialadm.net/okf/en/radar/prism-e-a-traducao-binaria-no-8713f0.md) - Published on September 10, 2026, this entry describes an article examining PRISM, the binary translation system used by Windows on Arm, and its effects on code.
* [Repositories opened for V4.1 Flash deployment](https://ai.exponencialadm.net/okf/en/radar/repositorios-abertos-facilitam-o-deploy-do-a4329c.md) - On September 10, 2026, the company announced that it opened three repositories to facilitate deployment of V4.1 Flash and later open models.
* [C*: programming and verification in C](https://ai.exponencialadm.net/okf/en/radar/c-unificando-programacao-e-verificacao-em-b02552.md) - Published on September 8, 2026, the article presents C*, an approach to bringing programming in C closer to formal verification.
* [TLS in Go: signing handshakes inside a TPM](https://ai.exponencialadm.net/okf/en/radar/assinando-handshakes-tls-em-go-com-bd8bd9.md) - Published on September 7, 2026, the article discusses client certificates whose private key stays in a TPM, Go's crypto/tls requirements, and per-handshake cost. It reports no quantitative results.
* [MaAI updates its English VAP model](https://ai.exponencialadm.net/okf/en/radar/maai-amplia-os-dados-de-treinamento-70633a.md) - On September 6, 2026, MaAI said its English VAP model was trained on a larger scale of data. The project predicts conversational cues and head movements in real time.
* [OmniVoice: speech synthesis and cloning in over 600 languages](https://ai.exponencialadm.net/okf/en/radar/omnivoice-sintese-e-clonagem-de-voz-7926ac.md) - A September 6, 2026 publication describes OmniVoice, a Python text-to-speech project with voice cloning, zero-shot synthesis, and generation controls.
* [ripwire builds structural repository context for coding agents](https://ai.exponencialadm.net/okf/en/radar/ripwire-mapeia-a-estrutura-de-repositorios-b2de0a.md) - Published on September 6, 2026, ripwire is a dependency-free C++23 CLI and MCP server that analyzes 21 languages and returns relevant symbols, relationships, and tests within a token limit.
* [Phoenix-VAD: semantic end-of-speech detection in streaming](https://ai.exponencialadm.net/okf/en/radar/phoenix-vad-explora-fim-semantico-de-a69180.md) - Published on September 4, 2026, the article presents an LLM-based model for semantically detecting the end of speech in full-duplex voice interactions, using sliding-window training.
* [FRMs apply recurrent flows to structured reasoning](https://ai.exponencialadm.net/okf/en/radar/frms-aplicam-fluxos-recorrentes-ao-raciocinio-e6560a.md) - A post dated September 3, 2026 describes flow-based reasoning models that refine outputs on discrete tasks such as Sudoku.
* [Microsoft introduces the MAI-Transcribe-2 speech model](https://ai.exponencialadm.net/okf/en/radar/microsoft-lanca-o-modelo-de-fala-8f5747.md) - Announced on September 3, 2026, MAI-Transcribe-2 is a transcription model reporting 2.0% AA-WER, about 411× real-time speed, and support for 60 languages.
* [Language models and control of their own attention](https://ai.exponencialadm.net/okf/en/radar/modelos-de-linguagem-e-controle-da-efd6c4.md) - A paper explores whether language models can control their own attention. The available material introduces the topic but does not report methods or results.
* [Launch announces streaming ASR with diarization](https://ai.exponencialadm.net/okf/en/radar/streaming-reune-transcricao-e-diarizacao-db8290.md) - A September 3, 2026 publication announces a streaming speech recognition model combining transcription and diarization. The text provides no performance metrics or details.
* [163 procedures for research agents across 16 fields](https://ai.exponencialadm.net/okf/en/radar/163-procedimentos-para-agentes-de-pesquisa-6678fc.md) - Published on September 2, 2026, the article presents an open library of research procedures and highlights domain-specific choices and caveats.
* [Depth-wise batching for recurrent models](https://ai.exponencialadm.net/okf/en/radar/depth-wise-batching-em-modelos-recorrentes-47fa46.md) - A September 2, 2026 post introduces depth-wise batching: overlapping requests at different recurrent depths to potentially improve utilization with adaptive compute.
* [NoRA changes LoRA initialization](https://ai.exponencialadm.net/okf/en/radar/nora-muda-a-inicializacao-do-lora-f644db.md) - A September 2, 2026 post describes NoRA, which normalizes the columns of LoRA matrix A at initialization, and reports higher average results than several adaptation methods in SFT and RLVR.
* [Photon 2.1 expands streaming speech recognition options](https://ai.exponencialadm.net/okf/en/radar/photon-2-1-amplia-opcoes-de-217f29.md) - The update adds automatic speech recognition for three models and support for the NVIDIA B200 GPU. The company says it outperformed each model’s reference engine in comparable tests.
* [Astra and the challenges of monitoring recurrent reasoning](https://ai.exponencialadm.net/okf/en/radar/profundidade-recorrente-levanta-duvidas-sobre-monitoramento-93d239.md) - A September 2, 2026 publication says Astra uses “recurrent depth,” an approach that researchers reportedly fear could hinder monitoring and affect costs and performance.
* [Zvec team launches local search tool zg](https://ai.exponencialadm.net/okf/en/radar/zg-reune-modos-de-busca-para-3f3ecc.md) - On September 2, 2026, Alibaba's Zvec team said it released zg as open source, a local search tool for developers and AI agents.
* [Checks reduce errors in AI-generated papers](https://ai.exponencialadm.net/okf/en/radar/checagens-contra-logs-reduzem-erros-em-f2c7ba.md) - Published on September 1, 2026, this Radar item summarizes an arXiv paper: checking manuscript claims against execution logs reduced the reported rate of serious hallucinations to 4%.
* [Hy4 Preview reduced to 214 GB with 1.25-bit quantization](https://ai.exponencialadm.net/okf/en/radar/hy4-preview-reduz-modelo-para-214-25ff62.md) - A Tencent publication says Sherry quantization reduces Hy4 Preview from 1.5 TB to 214 GB at 1.25 bits per weight, and describes running the model on GPUs distributed across multiple machines.
* [CERN slides introduce LLM training](https://ai.exponencialadm.net/okf/en/radar/slides-do-cern-apresentam-treinamento-de-3716ca.md) - A post dated August 31, 2026, says slides from a CERN talk on LLM training are available as a PDF and that an online recording is available.
* [Article reviews policy gradients and variants](https://ai.exponencialadm.net/okf/en/radar/derivacao-de-policy-gradients-e-suas-078a85.md) - An article on reinforcement learning foundations derives the policy gradient algorithm and discusses variants, providing review material for engineers.
* [KV Cache: Queries Are Not Kept](https://ai.exponencialadm.net/okf/en/radar/kv-cache-queries-nao-sao-mantidas-935f0d.md) - A post dated August 30, 2026 distinguishes the one-time use of queries from the retention of keys and values in the KV cache during inference.
* [Hy4: calibration-guided low-bit compression](https://ai.exponencialadm.net/okf/en/radar/hy4-compressao-low-bit-guiada-por-878785.md) - A post reports reducing the Hy4 preview from 1.5 TB to 200 GB with per-layer precision settings, and compares results on two benchmarks.
* [Wireshark: decrypting HTTPS traffic with TLS 1.3](https://ai.exponencialadm.net/okf/en/radar/inspecione-https-com-tls-1-3-ad2020.md) - Published on August 29, 2026, the article explains using SSLKEYLOGFILE and Wireshark to inspect HTTP requests in captured HTTPS traffic.
* [Colibri streams MoE experts from disk in pure C](https://ai.exponencialadm.net/okf/en/radar/colibri-transmite-experts-moe-do-disco-7fa2be.md) - The Colibri project says it can run seven families of open MoE models, from 744 billion to 2.8 trillion parameters, by streaming experts from disk on user hardware.
* [Phonon-1 launches for speech recognition](https://ai.exponencialadm.net/okf/en/radar/phonon-1-reconhecimento-de-fala-em-c4b524.md) - A post dated August 28, 2026, describes Phonon-1, a speech recognition model with 782 million parameters, a 415 MB download, and an Apache 2.0 license.
* [Reported 1-bit quantization results for GLM-5.3](https://ai.exponencialadm.net/okf/en/radar/quantizacao-dinamica-de-1-bit-reduz-0d9585.md) - In a post dated August 28, 2026, an engineer reports dynamically quantizing GLM-5.3 to 1 bit: 217 GB versus 1.5 TB in BF16, with about 76% top-1 accuracy.
* [TurnBench measures turn-taking in spoken dialogue](https://ai.exponencialadm.net/okf/en/radar/turnbench-mede-a-alternancia-de-turnos-1ac39e.md) - Published on August 28, 2026, TurnBench includes 30 hours of dyadic speech annotated by three people, a scoring protocol based on conversation analysis, and results for 14 systems.
* [WikiSkill reuses knowledge across agent iterations](https://ai.exponencialadm.net/okf/en/radar/wikiskill-conhecimento-persistente-para-skills-de-bd7137.md) - The approach separates execution traces, a persistent knowledge wiki, and executable skills. Experiences consolidated in the wiki guide later skill updates.
* [Sampling to optimize objectives without RL](https://ai.exponencialadm.net/okf/en/radar/amostragem-para-otimizar-objetivos-sem-rl-5ed12c.md) - Published on August 27, 2026, the post opens a series on alternatives to reinforcement learning for optimizing evaluable objectives. Part I introduces sampling methods.
* [SlimServe fork enables PCIe P2P on GeForce](https://ai.exponencialadm.net/okf/en/radar/fork-do-slimserve-habilita-p2p-via-397e71.md) - A post dated August 27, 2026 reports a graphics-driver change enabling peer-to-peer transfers over PCIe on GeForce cards while optimizing SlimServe. The author has not tested the proposed use on the RTX Pro 6000.
* [JIT-Agent adapts agent harnesses to tasks](https://ai.exponencialadm.net/okf/en/radar/jit-agent-cria-harnesses-de-agentes-579824.md) - An article published August 27, 2026 describes JIT-Agent, which synthesizes agent harnesses on demand and can repair and evolve them.
* [FireRedAudio combines speech tasks in one model](https://ai.exponencialadm.net/okf/en/radar/modelo-de-audio-reune-cinco-tarefas-d5ea1a.md) - A record published on August 27, 2026 describes an audio model combining speech recognition, audio understanding, speech synthesis, and editing.
* [Architecture overview of GLM-5.3-Flash](https://ai.exponencialadm.net/okf/en/radar/arquitetura-hibrida-do-glm-5-3-1b58ac.md) - A publication dated August 26, 2026 identifies Ox Alpha as GLM-5.3-Flash and outlines components of its hybrid architecture.
* [Post describes hybrid attention in GLM-5.3](https://ai.exponencialadm.net/okf/en/radar/atencao-linear-e-esparsa-a-afirmacao-321174.md) - Published on August 26, 2026, the post claims GLM-5.3 alternates linear and sparse attention, and compares this design with other hybrid approaches.
* [Breeze TTS 2 releases open weights](https://ai.exponencialadm.net/okf/en/radar/breeze-tts-2-libera-pesos-abertos-f1a100.md) - A post dated August 26, 2026, says Breeze TTS 2 weights are open and claims the model leads a ranking of TTS provider voices.
* [Practical resources on GRPO and OPD](https://ai.exponencialadm.net/okf/en/radar/grpo-e-opd-recursos-para-post-fed15f.md) - A publication dated August 26, 2026 shares two resources on post-training with reinforcement learning beyond SFT: one on scaling vanilla GRPO and another on OPD.
* [Pocket TTS runs on Android with LiteRT](https://ai.exponencialadm.net/okf/en/radar/pocket-tts-roda-no-android-com-b87143.md) - A publication dated August 26, 2026, reports that the 100-million-parameter Pocket TTS model, with six included voices, runs on-device with LiteRT on current Android devices.
* [Recuris proposes agent memory without changing the model](https://ai.exponencialadm.net/okf/en/radar/recuris-explora-memoria-de-agentes-sem-30ec11.md) - In a publication dated August 26, 2026, Recuris describes two loops for long-horizon agents: working and experiential memories guide tasks, while a fixed meta-agent analyzes execution traces and proposes targeted repairs.
* [iPhone OCR with Apple’s Vision framework](https://ai.exponencialadm.net/okf/en/radar/servidor-de-ocr-para-ios-com-1ee9a0.md) - Published on August 26, 2026, this entry describes an open-source project that provides OCR on an iPhone through an API accessible to devices on the same network.
* [Post highlights efficient designs in Chinese models](https://ai.exponencialadm.net/okf/en/radar/tecnicas-de-eficiencia-em-modelos-chineses-ab4676.md) - A post dated August 26, 2026, says Chinese frontier models share attention, residual-connection, and optimization techniques, but provides no supporting links for its broad claims.
* [Tencent launches the multimodal WeMM-Embedding family](https://ai.exponencialadm.net/okf/en/radar/tencent-lanca-a-familia-multimodal-wemm-c84a3a.md) - Tencent’s WeChat Vision team announced multimodal embedding models for understanding and retrieving text, images, and video. The post also says engineers can evaluate them on multimodal retrieval tasks.
* [Weixin announces multimodal WeMM-Embedding model](https://ai.exponencialadm.net/okf/en/radar/weixin-apresenta-o-embedding-multimodal-wemm-b8476f.md) - Announcement dated August 26, 2026: Weixin says its 9B embedding model supports search and recommendations across text, images, and video, and led two evaluations.
* [Breeze TTS 2: Open Weights and Streaming Speech](https://ai.exponencialadm.net/okf/en/radar/breeze-tts-2-voz-em-streaming-8f70a1.md) - Published on August 25, 2026, the entry describes Breeze TTS 2 as a text-to-speech model supporting 50 languages, speech generation from text prompts, and streaming output.
* [Engram retrieves token n-gram embeddings](https://ai.exponencialadm.net/okf/en/radar/consulta-de-n-gramas-em-arquiteturas-9688a5.md) - Published on August 25, 2026, the text describes a module that hashes the last N tokens to retrieve multi-head embeddings and add them to a layer's hidden state.
* [CTM-AI explores an architecture inspired by a consciousness model](https://ai.exponencialadm.net/okf/en/radar/ctm-ai-explora-consciencia-como-guia-036ce5.md) - An early project applies the Conscious Turing Machine to AI architecture. Its preview describes the proposal but reports no benchmark results.
* [Mac Studio M5 Ultra and RTX PRO 6000 for local AI](https://ai.exponencialadm.net/okf/en/radar/mac-studio-m5-ultra-ou-rtx-cfd014.md) - A comparison published on August 25, 2026 examines price, memory, prefill performance, CUDA support, and noise across two hardware options for local AI development.
* [Mol-JEPA combines modalities for molecular training](https://ai.exponencialadm.net/okf/en/radar/mol-jepa-explora-aprendizado-molecular-multimodal-42effc.md) - Published on August 25, 2026, the post describes a molecular model trained on multiple modalities and reports out-of-distribution gains on small ADME datasets.
* [Notebooks for exploring agent design patterns](https://ai.exponencialadm.net/okf/en/radar/notebooks-para-explorar-padroes-de-design-316ba4.md) - A guide provides code notebooks for hands-on exploration of prompt chaining, routing, reflection, tool use, planning, and multi-agent systems.
* [ComfyUI plugin adds VoxCPM2 and LoRA](https://ai.exponencialadm.net/okf/en/radar/plugin-do-comfyui-oferece-inferencia-e-3b01fa.md) - Published on August 25, 2026, ComfyUI_VoxCPM_SM offers VoxCPM2 inference and LoRA fine-tuning with custom voice data, with stated support for low VRAM and quantized models.
* [Kyutai opens the Pocket TTS training stack](https://ai.exponencialadm.net/okf/en/radar/pocket-tts-abre-seu-stack-de-8fd340.md) - Published on August 25, 2026, the stack includes a data pipeline, training recipes, and evaluations for a text-to-speech model designed to run on CPU.
* [Scroll treats agent context as an executable environment](https://ai.exponencialadm.net/okf/en/radar/scroll-trata-o-contexto-de-agentes-a6d8ca.md) - Published on August 25, 2026, the item presents Scroll, which combines an append-only event log with a persistent, isolated Python kernel to manage long-term agent context.
* [WeMM-Embedding-9B maps modalities into a shared space](https://ai.exponencialadm.net/okf/en/radar/wemm-embedding-9b-aproxima-texto-e-b99ac2.md) - Published on August 25, 2026, the post describes a model that represents text, images, videos, and visual documents in a shared space, and claims state-of-the-art results on MMEB-v2 and MMEB-v3.
* [Crescent Island listed with 32 Xe3P cores](https://ai.exponencialadm.net/okf/en/radar/crescent-island-listado-com-32-nucleos-501869.md) - A publication dated August 24, 2026 lists Intel Crescent Island with 32 Xe3P cores, 480 GB of LP5X memory, and 350 W power consumption.
* [Guide connects RL with LLM training methods](https://ai.exponencialadm.net/okf/en/radar/guia-de-reinforcement-learning-para-llms-670a11.md) - Published on August 24, 2026, the guide covers reinforcement learning fundamentals and formulations used in training language models.
* [Xring O3 reportedly tops 15,000 in Geekbench 6](https://ai.exponencialadm.net/okf/en/radar/xring-o3-teria-superado-15-mil-2d61e7.md) - A publication dated August 24, 2026 reports 3,945 single-core points and 15,221 multi-core points for Xiaomi's Xring O3 chip; its multi-core score reportedly surpasses Apple's M4.
* [AI R&D agents rarely revisit training strategy](https://ai.exponencialadm.net/okf/en/radar/agentes-de-p-d-raramente-revisam-c83b9b.md) - A study of 1,338 post-training trajectories reports that the overall strategy changed in only 74 of 3,557 adjacent experiments.
* [Pandora’s Router weighs cost and quality in model selection](https://ai.exponencialadm.net/okf/en/radar/pandora-s-router-pondera-custo-e-e2f4b2.md) - Published on August 22, 2026, the paper frames routing as a decision about when better estimates are worth paying for. On MATH, RAG, and EmbedLLM, it achieved or tied for the lowest average combined inspection cost and routing regret.
* [Active inference for context acquisition by agents](https://ai.exponencialadm.net/okf/en/radar/quando-agentes-devem-buscar-mais-contexto-4adb88.md) - An article published on August 22, 2026 models how interactive agents choose between acting on assumptions and seeking context, accounting for the costs of those choices.
* [AgentSysBench examines serving for agents](https://ai.exponencialadm.net/okf/en/radar/agentsysbench-investiga-gargalos-em-sistemas-para-3af61e.md) - Published on August 21, 2026, the paper presents a benchmark covering 10 agentic applications and reports that model inference is often not the main serving bottleneck.
* [FIBER separates GPU registers from execution](https://ai.exponencialadm.net/okf/en/radar/fiber-separa-a-propriedade-de-registradores-a5e656.md) - Published on August 21, 2026, the article presents FIBER, a GPU architecture that decouples register ownership from parallel execution instances. It describes ISA extensions, lightweight microarchitectural changes, and a programming model compatible with CUDA.
* [Marin details materials on model training](https://ai.exponencialadm.net/okf/en/radar/o-que-os-artefatos-de-treinamento-b1d14b.md) - A publication dated August 21, 2026 describes visualizations of pretraining mixtures by domain, sampled documents, live training loss, configurations, and scaling laws.
* [Slides on post-training and real-world adaptation](https://ai.exponencialadm.net/okf/en/radar/post-training-e-adaptacao-de-modelos-90dae7.md) - Introductory lecture slides for economists cover post-training methods and research questions about agents in changing environments.
* [A year of LLM serving workloads in production](https://ai.exponencialadm.net/okf/en/radar/um-ano-de-serving-de-llm-f5c8fb.md) - A paper examining a year of production LLM traffic reports workload changes, short-lived prefix reuse, and tradeoffs between KV-cache locality and load balancing.
* [A diary documents pretraining a MoE model](https://ai.exponencialadm.net/okf/en/radar/diario-de-pre-treinamento-de-um-f1f77c.md) - Published on August 20, 2026, the free book recounts an attempt to pretrain a 1.02-billion-parameter MoE model on an H200.
* [dstack announces Presets for inference optimization](https://ai.exponencialadm.net/okf/en/radar/dstack-apresenta-presets-para-otimizacao-portatil-a4e38f.md) - In an announcement dated August 20, 2026, dstack presents Presets, an open-source toolkit that uses agents to simplify inference optimization and provides a portable format for its results.
* [CuTe: layout representation and algebra for GPUs](https://ai.exponencialadm.net/okf/en/radar/cute-representacao-e-algebra-de-layouts-3b18d4.md) - An article covers CuTe layout representation and algebra, a relevant topic for engineers working with layout abstractions in GPU programming.
* [Movfuscator compiles C using only MOV](https://ai.exponencialadm.net/okf/en/radar/movfuscator-compila-c-usando-apenas-mov-bed72b.md) - Movfuscator turns C into Turing-complete assembly using only the MOV instruction, showing how restricting a compiler’s output can complicate code analysis.
* [Report points to a shell prompt mismatch in DSH](https://ai.exponencialadm.net/okf/en/radar/prompt-divergente-pode-atrasar-comandos-no-5be125.md) - A post dated August 17, 2026 reports that different prompts in two DSH tools triggered a 3.5-second timeout; after a constant was changed, one command reportedly fell from 3,600 ms to 158 ms.
* [Study reports code overlap in filtering systems](https://ai.exponencialadm.net/okf/en/radar/estudo-aponta-codigo-sobreposto-em-sistemas-152ae0.md) - Authors of a USENIX Security paper report overlap between leaked Geedge Networks source code and China’s Great Firewall, including DNS and RST injection behavior.
* [RL replay buffer with top-k in SQL](https://ai.exponencialadm.net/okf/en/radar/replay-buffer-de-rl-com-amostragem-cde5d4.md) - A post dated August 16, 2026 describes weighted top-k SQL sampling for a reinforcement-learning replay buffer and reports selecting 8,192 rows from 50 million in about 1.1 seconds on a laptop.
* [datatrove 0.10.0 expands pipeline execution options](https://ai.exponencialadm.net/okf/en/radar/datatrove-0-10-0-leva-pipelines-df6a9d.md) - Version 0.10.0 introduces an executor for pipelines on a jobs service, support for storage buckets as data folders, and preservation of reasoning outputs in inference results.
* [Tracing an LDG Instruction on the RTX 4090](https://ai.exponencialadm.net/okf/en/radar/o-caminho-de-uma-leitura-ldg-dddba8.md) - An article published on August 13, 2026, follows an LDG.E SASS instruction through the RTX 4090’s hardware units and memory hierarchy, including reverse engineering of undocumented details.
* [Swizzling and bank conflicts in CUDA shared memory](https://ai.exponencialadm.net/okf/en/radar/swizzling-e-conflitos-de-bancos-na-d13872.md) - A technical article published on August 13, 2026, discusses using swizzling to address bank conflicts in CUDA shared memory, a memory-access issue that can affect CUDA kernel performance.
* [IO Factory simulates AI-driven influence campaigns](https://ai.exponencialadm.net/okf/en/radar/io-factory-simula-campanhas-de-influencia-58dbe0.md) - An article dated August 12, 2026 presents an AI framework for simulating coordinated, adaptive influence campaigns.
* [A CUDA challenge path for GPU programming practice](https://ai.exponencialadm.net/okf/en/radar/uma-trilha-pratica-para-desafios-de-e1df81.md) - Published on August 12, 2026, the bookmark recommends LeetGPU challenges for CUDA practice and outlines a progression from naive SIMD kernels to more advanced techniques and tools.
* [Inference hooks control prompts](https://ai.exponencialadm.net/okf/en/radar/hooks-de-inferencia-controlam-prompts-antes-739a55.md) - Published on August 11, 2026, the text describes hooks that send prompts for security checks before inference, allowing an organization to permit or deny each request.
* [ExploitGym evaluates AI agents on vulnerability exploitation](https://ai.exponencialadm.net/okf/en/radar/exploitgym-avalia-agentes-de-ia-em-d8411c.md) - Research published on August 10, 2026 describes an evaluation of AI agents’ ability to turn vulnerabilities into concrete impacts, such as unauthorized file access or code execution.
* [ResidencyRL trains clinical agents in simulated consultations](https://ai.exponencialadm.net/okf/en/radar/residencyrl-treina-agentes-clinicos-em-consultas-b4d05a.md) - Research explores reinforcement learning in simulated clinical consultations, involving patient history, diagnostic hypotheses, and decisions under uncertainty.
* [Benchmark compares seven self-hosted memory providers](https://ai.exponencialadm.net/okf/en/radar/benchmark-compara-sete-provedores-de-memoria-8932a1.md) - Published on August 9, 2026, this community benchmark compares seven self-hosted memory providers for Hermes Agent, using 71,060 conversation turns and 3,750 questions about facts that change over time.
* [EnvACE proposes world rehearsals for agent training](https://ai.exponencialadm.net/okf/en/radar/envace-treina-agentes-com-ensaios-internos-c3fdac.md) - Published on August 9, 2026, the entry describes EnvACE, a reinforcement-learning method that alternates tool-call generation and internal world rehearsals during training.
* [obfus.h: compile-time C obfuscation](https://ai.exponencialadm.net/okf/en/radar/obfus-h-ofuscacao-de-c-em-e30f71.md) - Published on August 9, 2026, obfus.h is described as a macro header for compile-time C obfuscation on Windows x86/x64 using tcc.
* [Reverse-proxy phishing can capture sessions after MFA](https://ai.exponencialadm.net/okf/en/radar/phishing-com-proxy-pode-roubar-sessoes-9fa0bb.md) - A post published on August 9, 2026 describes kits that relay login and MFA interactions with a real service to intercept the resulting session cookie. It cites Evilginx, Modlishka, and Muraena.
* [Survey maps self-evolving coding agents](https://ai.exponencialadm.net/okf/en/radar/survey-mapeia-coding-agents-que-evoluem-851142.md) - A survey analyzes 108 systems that learn from software development interactions and use those experiences to change memory, workflows, tools, or model parameters.
* [Argus reports 78% on SWE-Bench Pro with a verified runtime](https://ai.exponencialadm.net/okf/en/radar/argus-verifica-mudancas-em-agentes-de-88153d.md) - Published on August 8, 2026, the account describes Argus, a runtime for long-horizon agentic tasks, and reports 78% on SWE-Bench Pro versus 59% for Direct Copilot.
* [libuv guide: event loops and asynchronous code](https://ai.exponencialadm.net/okf/en/radar/guia-do-libuv-event-loops-e-01c785.md) - Published on August 8, 2026, the guide introduces the libuv event loop, notes conceptual similarities with other systems, and explains its usefulness for analyzing asynchronous code in C and C++.
* [Semantica proposes graph-native infrastructure for agents](https://ai.exponencialadm.net/okf/en/radar/semantica-leva-grafos-de-conhecimento-a-597d60.md) - Published on August 8, 2026, the item presents Semantica, an open-source project that turns data into knowledge graphs and records agent decisions for later tracing.
* [Practical attacks on 4G/LTE privacy and availability](https://ai.exponencialadm.net/okf/en/radar/ataques-praticos-a-privacidade-e-disponibilidade-b641ce.md) - A 2015 article examines practical attacks on privacy and availability in 4G/LTE mobile communications, a topic relevant to engineers assessing mobile network risks.
* [Ostris AI Toolkit Adds LoRA Training for MiniMax H3](https://ai.exponencialadm.net/okf/en/radar/ostris-ai-toolkit-passa-a-treinar-23f31d.md) - The toolkit now supports LoRA training for MiniMax H3, initially limited to text-to-video and image-to-video. The announcement says it uses quantized Comfy weights.
* [Searchable reference for MSLK kernels and APIs](https://ai.exponencialadm.net/okf/en/radar/referencia-pesquisavel-de-kernels-e-apis-6e649f.md) - Published on August 2, 2026, the bookmark describes a searchable MSLK reference covering kernels, APIs, and developer utilities, created with agents, according to its author.
* [Transformers, attention and computational costs](https://ai.exponencialadm.net/okf/en/radar/transformers-e-atencao-conceitos-e-custos-2c4800.md) - An article published on August 2, 2026 explains the applied mathematics behind Transformers and attention mechanisms, including methods aimed at reducing computation and memory costs.
* [Study links safety tuning to mind attribution](https://ai.exponencialadm.net/okf/en/radar/safety-tuning-pode-alterar-atribuicoes-de-c64d69.md) - A publication dated August 1, 2026 reports that safety adjustments about consciousness also reduced mind attribution to other entities in three models; mechanistic interventions reportedly restored these tendencies.
* [PR configures checkpoint 0731 on two DGX Sparks](https://ai.exponencialadm.net/okf/en/radar/pr-configura-checkpoint-0731-em-dois-3c0e4e.md) - A change updates the default checkpoint to model version 0731 and installs its encoder on both ranks in a two-DGX-Spark inference setup.
* [GLM-5.2 EXL3 serving profile on four GPUs](https://ai.exponencialadm.net/okf/en/radar/perfil-de-serving-do-glm-5-5c22cb.md) - A GitHub repository presents a reproducible GLM-5.2 EXL3 MTP3 serving configuration and a thermal profile for four RTX PRO 6000 Blackwell GPUs.
* [Buzz publishes NIP-AE memory specification](https://ai.exponencialadm.net/okf/en/radar/buzz-publica-especificacao-formal-para-seu-0fcdf3.md) - Published on July 28, 2026, the NIP-AE document specifies Buzz's memory feature. The project says its formal specifications are intended to support implementation and interoperability.
* [Catalog collects silent failures in LLM serving](https://ai.exponencialadm.net/okf/en/radar/falhas-silenciosas-no-serving-de-llms-96e45d.md) - Published on July 28, 2026, a GitHub repository collects serving failures that can go unnoticed and describes symptoms, mechanisms, checks, and fixes.
* [Chronological readings on Looped Transformers](https://ai.exponencialadm.net/okf/en/radar/leituras-cronologicas-sobre-looped-transformers-1c4872.md) - A reading list covers papers on Universal Transformers, Relaxed Recursive Transformers, Mixture-of-Recursions, and DeepLoop to compare recurrent computation and parameter sharing.
* [Radar: OCR test announced](https://ai.exponencialadm.net/okf/en/radar/como-avaliar-benchmarks-de-ocr-8a4965.md) - A July 25, 2026 post reports adding Baidu Unlimited OCR to mlx-vlm and plans to evaluate it with an OCR benchmark.
* [Device-code phishing is not Wi-Fi-specific](https://ai.exponencialadm.net/okf/en/radar/phishing-por-device-code-nao-depende-04467a.md) - Published on July 25, 2026, the post warns that device-code authentication phishing can occur over Ethernet or the internet, not only on Wi-Fi networks.
* [A first-principles guide to JEPA's SIGReg](https://ai.exponencialadm.net/okf/en/radar/um-guia-de-primeiros-principios-sobre-e19ecd.md) - A publication dated July 24, 2026 presents a guide to mathematical concepts and a functional training loop related to JEPA's SIGReg.
* [C++ reflection and access to private data](https://ai.exponencialadm.net/okf/en/radar/reflection-em-c-modifica-campo-privado-ae0002.md) - Published on July 23, 2026, the example uses C++ reflection with an unchecked access context to increment Foo's non-static data members, including the private field secret, and print the result.
* [The Stack v3 gathers 5 trillion code tokens](https://ai.exponencialadm.net/okf/en/radar/the-stack-v3-reune-5-trilhoes-e02e7c.md) - Published on July 23, 2026, the announcement describes a dataset with about 5 trillion tokens across 713 languages and 224 million repositories, collected through August 2025.
* [The Stack v3 brings 5 trillion code tokens](https://ai.exponencialadm.net/okf/en/radar/the-stack-v3-reune-5t-tokens-97153d.md) - Published on July 23, 2026, the announcement describes an open dataset with about 5 trillion tokens of deduplicated and filtered source code across 713 languages. The post says code under restrictive licenses is excluded.
* [Awesome C++ gathers resources for engineers](https://ai.exponencialadm.net/okf/en/radar/awesome-c-um-ponto-de-partida-be5ac5.md) - Published on July 21, 2026, Awesome C++ is a curated list of C++ frameworks, libraries, resources, and other items. It is presented as a starting point for engineers looking for tools and materials related to the language.
* [ESP32 blocks 537,000 domains with about 50 KB of RAM](https://ai.exponencialadm.net/okf/en/radar/esp32-bloqueia-537-mil-dominios-usando-a3f497.md) - Published on July 20, 2026, the post describes ESP32 firmware that keeps a blocklist in flash and blocks requests in about 10 milliseconds, according to the post.
* [Packed FlashAttention speeds a Transformer vocoder](https://ai.exponencialadm.net/okf/en/radar/packed-flashattention-acelera-vocoder-transformer-8937f8.md) - A publication dated July 20, 2026 reports a 48.64% increase in QPS after optimizing waveform reconstruction in a Transformer vocoder. The available excerpt does not state the size of the latency reduction.
* [Survey maps self-improvement in AI agents](https://ai.exponencialadm.net/okf/en/radar/como-agentes-de-ia-podem-se-e5d9c3.md) - A survey of 239 papers examines how AI agents may be improved through updates to the model or surrounding components such as prompts, memory, and tools.
* [Benchmark compares models for vulnerability discovery](https://ai.exponencialadm.net/okf/en/radar/benchmark-compara-modelos-na-busca-por-346e45.md) - Published on July 18, 2026, the private benchmark uses the open-source DeepSec harness to compare models by recall, precision, and cost in an unnamed open-core application.
* [How Blackwell moves matmul accumulation into Tensor Memory](https://ai.exponencialadm.net/okf/en/radar/blackwell-leva-acumuladores-de-matmul-a-643953.md) - Radar: technical post describing Tensor Memory in Blackwell as a separate 256 KB per-SM space for Tensor Core results, and stating, according to the author, that this reduces matmul issue from 128 threads to one.
* [DeepLoop examines depth scaling in looped Transformers](https://ai.exponencialadm.net/okf/en/radar/deeploop-analisa-profundidade-em-transformers-com-8b72f4.md) - Published July 17, 2026, the article examines how looped Transformers reuse blocks to increase unfolded depth without adding stored parameters.
* [Dell Precision T7910 for multi-GPU computing](https://ai.exponencialadm.net/okf/en/radar/dell-precision-t7910-para-avaliar-computacao-4f0977.md) - A publication dated July 17, 2026 describes a Dell Precision T7910 configuration and highlights PCIe slots, ECC memory, and power limits as factors in assessing a used multi-GPU system.
* [Diffusing Blame trains networks under Dale’s principle](https://ai.exponencialadm.net/okf/en/radar/diffusing-blame-explora-treino-sem-backpropagation-5b0e11.md) - A July 17, 2026 post describes routing error signals to train networks with dedicated excitatory and inhibitory neurons, without backpropagation.
* [HarnessBank proposes evolving agent harnesses](https://ai.exponencialadm.net/okf/en/radar/harnessbank-evolucao-de-harnesses-de-agentes-4a45ba.md) - Published on July 17, 2026, the article presents a method for evolving agent harnesses—prompts, tools, and control loops—using semantic search and stepwise verification.
* [KTransformers explores LLMs with limited VRAM](https://ai.exponencialadm.net/okf/en/radar/ktransformers-explora-llms-com-vram-limitada-f7c2d8.md) - Published on July 17, 2026, the report describes an open framework for heterogeneous inference and fine-tuning of language models.
* [MemoHarness adapts harnesses with prior executions](https://ai.exponencialadm.net/okf/en/radar/memoharness-adapta-harnesses-de-agentes-pelo-cd52a5.md) - Published on July 17, 2026, MemoHarness adapts six dimensions of agent harnesses and retrieves lessons from similar past cases. On a shell-agent benchmark, it scored 0.806, versus 0.722 for the strongest fixed-harness baseline.
* [uv checks OSV before installing packages](https://ai.exponencialadm.net/okf/en/radar/uv-pode-checar-malware-antes-de-dfae34.md) - Published on July 17, 2026, the notice says that setting UV_MALWARE_CHECK=1 enables an OSV database check before installing packages from remote registries.
* [Survey examines self-improvement in modern agents](https://ai.exponencialadm.net/okf/en/radar/como-analisar-o-autoaperfeicoamento-de-agentes-48320f.md) - Published on July 16, 2026, the survey defines modern agents and describes self-improvement as adaptation that can accumulate capability gains from experience.
* [Linux x86_64 Assembly Guide Collected in a PDF](https://ai.exponencialadm.net/okf/en/radar/guia-de-assembly-para-linux-x86-0e78a3.md) - According to its author, the 0xAX Linux x86_64 assembly series is now available as a single PDF. The announcement points to GitHub release v1.0.0.
* [Study examines RL across LLM layers](https://ai.exponencialadm.net/okf/en/radar/como-o-reinforcement-learning-se-distribui-c736cb.md) - A preprint dated July 15, 2026 examines how reinforcement learning adaptation is distributed across transformer layers during LLM post-training and questions whether all layers contribute equally.
* [Standalone C headers for 8-bit emulators](https://ai.exponencialadm.net/okf/en/radar/headers-c-standalone-para-emuladores-de-4fa2f5.md) - Published on July 15, 2026, the floooh/chips repository contains emulators of chips and 8-bit systems implemented as standalone C headers, providing engineers with concrete code to study emulation in C.
* [Post reports an eight-day autoresearch experiment](https://ai.exponencialadm.net/okf/en/radar/experimento-de-autoresearch-relata-oito-dias-8fb40f.md) - Published on July 14, 2026, the post claims an autoresearch agent surpassed a manually tuned harness on benchmarks not used in training, after eight days of improvement.
* [GPU-Tile-Sim: GPU simulation for LLM co-design](https://ai.exponencialadm.net/okf/en/radar/gpu-tile-sim-propoe-simulacao-de-215492.md) - Published on July 14, 2026, this entry presents GPU-Tile-Sim (GTSim), a tile-based GPU simulation framework proposed to support hardware and software co-design for LLMs.
* [WANDR: an environment for evaluating agentic search](https://ai.exponencialadm.net/okf/en/radar/wandr-ambiente-para-avaliar-busca-agentic-5ec5ef.md) - On July 14, 2026, Perplexity said it was making available an evaluation and reinforcement-learning environment for agentic search, synthesized from production traces with weak human supervision.
* [Harness and costs of corporate agents](https://ai.exponencialadm.net/okf/en/radar/como-o-harness-afeta-custos-de-5bd751.md) - A July 13, 2026 publication reports that changing agent orchestration reduced cost and latency while maintaining equivalent quality across 22 tasks and six models.
* [StreamDQ brings dequantization closer to HBM memory](https://ai.exponencialadm.net/okf/en/radar/streamdq-aproxima-a-desquantizacao-da-memoria-a4d0e0.md) - Published on July 13, 2026, the summary presents StreamDQ, an architecture that dequantizes LLM weights in real time on the HBM base die while preserving conventional GPU load semantics.
* [Reference harness for security scanning and remediation](https://ai.exponencialadm.net/okf/en/radar/harness-de-referencia-para-analise-e-4fe259.md) - Published on July 12, 2026, the repository described offers skills for threat modeling, scanning, triage, and patching, plus a customizable harness for autonomous scanning.
* [REINFORCE: an estimator for policy gradients](https://ai.exponencialadm.net/okf/en/radar/policy-gradients-entendendo-o-estimador-reinforce-32bbdc.md) - Published on July 12, 2026, the first post in a series derives the REINFORCE estimator as an unbiased policy-gradient method and analyzes how variance grows with trajectory length.
* [Prompt injection in binaries analyzed with AI](https://ai.exponencialadm.net/okf/en/radar/strings-em-binarios-podem-influenciar-agentes-f0c173.md) - A July 12, 2026 bookmark describes Naval Postgraduate School research on prompt-injection strings in C binaries and their influence on LLM reverse-engineering agents in Ghidra.
* [Vulkan documentation explains compute shaders and GPUs](https://ai.exponencialadm.net/okf/en/radar/vulkan-compute-shaders-e-arquitetura-de-761092.md) - A Vulkan tutorial introduces compute shaders, GPU architecture and memory models. Published on July 12, 2026, it notes that much of the discussion also applies to DirectX 12.
* [OpenResearch reports automated paper reproduction](https://ai.exponencialadm.net/okf/en/radar/agentes-reproduzem-pesquisas-e-provisionam-gpus-af399e.md) - Published on July 11, 2026, the entry describes a tool that uses agents to reproduce research results and provision cloud GPUs. The author says it reproduced their paper end to end.
* [Colibri loads MoE experts on demand](https://ai.exponencialadm.net/okf/en/radar/colibri-carrega-experts-de-moe-sob-aad4a2.md) - Published on July 11, 2026, Colibri is described as a dependency-free C inference engine that loads MoE model experts from disk as needed. The post says it runs a 744-billion-parameter model on a laptop with 25 GB of RAM.
* [A hypothesis on how RL expands SFT](https://ai.exponencialadm.net/okf/en/radar/como-pensar-na-relacao-entre-rl-34f00d.md) - Published on July 11, 2026, the post presents a hypothesis: reward-guided variation during RL may separate useful, intertwined components in SFT solutions into reusable skills and routing rules.
* [Verifying code-agent rewards](https://ai.exponencialadm.net/okf/en/radar/como-verificar-recompensas-de-agentes-de-f9c119.md) - Published on July 11, 2026, the paper examines four designs for verifying code-agent rewards and finds no single solution.
* [Memory layout of multidimensional arrays](https://ai.exponencialadm.net/okf/en/radar/layout-de-memoria-e-desempenho-de-9e70a1.md) - Eli Bendersky’s article examines the memory layout of multidimensional arrays and includes benchmark code comparing the performance of different access patterns.
* [MiniCPM-SALA combines sparse and linear attention](https://ai.exponencialadm.net/okf/en/radar/minicpm-sala-combina-atencao-esparsa-e-02bc05.md) - Published on July 11, 2026, the report presents a 9-billion-parameter hybrid architecture for long contexts and reports good benchmark results, with scalable inference up to 1 million tokens.
* [What happens when a CUDA kernel runs](https://ai.exponencialadm.net/okf/en/radar/o-que-acontece-ao-executar-um-bc9119.md) - An introduction to the CPU–GPU communication required to launch a CUDA kernel and the work involved in that operation.
* [RLSD adjusts RLVR updates per token](https://ai.exponencialadm.net/okf/en/radar/rlsd-ajusta-o-aprendizado-por-token-ea0fa0.md) - A paper published on July 11, 2026 combines environmental rewards with self-distillation and reports higher average accuracy on five multimodal benchmarks.
* [SDM proposes sparse memory for RNNs](https://ai.exponencialadm.net/okf/en/radar/sparse-delta-memory-leva-enderecamento-esparso-879abc.md) - Published on July 11, 2026, the post describes Sparse Delta Memory: an explicit memory with sparse addressing used alongside an RNN that, according to the post, expands state capacity without raising compute per token.
* [ThetaEvolve explores test-time learning](https://ai.exponencialadm.net/okf/en/radar/thetaevolve-explora-aprendizado-em-tempo-de-5693fc.md) - Published on July 11, 2026, ThetaEvolve is an open-source framework for test-time learning on open-ended optimization problems. Its description outlines components, without reporting quantitative results.
* [AREAL2.0 proposes an architecture for agents with continuous learning](https://ai.exponencialadm.net/okf/en/radar/areal2-0-conecta-agentes-a-aprendizado-16ba83.md) - Published on July 10, 2026, the article presents an architecture for continuous learning by agents on real workloads, with step-by-step signals, task-to-training-data conversion, and automatic updates.
* [Triton Paged Attention Kernel Attributed to Grok](https://ai.exponencialadm.net/okf/en/radar/kernel-de-paged-attention-em-triton-ac90bf.md) - July 10, 2026 archive entry: a post shares a Triton paged attention kernel attributed to Grok and a solution file hosted on KernelBench.
* [KV cache quantization: keys and values](https://ai.exponencialadm.net/okf/en/radar/kv-cache-mais-bits-para-keys-207c86.md) - An ACL 2026 article reports that using 4 bits for keys and 2 for values may work well for KV cache compression; reversing those bit widths substantially degrades accuracy.
* [LED explores intermediate layers to broaden sampling](https://ai.exponencialadm.net/okf/en/radar/led-explora-camadas-intermediarias-no-decoding-3b7fb2.md) - Published on July 10, 2026, the report presents Latent Exploration Decoding (LED), a method requiring no additional training that uses intermediate-layer distributions to improve sampling in reasoning models.
* [Proactive memory for long-horizon agents](https://ai.exponencialadm.net/okf/en/radar/memoria-proativa-para-agentes-de-horizonte-3ab3a1.md) - A paper dated July 10, 2026 presents a separate memory agent that maintains structured information and decides when to offer reminders to an action agent.
* [Can metacognition lower the cost of LLM reasoning?](https://ai.exponencialadm.net/okf/en/radar/metacognicao-pode-reduzir-o-custo-do-82b993.md) - An article investigates whether models can balance accuracy, context size, compute cost, and latency without relying on long reasoning chains.
* [QuadRF tracks drones and detects Wi-Fi](https://ai.exponencialadm.net/okf/en/radar/quadrf-combina-raspberry-pi-5-e-5b2a34.md) - Published on July 10, 2026, Jeff Geerling’s account presents QuadRF, a phased-array radio built with a Raspberry Pi 5 and an FPGA board. It uses signal processing and beamforming to track drones and detect Wi-Fi through walls.
* [Hierarchical search agents: delegation and execution](https://ai.exponencialadm.net/okf/en/radar/como-distribuir-capacidade-em-agentes-de-9ec896.md) - An article published on July 9, 2026 examines how to distribute model capacity across delegation and execution roles in hierarchical search agents. The post identifies decomposition capacity as the main performance bottleneck.
* [Two-core GPU implemented on a US$30 FPGA](https://ai.exponencialadm.net/okf/en/radar/gpu-de-dois-nucleos-em-fpga-d2a9a4.md) - Published on July 9, 2026, the project describes a two-core GPU on a US$30 FPGA, with a 16-bit ISA, assembler, J++ language, compiler, and AI inference at 81 MHz.
* [Study compares costs and quality in agent orchestration](https://ai.exponencialadm.net/okf/en/radar/orquestracao-pode-mudar-custo-e-qualidade-f62b17.md) - Published July 9, 2026, the study evaluates 22 tasks across six models, changing only the orchestration layer. It reports lower costs, token use, and runtime, with equivalent task-completion quality.
* [Asynchronous RL rollouts for agentic training](https://ai.exponencialadm.net/okf/en/radar/rl-assincrono-rollouts-e-estabilidade-no-129fc7.md) - Published on July 9, 2026, the article discusses asynchronous RL for language-model post-training, with updates as rollouts arrive. It cites clipping adjustments in GLM 5.2 PPO and also points to VAPO.
* [Inference caching: evaluate performance and isolation](https://ai.exponencialadm.net/okf/en/radar/cache-de-inferencia-desempenho-e-isolamento-cc1af5.md) - Efficient memory can improve service capacity. The design must also prevent one session's context from appearing in another session.
* [CMDR assesses multimodal document retrieval across pages](https://ai.exponencialadm.net/okf/en/radar/cmdr-avalia-busca-multimodal-entre-paginas-0f2155.md) - The article presents CMDR and CMDR-Bench for retrieving relevant pages from multimodal documents when queries depend on context spanning multiple pages.
* [Diffusion language models: techniques in focus](https://ai.exponencialadm.net/okf/en/radar/como-estudar-modelos-de-linguagem-por-e3a4bd.md) - A Kuleshov Group post presents research on diffusion language models, covering masked diffusion, iterative refinement, post-training, and variable-length generation.
* [From nvcc to warps: tracing a CUDA kernel](https://ai.exponencialadm.net/okf/en/radar/da-compilacao-cuda-a-execucao-por-f8c813.md) - Published on July 8, 2026, the article follows a vector-sum kernel from compilation with nvcc to execution by GPU warps.
* [On-policy distillation routes supervision per token](https://ai.exponencialadm.net/okf/en/radar/destilacao-on-policy-roteia-supervisao-por-978145.md) - A paper describes a method that dynamically selects, token by token, between supervision based on a teacher and on the model itself.
* [ColBERT: lightweight vector compression fine-tuning](https://ai.exponencialadm.net/okf/en/radar/fine-tuning-leve-comprime-vetores-em-27c92e.md) - A paper published on July 8, 2026 presents pooling-aware fine-tuning and k-means to reduce token vectors in ColBERT documents, claiming to preserve accuracy.
* [NapMem proposes active navigation of memory](https://ai.exponencialadm.net/okf/en/radar/napmem-propoe-navegar-pela-memoria-conversacional-e738ba.md) - An article published on July 8, 2026 presents NapMem, a framework that organizes a user's history in a linked memory pyramid with multiple granularities.
* [NapMem explores long-term memory for agents](https://ai.exponencialadm.net/okf/en/radar/napmem-transforma-memoria-de-longo-prazo-694ece.md) - Published on July 8, 2026, NapMem organizes user history into memory layers and trains an agent to choose which layer to inspect. The proposal contrasts with relying only on evidence preselected by a retriever.
* [What theory reveals about MaxSim](https://ai.exponencialadm.net/okf/en/radar/o-que-a-teoria-revela-sobre-7282dc.md) - Published on July 8, 2026, the article examines MaxSim’s representational power and reports that it can exactly reproduce inner products between nonnegative k-sparse vectors.
* [Qwen 3.5 and Nemotron 3 use different hybrid architectures](https://ai.exponencialadm.net/okf/en/radar/qwen-3-5-e-nemotron-3-c9a68f.md) - Published on July 8, 2026, the post compares two hybrid architectural approaches and relates them to the convergent evolution of linear attention and state-space models.
* [Ridgeline: a roofline profiler for LLM inference](https://ai.exponencialadm.net/okf/en/radar/ridgeline-profiler-roofline-para-inferencia-de-b7fcf5.md) - Ridgeline is a small profiler built with PyTorch to help engineers analyze LLM inference workloads using a roofline model.
* [SearchEyes trains search agents in simulated worlds](https://ai.exponencialadm.net/okf/en/radar/searcheyes-treina-agentes-de-busca-em-6a4039.md) - Published on July 8, 2026, the summary presents SearchEyes, an approach that uses a typed knowledge graph as the basis for a simulated search world to train multimodal agents.
* [BrowserBC turns browsing traces into skills for agents](https://ai.exponencialadm.net/okf/en/radar/browserbc-transforma-traces-de-navegacao-em-728244.md) - The approach converts human browsing traces into natural-language skill cards that agents can retrieve and combine to navigate unfamiliar websites.
* [Relevance-Based Embeddings for Candidate Retrieval](https://ai.exponencialadm.net/okf/en/radar/embeddings-orientados-a-relevancia-para-recuperar-9a4afd.md) - A paper describes query and item embeddings for efficiently retrieving candidates when relevance computation is expensive. The reported approach uses relevance to selected support items or queries.
* [Harness engineering and AI self-improvement](https://ai.exponencialadm.net/okf/en/radar/harness-engineering-e-autoaperfeicoamento-de-ia-3400c0.md) - Published on July 7, 2026, Lilian Weng’s article examines recursive self-improvement and presents designing and optimizing a model harness as a practical starting point.
* [HiLS learns chunk selection for sparse attention](https://ai.exponencialadm.net/okf/en/radar/hils-aprende-a-selecionar-chunks-em-bb33a9.md) - Published on July 7, 2026, the material presents Hierarchical Landmark Sparse (HiLS) Attention, a method that learns chunk selection end to end using the language-modeling loss.
* [JetSpec speeds up speculative decoding with parallel trees](https://ai.exponencialadm.net/okf/en/radar/jetspec-explora-arvores-paralelas-na-decodificacao-54588d.md) - A post published on July 7, 2026 describes JetSpec, which uses a single forward pass with causal conditioning to create coherent draft trees. It reports speedups of up to 9.64× on MATH-500 and 4.58× in open-ended conversations.
* [Live stream presents new chapters on CUDA programming](https://ai.exponencialadm.net/okf/en/radar/live-stream-apresenta-novos-capitulos-de-6c8228.md) - Published on July 7, 2026, the notice describes a live stream about new chapters of Programming Massively Parallel Processors and an overview of the material.
* [Study traces training data linked to LLM circuits](https://ai.exponencialadm.net/okf/en/radar/rastreando-dados-de-treino-por-tras-608a41.md) - Published on July 7, 2026, the work presents a method for linking training samples to interpretable attention heads and reports effects of interventions on those circuits.
* [Rejection sampling for on-policy LLM alignment](https://ai.exponencialadm.net/okf/en/radar/rejection-sampling-no-alinhamento-on-policy-bee60a.md) - Published on July 7, 2026, the summary presents a proposal to convert off-policy tokens into tokens treated as on-policy through rejection sampling, addressing variance in LLM alignment.
* [Privileged-information distillation may affect long reasoning](https://ai.exponencialadm.net/okf/en/radar/self-distillation-pode-prejudicar-raciocinios-longos-ff99bd.md) - A study evaluates self-distillation with privileged information in reasoning models and reports degradation in long trajectories across five models tested on AIME.
* [A taxonomy of agentic recommender systems](https://ai.exponencialadm.net/okf/en/radar/tres-paradigmas-para-agentes-em-recomendacao-440c1b.md) - A survey published on July 7, 2026 proposes classifying recommender systems with LLM-based agents by their level of autonomy.
* [Autodata generates training data with agent loops](https://ai.exponencialadm.net/okf/en/radar/autodata-cria-dados-de-treino-com-a57316.md) - Published on July 6, 2026, this summary presents Autodata, a method that uses agents to create training and evaluation data by selecting questions that distinguish weak and strong solvers.
* [AutoMem trains agents to manage their own memory](https://ai.exponencialadm.net/okf/en/radar/automem-ensina-agentes-a-gerenciar-a-f58d63.md) - A post about a paper describes agents that write, search and organize notes, learning when memory is useful. According to the post, this improved an open 32B model's performance by 2x to 4x in long games.
* [Floating-point arithmetic from scratch](https://ai.exponencialadm.net/okf/en/radar/como-estudar-aritmetica-de-ponto-flutuante-9d0c36.md) - Published on July 6, 2026, the article presents building floating-point arithmetic from scratch as a high-difficulty project.
* [An Introduction to Compilers and Language Design](https://ai.exponencialadm.net/okf/en/radar/compiladores-e-design-de-linguagens-introducao-dbaf53.md) - A 2021 book by Douglas Thain introduces compilers and language design for engineers interested in building these technologies.
* [Functional Attention proposes linear-cost attention](https://ai.exponencialadm.net/okf/en/radar/functional-attention-reduz-o-custo-da-056f16.md) - Published on July 6, 2026, the paper summary describes learned bases and a compact matrix for representing attention, reports linear cost in the number of points, and cites results in three evaluation areas.
* [NVIDIA CUDA C++ guide updated in June 2026](https://ai.exponencialadm.net/okf/en/radar/guia-cuda-c-atualizado-em-junho-93710e.md) - A bookmark published on July 6, 2026 reports that NVIDIA's CUDA C++ Programming Guide was updated in June 2026 and describes it as a current reference for engineers working with CUDA C++.
* [HOLA combines recurrent state with a limited KV cache](https://ai.exponencialadm.net/okf/en/radar/hola-combina-estado-recorrente-e-kv-ed3f57.md) - In a post published July 6, 2026, HOLA combines recurrent state with a small KV cache for exact memory. The post reports a 16.1% reduction in perplexity and robust retrieval at 32k tokens.
* [Radar: the LLVM compiler infrastructure](https://ai.exponencialadm.net/okf/en/radar/llvm-inovacoes-e-impacto-na-infraestrutura-097724.md) - An article published on July 6, 2026 examines innovations, broad impact, and ongoing work around the LLVM compiler infrastructure, offering context for engineers in the field.
* [PACE estimates agent performance from few tasks](https://ai.exponencialadm.net/okf/en/radar/pace-preve-desempenho-de-agentes-com-d68420.md) - Published on July 6, 2026, the post presents PACE, which uses regression on a small set of non-agentic tasks to estimate agents’ full performance.
* [ReContext explores evidence in long contexts](https://ai.exponencialadm.net/okf/en/radar/recontext-reorganiza-evidencias-para-contextos-longos-f9f5b0.md) - Published on July 6, 2026, the paper presents ReContext, a training-free inference method that resubmits relevant evidence before generation. It reports results on eight 128K datasets and three model backbones.
* [Study examines in-context retrieval at one-million-token scale](https://ai.exponencialadm.net/okf/en/radar/retrieval-in-context-testado-em-escala-7f9c1a.md) - A study examines language models as in-context retrievers over one-million-token corpora and reports that BlockSearch generalizes up to ten times beyond its training length.
* [Radar: Meta storage architecture](https://ai.exponencialadm.net/okf/en/radar/armazenamento-localidade-e-desempenho-04aa40.md) - A July 5, 2026 publication describes a storage architecture and performance techniques, including data locality, cache unification, and pre-fetching.
* [EfficientRollout aims to speed up RL rollouts](https://ai.exponencialadm.net/okf/en/radar/efficientrollout-acelera-rollouts-de-rl-para-ae5fe3.md) - The paper describes system-adaptive self-speculative decoding for RL rollouts in LLMs. Its available summary reports up to 19.6% faster generation; another figure is cut off.
* [HASTE organizes skills for ML engineering agents](https://ai.exponencialadm.net/okf/en/radar/haste-testa-skills-em-camadas-para-193fc9.md) - Published on July 5, 2026, the entry describes HASTE, a hierarchical multi-agent system. In an ablation with 159 skills across eight competitions, hierarchical loading achieved a 100% medal rate, compared with 62.5% for flat loading.
* [HOLA adds a 64-token cache to linear attention](https://ai.exponencialadm.net/okf/en/radar/hola-adiciona-cache-de-64-tokens-6ce890.md) - A post describes a parameter-free token-surprise metric and a 64-token cache intended to address catastrophic forgetting in linear attention. It claims to outperform full-attention baselines.
* [Cosine taper lowers perplexity without expanding the budget](https://ai.exponencialadm.net/okf/en/radar/modelos-de-linguagem-distribuem-capacidade-entre-90cf3c.md) - The paper examines parameter allocation across layers and reports lower perplexity with a cosine taper across several architectures and scales, without increasing parameter count or FLOPs.
* [Pruning or training small LLMs from scratch](https://ai.exponencialadm.net/okf/en/radar/pruning-ou-treino-do-zero-para-a205f4.md) - A paper compares six pruning methods for Llama-3.1-8B with training from scratch under two token-budget scenarios. The advantage depends on how the budget is counted and on pruning granularity.
* [Local branching to scale LLMs at inference time](https://ai.exponencialadm.net/okf/en/radar/ramificacoes-locais-para-escalar-llms-na-b427bd.md) - Published on July 5, 2026, the material presents a local branching approach that previews possible next tokens and selects a path before generation continues.
* [Post-training RL: tuning one layer](https://ai.exponencialadm.net/okf/en/radar/rl-pos-treinamento-ajustar-uma-camada-9c1a2a.md) - A post about a paper reports that reinforcing high-contribution intermediate layers may match or outperform RL over all parameters while changing fewer parameters.
* [RubricEM uses rubrics to guide research agents](https://ai.exponencialadm.net/okf/en/radar/rubricem-ensina-agentes-de-pesquisa-a-29b1cf.md) - Published on July 5, 2026, the item describes a method that trains agents to create rubrics and use them for planning, searching, reviewing, and answering.
* [Cornell offers a workshop on GPU architecture](https://ai.exponencialadm.net/okf/en/radar/workshop-da-cornell-apresenta-arquitetura-de-dcfc0a.md) - The Cornell Virtual Workshop provides introductory material on GPU architecture, a subject that can help engineers analyze GPU performance.
* [Articles and talks on GPU optimization](https://ai.exponencialadm.net/okf/en/radar/artigos-e-palestras-sobre-otimizacao-de-9c27fd.md) - A collection of Vasily Volkov’s work on GPU performance, covering topics such as loop unrolling, latency hiding, dense linear algebra, and GPU memory.
* [Reward signals can diverge from human intent](https://ai.exponencialadm.net/okf/en/radar/como-avaliar-recompensas-de-coding-agents-1a6d42.md) - Published on July 4, 2026, the post describes four signals for evaluating coding agents and warns that they can drift from what people intend.
* [FlowTracer tracks information to guide RL](https://ai.exponencialadm.net/okf/en/radar/flowtracer-rastreia-informacao-para-orientar-o-a8cbe7.md) - The article presents a method that uses attention signals to trace information flow in language-model reasoning and identify influential tokens. Those signals can guide credit assignment in reinforcement learning.
* [Understanding latency hiding in GPUs](https://ai.exponencialadm.net/okf/en/radar/latency-hiding-como-gpus-mantem-o-f23db6.md) - A chapter by Vasily Volkov of UC Berkeley covers how GPUs hide latency and introduces concepts in GPU architecture and performance.
* [GPU Execution Model: a chapter for study](https://ai.exponencialadm.net/okf/en/radar/modelo-de-execucao-de-gpu-capitulo-ddda65.md) - A chapter in Modern GPU Programming for MLSys presents the GPU execution model as a study topic for people working with modern programming and machine-learning systems.
* [TinyRouter tests routing across open models](https://ai.exponencialadm.net/okf/en/radar/tinyrouter-explora-o-roteamento-entre-modelos-89abdc.md) - Published on July 4, 2026, the account describes a router with about 10,000 parameters and initial benchmark results, while noting that more experiments are needed.
* [HOLA proposes an exact KV cache for linear attention](https://ai.exponencialadm.net/okf/en/radar/hola-combina-atencao-linear-e-cache-3e0591.md) - Published on July 3, 2026, the article summary says HOLA combines a recurrent state based on the delta rule with a bounded, exact key-value cache. It reports improved Wikitext perplexity and robust needle recall on RULER up to 32k tokens.
* [Study examines in-context retrieval with millions of tokens](https://ai.exponencialadm.net/okf/en/radar/retrieval-in-context-em-escala-de-1a2c12.md) - Published on July 3, 2026, the report describes a study of answer retrieval from large in-context corpora and generalization to longer lengths.
* [How Statically Linked Programs Start on Linux](https://ai.exponencialadm.net/okf/en/radar/como-programas-linkados-estaticamente-iniciam-no-ac04a5.md) - Published on July 2, 2026, Eli Bendersky's piece explains how statically linked programs run on Linux, focusing on what happens during program startup.
* [Open-source real-time voice demo](https://ai.exponencialadm.net/okf/en/radar/demo-open-source-de-voz-em-63b4d5.md) - Published on July 2, 2026, the briefing describes a real-time voice demo with open-source models and code, supporting WebSocket or WebRTC.
* [Diffusion-GR2 aims to speed up reranking](https://ai.exponencialadm.net/okf/en/radar/diffusion-gr2-acelera-reranking-com-raciocinio-9fd635.md) - Published on July 2, 2026, the post presents Diffusion-GR2, a block diffusion model for generative-reasoning reranking. It reports ranking accuracy close to the autoregressive model and decoding 2.4–3.5 times faster.
* [Video introduces attention steps in Transformers](https://ai.exponencialadm.net/okf/en/radar/entenda-o-pipeline-de-attention-em-aba3d1.md) - Published on July 2, 2026, the video covers Q/K/V, attention scores, and encoder blocks. Its description explains how queries are compared with keys to weight value vectors.
* [Understanding Latency Hiding in GPUs](https://ai.exponencialadm.net/okf/en/radar/entendendo-o-ocultamento-de-latencia-em-1dc6ce.md) - Material by Vasily Volkov of UC Berkeley explores GPU architecture and performance concepts that are not covered in depth in a well-known parallel programming book.
* [LLM hard negatives for two-tower retrieval](https://ai.exponencialadm.net/okf/en/radar/hard-negatives-com-llm-em-retrieval-134940.md) - A paper proposes self-supervised hard-negative sampling for two-tower recommendation, using LLM-derived item clusters to generate harder examples in real time.
* [Shared QKV projections in Transformers](https://ai.exponencialadm.net/okf/en/radar/projecoes-compartilhadas-em-transformers-o-que-fdd85e.md) - Published on July 2, 2026, this bookmark describes a study of attention-projection sharing variants and reports that setting K equal to V was excessive for the author's small ternary LLMs.
* [RoPoLL proposes geometric median for LLM judge panels](https://ai.exponencialadm.net/okf/en/radar/ropoll-agrega-avaliacoes-de-paineis-de-9b1eef.md) - The paper formalizes LLM evaluator panels under a contamination model and proposes aggregating scores with the geometric median. It reports experiments with 13 judges and corruption rates up to 50%.
* [A reading list on Transformers and training](https://ai.exponencialadm.net/okf/en/radar/roteiro-de-leitura-sobre-transformers-e-9af6c3.md) - Published on July 2, 2026, the list presents ten papers in a suggested order on Transformers, scaling, few-shot learning, RLHF, LoRA, FlashAttention, chain-of-thought, and DPO.
* [Online DSpark training example in Speculators](https://ai.exponencialadm.net/okf/en/radar/treinamento-online-de-dspark-no-speculators-bd5faf.md) - Published on July 2, 2026, this entry highlights a Speculators shell script for online DSpark training with Qwen3 0.6B and ShareGPT.
* [Trie speeds up IR pipeline experiments](https://ai.exponencialadm.net/okf/en/radar/trie-reutiliza-etapas-em-pipelines-de-3aeb8f.md) - An article describes a radix trie execution plan for PyTerrier that reuses shared pipeline prefixes and reports experiment time reductions of up to 26%.
* [Radar: catalogue of compiler optimizations](https://ai.exponencialadm.net/okf/en/radar/um-catalogo-de-otimizacoes-para-compiladores-6e951d.md) - A record about Frances Allen’s 1971 article cataloguing compiler optimization transformations and listing passes attributed to Graydon Hoare.
* [Vocabulary Transfer for sparse retrieval](https://ai.exponencialadm.net/okf/en/radar/vocabulary-transfer-e-a-lacuna-de-efad92.md) - A July 2, 2026 brief on a proposal to address vocabulary gaps in learned sparse retrieval with advanced encoders.
* [EMP uses effective sample size to prune models](https://ai.exponencialadm.net/okf/en/radar/effective-model-pruning-usa-tamanho-amostral-b58c94.md) - Researchers at the University of Florida and Ohio University present Effective Model Pruning, a method that uses effective sample size to calculate how many components to retain in neural networks and language models.
* [Intel updates paper on memory protections](https://ai.exponencialadm.net/okf/en/radar/intel-atualiza-paper-sobre-protecoes-de-7e7dce.md) - On July 1, 2026, an update was announced to Intel’s technical paper on memory protections for confidential computing.
* [The Art of Debugging: an open book](https://ai.exponencialadm.net/okf/en/radar/livro-aberto-apresenta-metodologia-de-depuracao-91c7d0.md) - Published on July 1, 2026, this open book has five chapters and is available in PDF and EPUB. The author says its general debugging methodology applies broadly, despite a focus on Unix, Python, and PyTorch.
* [Synthesized programs approximate attention heads](https://ai.exponencialadm.net/okf/en/radar/programas-sintetizados-aproximam-attention-heads-41fe6e.md) - A post reports Python programs synthesized from attention maps, reaching 69–79% mean IoU on three models; replacing up to 30–40% of heads reportedly caused little decline in QA.
* [STEB evaluates style embeddings across 96 datasets](https://ai.exponencialadm.net/okf/en/radar/steb-mede-embeddings-de-estilo-em-657279.md) - The open-source benchmark compares style embeddings across 96 datasets and seven languages. The post reports that semantic embeddings perform worse on stylistic tasks and that no model dominates.
* [Agents and continual learning in ARC-AGI-3](https://ai.exponencialadm.net/okf/en/radar/arc-agi-3-e-o-aprendizado-90aef2.md) - A June 30, 2026 post says that learning during inference in the ARC-AGI-3 benchmark leads agents to build and update internal models of the environment’s rules and mechanics.
* [Local-first search: a framework for retrieval](https://ai.exponencialadm.net/okf/en/radar/busca-local-first-um-framework-para-415e57.md) - Published on June 30, 2026, the paper proposes organizing information retrieval systems with indexes, models, and inference on users’ devices, while making remote services optional.
* [C++26 standardizes Hazard Pointers](https://ai.exponencialadm.net/okf/en/radar/c-26-padroniza-hazard-pointers-para-cc1bb4.md) - Published on June 30, 2026, the post says C++26 standardizes Hazard Pointers, an approach that delays deleting nodes until concurrent readers finish using them.
* [CAMI evaluates RAG indexes under a budget](https://ai.exponencialadm.net/okf/en/radar/cami-avalia-indices-de-rag-sob-9d127f.md) - Published on June 30, 2026, the CAMI paper frames the construction of multiple semantic enrichment indexes as portfolio selection with multiple objectives and a limited budget.
* [RAG: study evaluates retrieval techniques with a reranker](https://ai.exponencialadm.net/okf/en/radar/como-avaliar-melhorias-de-retrieval-em-b691d1.md) - Published on June 30, 2026, the study examines RAG retrieval techniques with a strong cross-encoder reranker and tests whether reported gains hold across varied collections.
* [Reducing the Train–Inference Gap in RL for LLMs](https://ai.exponencialadm.net/okf/en/radar/como-reduzir-a-diferenca-entre-treino-c8f85c.md) - An article examines how separate engines can assign inconsistent probabilities to the same trajectories during LLM post-training, and describes TIS and rejecting checkpoint updates.
* [Stanford course covers GPUs, CUDA, and parallel programming](https://ai.exponencialadm.net/okf/en/radar/curso-de-stanford-aborda-gpus-cuda-c9714b.md) - Published on June 30, 2026, the listing presents a Stanford course with 19 lessons and 24 hours on GPU and CUDA architecture, performance optimization, and multicore architectures.
* [Two-stage distillation for multitask agentic LLMs](https://ai.exponencialadm.net/okf/en/radar/distilacao-em-duas-fases-para-llms-ad0dab.md) - Published on June 30, 2026, the summary describes consolidating reinforcement-learning specialists into a multitask model through off-policy initialization and on-policy refinement.
* [Dockerless checks patches without running code](https://ai.exponencialadm.net/okf/en/radar/dockerless-avalia-patches-de-agentes-de-d8aa57.md) - Published on June 30, 2026, the brief presents Dockerless, a verifier that assesses patches generated by coding agents without executing them.
* [EvoRec proposes agents to optimize recommender systems](https://ai.exponencialadm.net/okf/en/radar/evorec-agentes-para-otimizar-sistemas-de-a19033.md) - Published on June 30, 2026, the summary presents EvoRec, a multi-agent framework intended to automate recommender-system optimization and reduce manual iteration.
* [HydraHead combines attention by head](https://ai.exponencialadm.net/okf/en/radar/hydrahead-combina-attention-por-head-c7c76b.md) - Published on June 30, 2026, the post describes an architecture combining Full Attention and Linear Attention at the head level for long-context models.
* [IW-OPD reweights tokens in on-policy distillation](https://ai.exponencialadm.net/okf/en/radar/iw-opd-repondera-tokens-na-destilacao-49c0a2.md) - A June 30, 2026 publication presents IW-OPD, a method that gives more weight to early tokens. The post reports faster convergence and a 6.9-point advantage over standard OPD on AIME-2025.
* [KbSD proposes self-distillation to calibrate agentic search](https://ai.exponencialadm.net/okf/en/radar/kbsd-calibra-decisoes-em-agentic-search-15543c.md) - Published on June 30, 2026, the summary introduces KbSD, a self-distillation framework using dense token-level supervision for agentic search decisions. It reports no evaluation results.
* [Learning rate may scale nonlinearly](https://ai.exponencialadm.net/okf/en/radar/learning-rate-pode-nao-escalar-linearmente-530fbc.md) - A paper dated June 30, 2026 evaluates learning-rate scaling in GPT-2-style models and reports an upward trend for the optimal rate.
* [Tradecraft Garden publishes research on evasion and linking](https://ai.exponencialadm.net/okf/en/radar/pesquisa-sobre-evasao-e-ferramentas-de-c98287.md) - A June 30, 2026 post describes Crystal Palace, a linker with code transformation and randomization, YARA rule generation, and support for reusable tradecraft modules.
* [Branch prediction and software performance](https://ai.exponencialadm.net/okf/en/radar/predicao-de-desvios-e-desempenho-de-4f52a2.md) - Lesson 17 in the Digital Design and Computer Architecture series covers advanced branch prediction, a processor feature that can affect software performance.
* [Code agents: limits of reward signals](https://ai.exponencialadm.net/okf/en/radar/recompensas-de-agentes-de-codigo-perdem-a6caf9.md) - Published on June 30, 2026, the paper examines test approval, language-model judges, and execution traces as reward signals for code agents.
* [DiscoBench evaluates deep-search agents](https://ai.exponencialadm.net/okf/en/radar/discobench-testa-agentes-de-busca-diante-2a8d6d.md) - Published on June 29, 2026, DiscoBench is described as a benchmark that tests LLM search agents on vague, underspecified, or factually incorrect requests. The description provides no performance results.
* [From nvcc to CUDA kernel execution](https://ai.exponencialadm.net/okf/en/radar/do-nvcc-aos-warps-acompanhando-um-fdb1e4.md) - An article published on June 29, 2026 follows a vector-addition kernel from compilation with nvcc to execution by GPU warps.
* [Using programs to explain attention heads](https://ai.exponencialadm.net/okf/en/radar/estudar-attention-heads-com-programas-executaveis-a0a167.md) - An article published on June 29, 2026 proposes approximating deep-network components with executable programs, focusing on attention heads in Transformers. It calculates attention matrices on randomly selected training examples and uses a pretrained language model.
* [Paper Assistant and scaling scientific review](https://ai.exponencialadm.net/okf/en/radar/paper-assistant-e-a-revisao-cientifica-69c9d3.md) - An article treats the scalability of scientific review as a systems challenge and proposes four levels of collaboration between AI and people, while discussing Google Paper Assistant as an initial tool.
* [Paper Assistant and AI-assisted scientific review](https://ai.exponencialadm.net/okf/en/radar/paper-assistant-propoe-revisao-cientifica-assistida-6f0bdf.md) - Published on June 29, 2026, the item presents a paper proposing a taxonomy for AI-assisted scientific review and describes a system focused on detecting errors and reviewing manuscripts.
* [Paper explains attention heads with Python programs](https://ai.exponencialadm.net/okf/en/radar/programas-python-ajudam-a-explicar-heads-50735b.md) - An automated interpretability technique describes Transformer attention heads as Python programs; replacing about 40% of patterns in Llama-3B had almost no effect on task performance.
* [Red Queen Gödel Machine coevolves agents and evaluators](https://ai.exponencialadm.net/okf/en/radar/red-queen-godel-machine-coevolui-agentes-c2206b.md) - A self-improvement method evolves an agent and its evaluator, keeping evaluators fixed during each epoch and replacing them only after stronger performance on held-out ground truth.
* [SHIFT proposes activation gates for RAG conflicts](https://ai.exponencialadm.net/okf/en/radar/shift-ajusta-ativacoes-para-lidar-com-cc1ac6.md) - Published on June 29, 2026, the item describes SHIFT, which adds learnable gates to FFN activations to balance retrieved context and parametric knowledge when they conflict.
* [Strix: open-source AI pentesting tool](https://ai.exponencialadm.net/okf/en/radar/strix-pentest-com-ia-em-codigo-9adf8a.md) - Published on June 29, 2026, the entry describes Strix as an open-source AI penetration-testing tool for finding and fixing vulnerabilities in applications.
* [VulHunt: detecting vulnerabilities in binaries](https://ai.exponencialadm.net/okf/en/radar/vulhunt-pesquisa-para-detectar-vulnerabilidades-em-c61412.md) - Dated June 29, 2026, this bookmark presents VulHunt, a framework from Binarly's REsearch team, and mentions a post series with an example of detecting a remote code execution vulnerability in rsync.
* [Geometry of arithmetic representations in LLMs](https://ai.exponencialadm.net/okf/en/radar/a-geometria-interna-da-aritmetica-em-dab4ad.md) - Published on June 28, 2026, the paper studies representations during multi-operand addition and proposes a geometric account of carries and arithmetic errors in LLMs.
* [Black-box analysis of Xiaomi’s MJA1 secure chip](https://ai.exponencialadm.net/okf/en/radar/como-analisar-um-chip-seguro-sem-8b3138.md) - Published on June 28, 2026, Quarkslab’s work describes analysis of the MJA1 chip, I2C traffic and firmware, and evaluates its security properties.
* [Reasoning and knowledge stored in LLMs](https://ai.exponencialadm.net/okf/en/radar/como-o-raciocinio-pode-revelar-conhecimento-c724a1.md) - Published on June 28, 2026, the Google Research text examines how reasoning may help LLMs access knowledge stored in their weights.
* [ARM/AARCH64 CTF offers 29 exploitation techniques](https://ai.exponencialadm.net/okf/en/radar/ctf-arm-reune-29-tecnicas-de-77ebcf.md) - Updated on June 28, 2026, the exploit_me repository offers a vulnerable ARM/AARCH64 app for CTF-style exploitation practice. The author says obvious symbols were removed for LLM fine-tuning.
* [DreamSmooth smooths sparse rewards in model-based RL](https://ai.exponencialadm.net/okf/en/radar/dreamsmooth-suaviza-rewards-esparsos-em-model-c160cb.md) - Published on June 28, 2026, this item describes an ICLR proposal to smooth rewards over time between steps for model-based reinforcement learning agents.
* [FUTEX_PRIVATE_FLAG may reduce lock overhead](https://ai.exponencialadm.net/okf/en/radar/futex-private-flag-pode-reduzir-overhead-f213e9.md) - Published on June 28, 2026, the summary says FUTEX_PRIVATE_FLAG may reduce lock latency between threads in the same process by letting the kernel ignore a shared global hash table.
* [Red Queen Gödel Machine coevolves agents and evaluators](https://ai.exponencialadm.net/okf/en/radar/red-queen-godel-machine-coevolui-agentes-66a281.md) - A paper proposes including evaluators in recursive self-improvement of agents, rather than assuming that verifiers, benchmarks, or labeled datasets remain fixed.
* [A curated list of computing books](https://ai.exponencialadm.net/okf/en/radar/uma-lista-selecionada-de-livros-de-64c85d.md) - A page attributed to Brendan Gregg collects selected books across computing fields and may help engineers find reading recommendations.
* [BINEVAL makes LLM evaluations more interpretable](https://ai.exponencialadm.net/okf/en/radar/bineval-torna-avaliacoes-de-llms-mais-8b1ec2.md) - The method breaks criteria into yes-or-no questions and aggregates the answers into multidimensional scores. The report says it matched or outperformed two methods, without training, on three datasets.
* [Linux perf examples for profiling and debugging](https://ai.exponencialadm.net/okf/en/radar/exemplos-de-linux-perf-para-profiling-6c0255.md) - A practical reference for exploring Linux perf (perf_events) in performance analysis and debugging of Linux systems, with a range of event sources.
* [iLLaDA trains an 8B diffusion model](https://ai.exponencialadm.net/okf/en/radar/illada-treina-modelo-de-difusao-de-7c9350.md) - A paper reports training an 8-billion-parameter Transformer with bidirectional masked diffusion on 12 trillion tokens and compares its results with autoregressive models.
* [A packet’s path through the Linux kernel](https://ai.exponencialadm.net/okf/en/radar/o-caminho-de-um-pacote-pelo-9471a5.md) - A publication points to a PDF about how packets travel through the Linux kernel and how that can help analyze network behavior.
* [Why CPU utilization can mislead](https://ai.exponencialadm.net/okf/en/radar/por-que-a-utilizacao-de-cpu-61fce3.md) - In an article published on June 27, 2026, Brendan Gregg discusses the limitations of CPU utilization as a metric for evaluating system performance.
* [Survey maps agents that evolve with experience](https://ai.exponencialadm.net/okf/en/radar/survey-aborda-agentes-que-evoluem-com-7ca482.md) - Published on June 27, 2026, the survey and a related list examine how agents in production turn interaction traces into durable capabilities, from self-evolution to meta-evolution.
* [Windows tokens and privilege escalation](https://ai.exponencialadm.net/okf/en/radar/tokens-de-acesso-do-windows-e-d4858f.md) - A June 27, 2026 post explains the Windows kernel _TOKEN structure and its relationship to identity, security attributes, and privilege escalation.
* [Survey examines evaluation of agent memory](https://ai.exponencialadm.net/okf/en/radar/como-avaliar-benchmarks-de-memoria-para-0896b3.md) - The article reviews the evolution of memory in language-model agents and highlights limits in current evaluations, which often rely on end-to-end task metrics and treat memory as a black box.
* [Lecture covers linking, loading and position-independent code](https://ai.exponencialadm.net/okf/en/radar/linking-loading-e-codigo-independente-de-be451c.md) - A lecture by Anton Burtsev of the University of California, Irvine, covers linking and loading with position-independent code, a topic relevant to executable layout and security hardening.
* [NOVA explores agents for evolving recommendation systems](https://ai.exponencialadm.net/okf/en/radar/nova-explora-agents-com-verificacao-em-fe6986.md) - Published on June 26, 2026, the summary presents NOVA, an agent harness that proposes changes to recommendation architectures under domain constraints.
* [TAPA proposes token-aware positional encoding](https://ai.exponencialadm.net/okf/en/radar/tapa-propoe-atencao-de-fase-sensivel-a1e91c.md) - The article introduces Token-Aware Phase Attention (TAPA), a positional encoding method with a trainable phase function, and analyzes RoPE's distance-dependent bias. It reports a comparison of performance in long contexts.
* [YaFF proposes zero-copy serialization for Protobuf](https://ai.exponencialadm.net/okf/en/radar/yaff-propoe-serializacao-zero-copy-para-a13787.md) - The YaFF C++ library offers a zero-copy wire format for the Protobuf ecosystem. Engineers evaluating serialization options can consult its implementation and documentation.
* [Autodata proposes agents to create training and evaluation data](https://ai.exponencialadm.net/okf/en/radar/autodata-estrutura-agentes-para-criar-e-dc926c.md) - Published on June 25, 2026, the text presents Autodata, a staged method in which AI agents act as data scientists to create and analyze training and evaluation data, with meta-optimization.
* [AutoDecompiler explores RL for iterative decompilation](https://ai.exponencialadm.net/okf/en/radar/autodecompiler-explora-rl-para-decompilacao-iterativa-470638.md) - AutoDecompiler is described as a reinforcement-learning-optimized model that uses feedback across multiple steps for decompilation, unlike single-step approaches.
* [Radar: glibc malloc internals](https://ai.exponencialadm.net/okf/en/radar/entenda-os-internals-do-malloc-no-f82cf3.md) - A June 25, 2026 publication points to the glibc wiki page on malloc internals and notes its usefulness for analyzing memory allocation behavior.
* [Radar: inside glibc malloc](https://ai.exponencialadm.net/okf/en/radar/entendendo-os-internals-do-glibc-malloc-d8a7f5.md) - An article published on June 25, 2026, introduces aspects of the glibc malloc allocator, including how the heap is obtained from the kernel and how memory is managed.
* [Framework compares RAG approaches for semi-structured data](https://ai.exponencialadm.net/okf/en/radar/framework-compara-quatro-abordagens-de-rag-af2be0.md) - A paper published on June 25, 2026 presents a framework for comparing regular, Graph, Modular, and Agentic RAG on semi-structured knowledge bases, using nine standardized scenarios.
* [FuncAttn applies functional attention to continuous data](https://ai.exponencialadm.net/okf/en/radar/funcattn-aplica-attention-funcional-a-dados-da051c.md) - Published on June 25, 2026, the post introduces FuncAttn, a method applying attention between function spaces, and reports results in PDE solving, 3D segmentation, and regression.
* [iLLaDA: an 8B masked diffusion model](https://ai.exponencialadm.net/okf/en/radar/illada-difusao-mascarada-em-um-modelo-74df6b.md) - Published on June 25, 2026, the entry describes iLLaDA, an 8-billion-parameter masked diffusion language model trained from scratch with fully bidirectional attention.
* [Ornith-1.0 presents open models for agentic coding](https://ai.exponencialadm.net/okf/en/radar/ornith-1-0-apresenta-modelos-abertos-83bf14.md) - Published on June 25, 2026, Ornith-1.0 is a family of open models ranging from 9 billion to 397 billion parameters. Reinforcement learning is used to optimize both code solutions and the scaffolds that guide them.
* [PostTrainBench: GLM 5.2 tops the ranking](https://ai.exponencialadm.net/okf/en/radar/posttrainbench-poe-glm-5-2-no-aeadb3.md) - Published on June 25, 2026, the report gives GLM 5.2 (Max reasoning) a score of 34.29%, versus 34.08% for Opus 4.8 Max, and reports zero failures across 84 GLM 5.2 attempts.
* [Reversecore_MCP automates reverse engineering](https://ai.exponencialadm.net/okf/en/radar/reversecore-mcp-leva-agentes-de-ia-2657b1.md) - Listed on June 25, 2026, the entry describes an MCP server for AI agents with tools for binary, malware, and vulnerability analysis.
* [Reverse Engineering and Malware Analysis Roadmap](https://ai.exponencialadm.net/okf/en/radar/roadmap-de-engenharia-reversa-e-analise-5da265.md) - A briefing published on June 25, 2026, describes a GitHub roadmap covering topics and tools for reverse engineering and malware analysis.
* [Video introduces LoRA and fine-tuning methods](https://ai.exponencialadm.net/okf/en/radar/video-apresenta-lora-e-metodos-de-772f68.md) - Published on June 25, 2026, the video introduces LoRA, LoRA+, QLoRA, VeRA, and DoRA as related parameter-efficient fine-tuning methods.
* [Iterative code improves scientific forecasts](https://ai.exponencialadm.net/okf/en/radar/codigo-automatizado-aprimora-previsoes-cientificas-6a6c0c.md) - A Nature study assessed repeated code generation and testing for scientific forecasting across six fields. Of 55 combinations of existing methods, 24 outperformed both methods from which they originated.
* [NatureBench evaluates coding agents on 90 tasks](https://ai.exponencialadm.net/okf/en/radar/naturebench-avalia-agentes-de-codigo-em-ffd0d1.md) - Published on June 24, 2026, the briefing describes a benchmark based on 90 tasks from papers in the Nature family. According to the post, the top agent exceeded the published reference result on 17.8% of tasks.
* [Physical neural network of metasurfaces trained without digital models](https://ai.exponencialadm.net/okf/en/radar/rede-fisica-aprende-sem-modelo-digital-29635f.md) - A Nature Communications article describes a physical neural network with four programmable metasurfaces that learns through electromagnetic interactions, without relying on digital models.
* [Riot announces Vanguard On-Demand](https://ai.exponencialadm.net/okf/en/radar/riot-anuncia-o-vanguard-on-demand-24fdd9.md) - On June 24, 2026, Riot announced an option to run Vanguard only during gameplay, conditional on specific device security features being enabled.
* [Long-horizon RL: PPO or GRPO?](https://ai.exponencialadm.net/okf/en/radar/rl-de-longo-horizonte-quando-ppo-20012f.md) - A Zhihu post published on June 24, 2026 compares GRPO and PPO for reinforcement learning tasks with LLMs, highlighting credit-assignment challenges in long, noisy agentic rollouts.
* [An effective theory for understanding deep neural networks](https://ai.exponencialadm.net/okf/en/radar/teoria-efetiva-para-entender-redes-neurais-19f0b7.md) - A book published on June 24, 2026 presents a theoretical approach linking the structure and training of deep networks to model behavior.
* [TMAX trains terminal agents in Dockerized environments](https://ai.exponencialadm.net/okf/en/radar/tmax-treina-agentes-de-terminal-em-4a8741.md) - A post published June 24, 2026 describes 14,600 Dockerized reinforcement-learning environments and an outcome-only DPPO recipe. It reports a 9B model scoring 27% on Terminal-Bench 2.0.
* [Transformer Cookbook explores algorithms in transformer weights](https://ai.exponencialadm.net/okf/en/radar/transformer-cookbook-algoritmos-nos-pesos-bce29d.md) - Published on June 24, 2026, Transformer Cookbook presents an approach to encoding algorithms in transformer weights, following a paper on RASP.
* [Index compares Speech-to-Speech models](https://ai.exponencialadm.net/okf/en/radar/indice-compara-modelos-de-fala-em-811c30.md) - Published on June 23, 2026, the index combines three benchmarks to compare native Speech-to-Speech models on speech reasoning, conversational dynamics, and agentic performance.
* [Infrastructure for RL in trillion-parameter MoE models](https://ai.exponencialadm.net/okf/en/radar/infraestrutura-para-rl-em-moe-com-879171.md) - A briefing on infrastructure techniques for scaling reinforcement-learning workloads on large MoE models, including FP8 and parallelism.
* [Book covers modern GPU programming techniques for ML](https://ai.exponencialadm.net/okf/en/radar/livro-reune-tecnicas-modernas-de-programacao-713d7e.md) - An online resource based on a mini-series in CMU’s ML Systems course covers data layout swizzling, 3D TMA, and programming for Blackwell.
* [Long-document OCR: model processes more than 40 pages](https://ai.exponencialadm.net/okf/en/radar/ocr-de-documentos-longos-em-uma-eb6fb9.md) - Published on June 23, 2026, the report describes an open-source OCR model that processes documents over 40 pages in one pass and reports results on OmniDocBench v1.5 and v1.6.
* [Proposal explores reasoning without model retraining](https://ai.exponencialadm.net/okf/en/radar/reasoning-verificado-sem-atualizar-pesos-1d2157.md) - A June 23, 2026 post proposes retrieving verified reasoning from an external pool for a frozen LLM. Verification and deployment remain challenges, and it is unknown whether the approach can match RL gains.
* [Learning about the Linux kernel with ftrace](https://ai.exponencialadm.net/okf/en/radar/aprenda-tracing-do-kernel-linux-com-613e16.md) - A 2018 session by Steven Rostedt introduces Linux kernel tracing with ftrace and covers computer architecture, virtual memory, page tables, and how programs interact with the kernel.
* [LocalVQE Pi V1 promises audio cleanup on Raspberry Pi 5](https://ai.exponencialadm.net/okf/en/radar/localvqe-pi-v1-limpeza-de-audio-8cb000.md) - Announced on June 22, 2026, the 49,000-parameter model targets acoustic echo cancellation, noise suppression, and reverberation reduction. The post claims 21× real-time performance on one Raspberry Pi 5 core.
* [PPO and GRPO for long-horizon tasks](https://ai.exponencialadm.net/okf/en/radar/ppo-ou-grpo-em-tarefas-de-9c22ee.md) - Published on June 21, 2026, the post argues that PPO may suit long tasks better than GRPO, citing synchronization challenges and reduced effectiveness of group-based variance reduction.
* [PufferLib 4.0 reports multi-GPU scaling results](https://ai.exponencialadm.net/okf/en/radar/pufferlib-4-0-relata-escala-multi-889eb3.md) - A June 21, 2026 publication reports 6.6× scaling on eight GPUs in a worst-case test and throughput above 100 million steps per second on eight RTX 5090 GPUs.
* [Shared libraries: guidance for engineers](https://ai.exponencialadm.net/okf/en/radar/shared-libraries-orientacoes-para-engenheiros-f7110d.md) - A publication highlights Ulrich Drepper’s “How to Write Shared Libraries” as a reference on creating and using shared libraries.
* [Turn-PPO for multi-turn agents](https://ai.exponencialadm.net/okf/en/radar/turn-ppo-e-o-treinamento-de-98e63b.md) - An article introduces Turn-PPO, a per-turn advantage estimation method using PPO for multi-turn reinforcement learning in agentic LLMs.
* [JSON Parsing and SIMD at High Throughput](https://ai.exponencialadm.net/okf/en/radar/parsing-de-json-e-simd-em-1f2594.md) - A publication recommends exploring JSON parsing at gigabytes-per-second speeds and a companion text explaining SIMD, a topic relevant to engineers interested in high-throughput processing.
* [Talk covers ABIs and interoperability between languages](https://ai.exponencialadm.net/okf/en/radar/abis-como-linguagens-se-comunicam-4fefce.md) - In a talk, Shung-Hsi Yu discusses how application binary interfaces (ABIs) enable communication between languages, including data layout changes, symbol versioning, and FFI.
* [ARGUS examines slowness in large-scale training](https://ai.exponencialadm.net/okf/en/radar/argus-busca-diagnosticar-lentidao-em-treinamentos-ee2d8d.md) - A paper published on arXiv on June 19, 2026 describes ARGUS, which combines CPU, framework, and GPU-kernel data to diagnose slowness, reporting continuous overhead below 2%.
* [A proposed test for reward-hacking mitigations](https://ai.exponencialadm.net/okf/en/radar/como-comparar-mitigacoes-de-reward-hacking-ce2ad5.md) - Published on June 19, 2026, the item raises a question about comparing two ways to mitigate reward hacking; it reports no test results.
* [Radar: How to Write Shared Libraries](https://ai.exponencialadm.net/okf/en/radar/como-escrever-bibliotecas-compartilhadas-um-documento-5786c4.md) - A document dated June 19, 2026, about writing shared libraries, presented as useful context for engineers who work with this kind of library.
* [OpenMythos proposes a recurrent-depth Transformer](https://ai.exponencialadm.net/okf/en/radar/openmythos-explora-transformers-de-profundidade-recorrente-73a623.md) - Published on June 19, 2026, the project presents an inspectable theoretical reconstruction of hypotheses about recurrent depth, sparse experts, and adaptive per-token halting.
* [Post compares OPD and RL efficiency](https://ai.exponencialadm.net/okf/en/radar/post-compara-eficiencia-de-opd-e-ed494e.md) - Published on June 19, 2026, a post claims OPD uses compute and samples more efficiently than RL. The available summary provides no metrics or detailed results.
* [slime: an open framework for LLM post-training](https://ai.exponencialadm.net/okf/en/radar/slime-framework-aberto-para-pos-treinamento-1196a1.md) - Published on June 19, 2026, the slime project from THUDM describes an open-source framework for post-training language models with reinforcement learning and scaling.
* [VoiceGate proposes multilingual dubbing in ComfyUI](https://ai.exponencialadm.net/okf/en/radar/voicegate-reune-etapas-de-dublagem-multilingue-f1e7f8.md) - Published on June 19, 2026, the project combines transcription, translation, speech synthesis, and SRT-timestamp alignment for video dubbing.
* [AutoResearch automates GPU-based RL experiments](https://ai.exponencialadm.net/okf/en/radar/autoresearch-automatiza-experimentos-de-rl-244096.md) - Published on June 18, 2026, the account describes a framework that planned GPU experiments and automated steps in a reinforcement-learning pipeline.
* [Chapter 8 introduces Transformer components](https://ai.exponencialadm.net/okf/en/radar/como-estudar-a-arquitetura-transformer-pelo-0182c3.md) - Published on June 18, 2026, chapter 8 of Speech and Language Processing surveys Transformer components, moving from intuition to matrix formulation.
* [Research flags plaintext traffic on GEO satellite links](https://ai.exponencialadm.net/okf/en/radar/pesquisa-alerta-para-trafego-sem-criptografia-2a0194.md) - Research published on June 18, 2026, reports that a significant share of GEO satellite traffic may be intercepted because it is sent in plaintext, using relatively accessible equipment.
* [Post announces bootrom exploit for Apple chips](https://ai.exponencialadm.net/okf/en/radar/post-anuncia-exploit-de-bootrom-para-b27234.md) - A post dated June 18, 2026 claims a bootrom exploit has been released for A12, A13, S4/S5, and A12X/Z devices; those devices had not yet been tested.
* [Ubuntu announces Project Myna for desktop dictation](https://ai.exponencialadm.net/okf/en/radar/project-myna-mira-ditado-local-no-3502b7.md) - Announced on June 18, 2026, Project Myna aims to bring voice dictation to Ubuntu Desktop, inserting text into the active application. The workflow is planned for Ubuntu 26.10, and the project description says processing will run locally.
* [C++26 Reflection for a universal printer](https://ai.exponencialadm.net/okf/en/radar/reflection-do-c-26-para-criar-723b7e.md) - Lieven de Cock demonstrates using C++26 Reflection to inspect and print information about class members.
* [RL for beneficial behaviors across domains](https://ai.exponencialadm.net/okf/en/radar/rl-para-comportamentos-beneficos-em-diferentes-19250c.md) - A report on reinforcement learning says that focusing on beneficial behaviors improved alignment across domains and resisted adversarial pressure.
* [Radar: Cornell self-paced compiler course](https://ai.exponencialadm.net/okf/en/radar/curso-autodirigido-de-compiladores-avancados-115bab.md) - A listing dated June 17, 2026 presents Cornell's self-paced CS 6120 course on advanced compilers for engineers to study at their own pace.
* [Deli AutoResearch automates RL experiments](https://ai.exponencialadm.net/okf/en/radar/deli-autoresearch-automatiza-experimentos-de-rl-55a1a1.md) - An open-source agent framework says it automates several stages of reinforcement learning research workflows, including GPU experiments.
* [Autonomous vulnerability hunting with MCP](https://ai.exponencialadm.net/okf/en/radar/hunting-autonomo-de-vulnerabilidades-com-mcp-be1d39.md) - A ZephrSec article discusses using MCP for autonomous vulnerability hunting with LLMs, a topic for engineers exploring approaches to vulnerability research.
* [Compiler notes: Cornell's CS 4120 course](https://ai.exponencialadm.net/okf/en/radar/notas-de-compiladores-curso-cs-4120-8f34c7.md) - Cornell's CS 4120/ENGRD 4120 notes are described as an introduction to compilers and a course resource for engineers learning the subject.
* [PaddleOCR 3.7 expands OCR execution options](https://ai.exponencialadm.net/okf/en/radar/paddleocr-3-7-amplia-opcoes-de-cb1cfa.md) - Published on June 17, 2026, the PaddleOCR 3.7 announcement highlights an ONNX Runtime backend and CPU and GPU inference options. It says PP-OCRv6 models can be up to 3.9 times faster on CPU.
* [Efficient Attention in Hybrid Architectures](https://ai.exponencialadm.net/okf/en/radar/atencao-eficiente-em-arquiteturas-hibridas-0bb0b6.md) - An analysis examines how sliding-window attention and recurrent sequence mixers affect capabilities, scalability, and architectural choices in hybrid language models.
* [NextLat predicts transformers’ latent states](https://ai.exponencialadm.net/okf/en/radar/nextlat-preve-estados-latentes-de-transformers-3d57c3.md) - Published on June 16, 2026, Next-Latent Prediction (NextLat) is presented as a self-supervised learning method that trains transformers to predict their next latent state.
* [CVE-Genie aims to automate exploit generation](https://ai.exponencialadm.net/okf/en/radar/cve-genie-promete-automatizar-a-criacao-36bbe7.md) - A June 15, 2026 publication describes CVE-Genie as a tool that turns known vulnerabilities into functional exploits using a few API calls. The available preview provides no further details to verify the claim.
* [Dynamic Linear Attention adapts memory to token content](https://ai.exponencialadm.net/okf/en/radar/dynamic-linear-attention-adapta-memoria-aos-7c5e28.md) - Published on June 15, 2026, the account describes a technique that adjusts memory according to token information and reports benchmark, throughput, and memory-use gains.
* [CppCon 2020: a talk on pointers and memory in C++](https://ai.exponencialadm.net/okf/en/radar/ponteiros-e-memoria-em-c-fundamentos-2ba202.md) - A record of Ben Saks’s CppCon 2020 talk on pointers and memory, foundational concepts for C++ engineers.
* [Parallel programming: challenges and approaches](https://ai.exponencialadm.net/okf/en/radar/programacao-paralela-como-enfrentar-os-desafios-cc2377.md) - A resource published on June 15, 2026, covers challenges in parallel programming and ways to address them. It notes that parallel programming concepts help engineers reason about concurrency in GPU kernels.
* [SFT in LLMs: effects explained through interactions](https://ai.exponencialadm.net/okf/en/radar/sft-em-llms-efeitos-vistos-pelas-c40b4e.md) - An article published on June 15, 2026, examines how changes in token interactions may help explain why supervised fine-tuning benefits smaller networks but has inconsistent or harmful effects in LLMs.
* [How speculative decoding verifies proposed tokens](https://ai.exponencialadm.net/okf/en/radar/speculative-decoding-tokens-propostos-e-verificados-4bd487.md) - A June 15, 2026 post explains the approach: a fast draft model proposes tokens, and a larger target model verifies them in parallel.
* [Surfshark presents the Dausos VPN protocol](https://ai.exponencialadm.net/okf/en/radar/surfshark-apresenta-o-protocolo-vpn-dausos-23711d.md) - Surfshark describes Dausos as a new VPN protocol aimed at speed, security, and efficiency, using AEGIS-256X2.
* [Introduction to parallel algorithms](https://ai.exponencialadm.net/okf/en/radar/algoritmos-paralelos-conceitos-e-otimizacao-de-d16739.md) - Bookmark published on June 14, 2026: a post points to an introductory PDF and recommends Aalto’s free course on programming parallel computers.
* [Aalto course covers parallel computer programming](https://ai.exponencialadm.net/okf/en/radar/curso-da-aalto-aborda-programacao-paralela-68c784.md) - Published on June 14, 2026, Aalto’s practical course covers algorithm engineering for modern parallel computers, including multicore CPUs and GPUs.
* [Index collects research on LLMs and vulnerabilities](https://ai.exponencialadm.net/okf/en/radar/pesquisas-sobre-llms-na-deteccao-de-86a21e.md) - Published on June 14, 2026, the index gathers research on using language models to detect software vulnerabilities, from function- and repository-level analysis to agentic approaches.
* [Valgrind for debugging and profiling](https://ai.exponencialadm.net/okf/en/radar/valgrind-debugging-e-profiling-de-programas-8dda23.md) - The Valgrind tool suite helps engineers investigate memory-management and threading bugs, and perform detailed program profiling.
* [Memory allocation strategies](https://ai.exponencialadm.net/okf/en/radar/estrategias-de-alocacao-de-memoria-31bba7.md) - A series by gingerBill introduces memory allocation strategies and may help engineers deepen their understanding of the topic.
* [C++: constexpr and compile-time programming](https://ai.exponencialadm.net/okf/en/radar/c-constexpr-programacao-em-tempo-de-208f57.md) - Published on June 12, 2026, the MC++ Blog article covers using constexpr for compile-time programming, drawing on the C++ Core Guidelines.
* [Attention Matching rapidly compresses KV caches](https://ai.exponencialadm.net/okf/en/radar/compactacao-de-kv-cache-acelera-o-cf7d7f.md) - A June 12, 2026 publication reports a method that creates compact KV caches in latent space while preserving per-head attention outputs. On some datasets, compression reached up to 50× in seconds, with minimal quality loss.
* [Abseil C++ Performance Hints](https://ai.exponencialadm.net/okf/en/radar/dicas-de-performance-em-c-do-696b6c.md) - A resource for people working on C++ performance. Its preview describes Abseil as an open-source collection of core C++ libraries.
* [NIST highlights continuous monitoring for AI security](https://ai.exponencialadm.net/okf/en/radar/nist-defende-monitoramento-continuo-da-seguranca-b158eb.md) - An article published on June 12, 2026 presents a mathematical proof supporting an AI systems security model based on continuous monitoring and updates, as an alternative to static controls.
* [PP-OCRv6 arrives in three model sizes](https://ai.exponencialadm.net/okf/en/radar/pp-ocrv6-traz-modelos-de-ocr-4df483.md) - Announced on June 12, 2026, the PP-OCRv6 series includes Tiny, Small, and Medium models ranging from 1.5 million to 34.5 million parameters. PaddlePaddle reports support for 50 languages and CPU inference up to 5.2 times faster with OpenVINO.
* [Microsoft Research explores encrypted collaboration](https://ai.exponencialadm.net/okf/en/radar/encrypted-spaces-explora-colaboracao-criptografada-8e8f07.md) - Published June 11, 2026, the entry describes Encrypted Spaces, a research project on collaborative applications with encrypted data and cryptographically verifiable operations.
* [Linux patch proposes lock-free MPSC queue for io_uring](https://ai.exponencialadm.net/okf/en/radar/patch-propoe-fila-fifo-mpsc-sem-65866a.md) - Published on June 11, 2026, the Linux kernel patch proposes adding a lockless FIFO queue with multiple producers and one consumer to io_uring.
* [PP-OCRv6 expands options for multilingual OCR](https://ai.exponencialadm.net/okf/en/radar/pp-ocrv6-amplia-opcoes-para-ocr-bd08ef.md) - PP-OCRv6 comes in 1.5-million, 7.7-million, and 34.5-million-parameter versions, with declared support for more than 48 languages and a focus on edge deployment.
* [cvs-fast-export 2.1 fixes error in binary blobs](https://ai.exponencialadm.net/okf/en/radar/cvs-fast-export-2-1-corrige-4a7788.md) - Published on June 10, 2026, cvs-fast-export 2.1 fixes an “invalid delta” error when processing binary blobs. The tool converts RCS or CVS histories into a fast-import stream.
* [Kernel locks in Real-time Ubuntu](https://ai.exponencialadm.net/okf/en/radar/locks-do-linux-kernel-no-real-67a465.md) - Real-time Ubuntu documentation explains how kernel locks protect critical regions from context switches or preemption, including on SMP systems. The Radar entry is dated June 10, 2026.
* [Cohere reports top result in speech benchmark](https://ai.exponencialadm.net/okf/en/radar/modelo-aberto-de-fala-lidera-benchmark-2b815e.md) - On June 10, 2026, Cohere said its open speech recognition model, Cohere Transcribe, ranked first in the new Far-Field ASR benchmark. The result may help engineers evaluate speech recognition models.
* [Self-Harness: agents refine their own harnesses](https://ai.exponencialadm.net/okf/en/radar/self-harness-agentes-refinam-seus-proprios-81d479.md) - A paper examines agents that analyze failures and propose small harness changes, keeping them only after regression tests pass. It reports higher success on unseen Terminal-Bench-2.0 data.
* [TTS improves reading of structured data](https://ai.exponencialadm.net/okf/en/radar/tts-e-a-pronuncia-de-dados-fd5a42.md) - A TTS model update reports improved pronunciation of spelled-out words, acronyms, emails, phone numbers, and codes. The announcement cites real-time benchmark comparisons with the previous version and other models.
* [Guide covers undefined behavior in C and C++](https://ai.exponencialadm.net/okf/en/radar/comportamento-indefinido-em-c-e-c-a1c802.md) - Published on June 9, 2026, the article introduces a guide to undefined behavior in C and C++. Its description gives no specific examples or conclusions.
* [Radar: Linux Kernel locking documentation](https://ai.exponencialadm.net/okf/en/radar/documentacao-de-locking-do-linux-kernel-6c3bc9.md) - Dated June 9, 2026, the item describes Linux Kernel documentation as a guide to kernel locking and concurrency practices.
* [DS4 fork enables distributed inference across multiple machines](https://ai.exponencialadm.net/okf/en/radar/fork-amplia-inference-distribuida-em-varias-c38e57.md) - Published on June 9, 2026, the described fork combines --tensor-parallel with --role to run distributed inference across multiple machines and GPUs.
* [Sparse retrieval reduces KV cache use in long contexts](https://ai.exponencialadm.net/okf/en/radar/retrieval-esparso-reduz-o-kv-cache-879556.md) - A June 9, 2026 post summarizes a paper that predicts relevant KV cache chunks: selected chunks stay on the GPU while the others are offloaded. The account reports an average KV cache footprint of 13.5% and up to 90% memory reduction at 500K context.
* [Radar: semaphores and systems concurrency](https://ai.exponencialadm.net/okf/en/radar/semaforos-um-guia-para-estudar-concorrencia-7b4b4f.md) - Published on June 9, 2026, the entry shares a PDF of Allen B. Downey’s *The Little Book of Semaphores*, about semaphores, a concurrency topic relevant to systems programming.
* [Live translation via API: what the documentation covers](https://ai.exponencialadm.net/okf/en/radar/traducao-ao-vivo-via-api-o-d2d935.md) - Documentation for a real-time conversational API covers live translation. The topic may interest engineers exploring API-based multilingual experiences.
* [UNITTS brings inference and benchmarking to open-source TTS](https://ai.exponencialadm.net/okf/en/radar/unitts-reune-inferencia-e-benchmarking-de-bc0d6e.md) - Published on June 9, 2026, the announcement presents UNITTS, a toolkit for running inference and benchmarking open-source text-to-speech models.
* [Dispatch performance: compiler and standard library effects](https://ai.exponencialadm.net/okf/en/radar/dispatch-em-c-compiler-muda-o-4d8859.md) - A post compares std::variant and virtual dispatch across GCC versions. Its preview reports a shift from 28% slower with GCC 11 to 50% faster with GCC 12, changing only the compiler.
* [CAI: an open-source framework for AI security operations](https://ai.exponencialadm.net/okf/en/radar/cai-framework-aberto-para-seguranca-com-400d67.md) - Published on June 7, 2026, CAI is described as an open-source framework for AI-powered cybersecurity operations, including multi-agent workflows, security tools, and CTF automation.
* [An organized C++ reference on cppreference](https://ai.exponencialadm.net/okf/en/radar/como-consultar-uma-referencia-organizada-de-e9f677.md) - The cppreference reference page covers the C++ language and is described by the author as well organized. Engineers can consult it when working with language features.
* [Bootloaders: risks, detection, and defenses](https://ai.exponencialadm.net/okf/en/radar/bootloaders-riscos-deteccao-e-defesas-53b8a2.md) - An article published on June 5, 2026, covers attack surfaces, vulnerability detection, and defenses in firmware, operating systems, and monolithic bootloaders.
* [Concurrency in C++: a two-part talk](https://ai.exponencialadm.net/okf/en/radar/concorrencia-em-c-palestra-em-duas-32b6eb.md) - A publication recommends Fedor Pikus’s “Concurrency in C++: A Programmer’s Overview,” a talk focused on concurrency in C++.
* [C++ Concurrency: An Overview for Programmers](https://ai.exponencialadm.net/okf/en/radar/concorrencia-em-c-uma-visao-geral-bd5d2d.md) - A two-part talk published on June 5, 2026, offers engineers an overview of concurrency in C++.
* [Handles as an alternative to pointers](https://ai.exponencialadm.net/okf/en/radar/handles-como-alternativa-a-ponteiros-9b2a8d.md) - A post published on June 5, 2026 discusses handles as an alternative to pointers, using a per-slot generation counter to avoid collisions in a memory-management approach relevant to data-oriented design.
* [Gradbot tests parallel replies for voice agents](https://ai.exponencialadm.net/okf/en/radar/modelos-em-paralelo-reduzem-pausas-em-b479b8.md) - Published on June 5, 2026, the item describes a community contribution to Gradbot: one model streams a short reply to a voice system while another reasons and calls tools in the background.
* [GPS field may expose cryptographic logistics metadata](https://ai.exponencialadm.net/okf/en/radar/campo-do-gps-pode-expor-metadados-df8ee5.md) - A publication claims that a public 176-bit field in GPS navigation has carried high-entropy payloads for years. The proposed encrypted control or key-distribution channel remains unproven.
* [Algorithmica guide to RAM and CPU caches](https://ai.exponencialadm.net/okf/en/radar/ram-e-caches-de-cpu-um-f4f03a.md) - Published on June 4, 2026, Algorithmica’s guide covers RAM and CPU caches, topics related to memory performance.
* [Watchdog timers and embedded reliability](https://ai.exponencialadm.net/okf/en/radar/watchdog-timers-e-confiabilidade-embarcada-37bb8e.md) - A brief reference highlights watchdog timers as a practical reliability topic for embedded software engineers, without detailing configurations or procedures.
* [Post warns of AI-orchestrated intrusion workflows](https://ai.exponencialadm.net/okf/en/radar/alerta-workflows-de-intrusao-orquestrados-por-7921e1.md) - Published on June 3, 2026, the post warns that orchestrating models with scanners, shells, and other tools could make intrusion workflows more repeatable and less costly.
* [Why do code reviews?](https://ai.exponencialadm.net/okf/en/radar/code-reviews-qual-e-o-proposito-cef974.md) - The first part of Arne Mertz’s series explores reasons to conduct code reviews—and reasons not to—and gives engineers a basis for assessing their purpose.
* [CppCon 2022: strategies for efficient C++](https://ai.exponencialadm.net/okf/en/radar/como-estudar-estrategias-de-eficiencia-em-a9ee57.md) - A record published on June 3, 2026, about Jan Bielak's CppCon 2022 presentation on strategies for writing efficient C++ programs.
* [Building a Debugger from Scratch: An Eight-Part Series](https://ai.exponencialadm.net/okf/en/radar/criando-um-debugger-do-zero-uma-bf2035.md) - An eight-part blog series covers building a debugger, from connecting to a process to working with source code and symbols.
* [Series explores building a debugger from scratch](https://ai.exponencialadm.net/okf/en/radar/debugger-do-zero-uma-serie-em-0c082b.md) - Published on June 3, 2026, the eight-part series covers building a debugger, from connecting to a process to handling source code and symbols.
* [DFlash and MTP on Qwen3.6: benchmarks](https://ai.exponencialadm.net/okf/en/radar/dflash-e-mtp-speculative-decoding-no-526e9c.md) - The article compares speculative decoding with DFlash and MTP, using vLLM and llama.cpp for math, coding, and chat. Performance varies by model and workload.
* [Research worm adapts attacks to vulnerability classes](https://ai.exponencialadm.net/okf/en/radar/worm-de-pesquisa-com-ia-adapta-6b8c74.md) - A June 3, 2026 publication describes a research worm that uses a language model on compromised machines and adapts its attack logic. In reported tests, it exploited systems and replicated across Linux, Windows, and IoT targets.
* [Collision attacks on SHA-2 advance to 39 steps](https://ai.exponencialadm.net/okf/en/radar/ataques-de-colisao-contra-sha-2-337393.md) - A preview describes an improved search procedure for SHA-2 collision attacks, advancing from earlier results of up to 37 steps to 39. This does not amount to a full attack on the algorithm.
* [Eight-part series teaches debugger construction](https://ai.exponencialadm.net/okf/en/radar/como-estudar-a-construcao-de-um-5bfe46.md) - Published on June 2, 2026, the series covers building a debugger, from connecting to a process to handling source code and symbols, with a practical look at internals relevant to C++ development.
* [List gathers cybersecurity resources](https://ai.exponencialadm.net/okf/en/radar/lista-reune-recursos-de-ciberseguranca-2e272e.md) - Published on June 2, 2026, this GitHub list gathers cybersecurity blog posts, writeups, papers, and tools as a starting point for engineers.
* [Verifying stack limits in C](https://ai.exponencialadm.net/okf/en/radar/verificacao-formal-de-limites-de-stack-df81f7.md) - A 2014 article addresses end-to-end verification of stack usage limits in C programs, a topic relevant to engineers studying formal verification of memory resource bounds.
* [C++: Agner Fog’s guide to memory access](https://ai.exponencialadm.net/okf/en/radar/c-guia-de-agner-fog-sobre-87b606.md) - The publication recommends Chapter 9 of Agner Fog’s *Optimizing Software in C++*, focused on optimizing memory access. Here is a practical way to study the topic.
* [Loop transformer reports 100% on Sudoku-Extreme](https://ai.exponencialadm.net/okf/en/radar/transformer-compacto-testa-deducao-em-sudoku-17c8ab.md) - A June 1, 2026 publication describes a compact, 800,000-parameter loop transformer that it says achieved 100% on Sudoku-Extreme after 15 minutes of training.
* [Radar: reading complex declarations in C](https://ai.exponencialadm.net/okf/en/radar/como-interpretar-declaracoes-complexas-em-c-337310.md) - A May 31, 2026 publication presents the right-left rule for interpreting a complex C declaration and recommends a reference book.
* [Benchmarks measure the cost of rare conditions](https://ai.exponencialadm.net/okf/en/radar/como-medir-o-custo-de-condicoes-ec8133.md) - Published on May 31, 2026, Cloudflare's article examines the performance cost of rarely executed `if` clauses and presents benchmarks on x86 and M1 to assess branch prediction and conditional code.
* [MogVMP devirtualizes VMProtect 3.0–3.5](https://ai.exponencialadm.net/okf/en/radar/mogvmp-converte-codigo-vmprotect-em-llvm-ec2a99.md) - Published on May 31, 2026, MogVMP is described as a static devirtualizer that converts VMProtect-virtualized code to LLVM and removes the virtual machine layer through optimization.
* [C++ Annotations: a comprehensive C++ reference](https://ai.exponencialadm.net/okf/en/radar/c-annotations-referencia-ampla-para-estudar-6c6bb6.md) - Published on May 30, 2026, Frank B. Brokken’s book is described as an up-to-date C++ reference spanning more than 1,100 pages.
* [Linux kernel scheduler documentation](https://ai.exponencialadm.net/okf/en/radar/documentacao-dos-schedulers-do-kernel-linux-650439.md) - Linux kernel documentation describes the schedulers available in the kernel and serves as a reference for investigating process scheduling and system performance.
* [Book teaches ARM compilation with TypeScript](https://ai.exponencialadm.net/okf/en/radar/compiladores-na-pratica-gerando-codigo-para-c0301b.md) - Published on May 29, 2026, Vladimir Keleshev’s book introduces compiler fundamentals and Assembly through a TypeScript subset that generates code for 32-bit ARM.
* [Practical guide to bare-metal C++](https://ai.exponencialadm.net/okf/en/radar/guia-pratico-de-c-bare-metal-69b9d2.md) - A resource focused on bare-metal C++ is recommended for engineers studying low-level programming. The source does not describe the guide’s technical contents.
* [Book introduces ARM64 Assembly programming](https://ai.exponencialadm.net/okf/en/radar/introducao-pratica-a-programacao-em-assembly-3f801f.md) - Published on May 29, 2026, Perry Kivolowitz's book introduces Assembly programming for the 64-bit ARM ISA, with practices and ideas that also apply to high-level languages.
* [Browser SSD timing may reveal user activity](https://ai.exponencialadm.net/okf/en/radar/timing-de-ssd-no-browser-pode-49964f.md) - Published on May 29, 2026, the report describes measuring SSD latency through the browser's Origin Private File System in an ordinary tab. It reports 88% accuracy in identifying visited sites and 95% in detecting launched desktop applications.
* [More C++ Idioms: a community reference](https://ai.exponencialadm.net/okf/en/radar/more-c-idioms-referencia-comunitaria-c3e770.md) - Published on May 28, 2026, the briefing describes a Wikibooks reference with more than 80 community-maintained C++ design patterns.
* [PaddleOCR-VL 1.6 reports 96.33% on OmniDocBench v1.6](https://ai.exponencialadm.net/okf/en/radar/paddleocr-vl-1-6-reporta-96-be6997.md) - A May 28, 2026 publication announces PaddleOCR-VL 1.6 and reports 96.33% on the OmniDocBench v1.6 benchmark, describing the result as a new state of the art.
* [Seasonic to show 5,200 W CRPS power supply at Computex](https://ai.exponencialadm.net/okf/en/radar/seasonic-anuncia-fonte-crps-de-5-d4175a.md) - Seasonic plans to present a 5,200 W CRPS power supply with 80 PLUS Ruby certification at Computex 2026, for engineers assessing power for high-consumption hardware.
* [Free tutorial covers modern C++, from C++11 to C++26](https://ai.exponencialadm.net/okf/en/radar/tutorial-aberto-de-c-moderno-do-e5137a.md) - Ou Changkun’s open-access book presents features from several modern C++ versions, spanning C++11 to C++26, as learning material for engineers. The page was published on May 28, 2026.
* [Arrays in C: a perspective on their types](https://ai.exponencialadm.net/okf/en/radar/arrays-em-c-uma-perspectiva-sobre-2ab81f.md) - An article presents what its author considers unusual about C array types, what they would do differently, and related topics.
* [Radar: HDD and SSD firmware analysis and modification](https://ai.exponencialadm.net/okf/en/radar/como-estudar-firmware-de-hdds-e-0ad680.md) - Published on May 27, 2026, the series covers dumps, reverse engineering, firmware modification for HDDs and SSDs, and JTAG debugging.
* [Radar: an introduction to linkers](https://ai.exponencialadm.net/okf/en/radar/linkers-guia-para-entender-o-linking-40c47f.md) - Published on May 27, 2026, this material introduces the linking process and its relevance to diagnosing build and linker problems in C and C++.
* [GPU service aims to withstand failures](https://ai.exponencialadm.net/okf/en/radar/servico-de-gpu-busca-resistir-a-c0a9ff.md) - Research describes fault isolation in a multi-process GPU service and VMM-based active-standby recovery; the authors report zero runtime overhead for the isolation mechanism.
* [Free C++ tutorials organized by topic](https://ai.exponencialadm.net/okf/en/radar/tutoriais-gratuitos-de-c-organizados-por-2422d2.md) - The publication presents free C++ tutorials grouped into categories, which can help engineers structure their learning of the language.
* [Under the Hood of C++: Class Internals](https://ai.exponencialadm.net/okf/en/radar/c-por-baixo-dos-panos-mecanismos-6cbcab.md) - A talk by Chris Ryan explores how C++ classes work internally. It is recommended for beginners, though experienced developers may also learn something from it.
* [An epoll use-after-free race fixed with RCU](https://ai.exponencialadm.net/okf/en/radar/rcu-adia-a-liberacao-em-uma-7f1509.md) - Published May 26, 2026, the post describes a use-after-free race in fs/eventpoll.c and commit 07712db80857, which replaces kfree() with kfree_rcu().
* [Cache: design, management and tradeoffs](https://ai.exponencialadm.net/okf/en/radar/cache-design-gerenciamento-e-tradeoffs-6240f4.md) - A video by Onur Mutlu covers cache design and management, topics that help explain system performance. The available description does not specify particular results or recommendations.
* [Large primes in LLVM IR: a video explores the calculation](https://ai.exponencialadm.net/okf/en/radar/como-explorar-numeros-primos-grandes-com-47055c.md) - A video published on May 25, 2026 explores how to calculate the largest possible prime number in one second, without precomputation, covering LLVM IR, sieves, number theory, and category theory.
* [How C++ Debuggers Work](https://ai.exponencialadm.net/okf/en/radar/como-funcionam-os-debuggers-de-c-7cd85f.md) - A video by Sy Brand introduces how C++ debuggers work internally and why understanding that can help engineers use debugging tools.
* [C, Extensions, and Portability Across Compilers](https://ai.exponencialadm.net/okf/en/radar/extensoes-de-c-e-os-limites-869e62.md) - Published May 25, 2026, the article covers C extensions, portability, and alternative compilers, noting reliance on nonstandard behavior in real-world code.
* [Linux Device Drivers: communicating with hardware](https://ai.exponencialadm.net/okf/en/radar/linux-device-drivers-uma-introducao-a-db00dd.md) - Published on May 25, 2026, the post recommends Linux Device Drivers as an introduction to how the Linux kernel communicates with hardware. It says readers should know some C and basic Unix system calls.
* [A first look at C++ patterns](https://ai.exponencialadm.net/okf/en/radar/primeiros-passos-com-patterns-em-c-736b78.md) - An introductory overview of design patterns and architectural patterns gives engineers a starting point for exploring software design patterns in C++.
* [Community recipes for serving LLMs on RTX GPUs](https://ai.exponencialadm.net/okf/en/radar/receitas-comunitarias-para-servir-llms-em-1c5745.md) - Published May 24, 2026, the repository collects model-independent community recipes for serving LLMs on RTX 3090, 4090, and 5090 GPUs, using vLLM, llama.cpp, and ik_llama.
* [TokenSpeed targets agentic workloads](https://ai.exponencialadm.net/okf/en/radar/tokenspeed-mira-workloads-agentic-93f914.md) - A May 24, 2026 entry describes an LLM inference engine for agentic workloads, highlighting scheduling, parallelism, and KV resource management features.
* [MC++ blog index gathers C++ articles](https://ai.exponencialadm.net/okf/en/radar/indice-do-blog-mc-reune-artigos-232ab3.md) - Published on May 23, 2026, the MC++ blog index gathers links to articles about C++ and may help engineers locate content.
* [Reports point to GPU Direct Storage plans for Vera Rubin](https://ai.exponencialadm.net/okf/en/radar/vera-rubin-e-o-plano-relatado-13e1ed.md) - A May 23, 2026 entry reports that NVIDIA plans GPU Direct Storage for Vera Rubin and identifies CPU-managed storage I/O as a possible bottleneck in GPU workloads.
* [MIT features a practical course on performance engineering](https://ai.exponencialadm.net/okf/en/radar/curso-do-mit-sobre-engenharia-de-ceda61.md) - Published on May 22, 2026, this entry describes MIT OpenCourseWare's 6.172 course on scalable, high-performance systems, covering performance analysis, optimization, caching, parallel programming, and C.
* [Virtual memory: concepts for beginners](https://ai.exponencialadm.net/okf/en/radar/memoria-virtual-conceitos-para-iniciantes-856c70.md) - Published on May 22, 2026, this article introduces paging, address translation, and the role of the MMU—concepts useful to C and C++ programmers.
* [C++ design patterns: common misconceptions](https://ai.exponencialadm.net/okf/en/radar/design-patterns-em-c-equivocos-comuns-7715bb.md) - A video by Klaus Iglberger discusses common misconceptions about design patterns in C++. Published on May 21, 2026, the listing describes a focused discussion for engineers.
* [Radar: three readings on memory safety](https://ai.exponencialadm.net/okf/en/radar/radar-tres-leituras-sobre-seguranca-de-e44f5a.md) - Published on May 21, 2026, this entry lists three materials on memory corruption and memory safety: a history of attacks, a work on memory errors past, present and future, and a page asking why memory safety remains a concern.
* [Community reference for modern C++ patterns](https://ai.exponencialadm.net/okf/en/radar/repositorio-comunitario-reune-patterns-de-c-b0dc4c.md) - Published on May 21, 2026, this bookmark presents a community repository of modern, idiomatic C++ code patterns for engineers to consult.
* [More than 700 C++ videos organized by topic](https://ai.exponencialadm.net/okf/en/radar/mais-de-700-videos-de-c-e1dc94.md) - Published on May 18, 2026, the briefing presents a Hacking C++ list of more than 700 educational videos, including conference talks, tutorials, courses, and explanations organized by subject.
* [Guide surveys 13 open-source tools for model deployment](https://ai.exponencialadm.net/okf/en/radar/13-ferramentas-open-source-para-implantar-3cd678.md) - Published on May 17, 2026, Turing Post’s guide covers 13 open-source tools for deploying, serving, and running foundation models, from local LLMs to high-throughput production inference.
* [llama.cpp adds Multi-Token Prediction support](https://ai.exponencialadm.net/okf/en/radar/llama-cpp-adiciona-suporte-a-multi-544a3d.md) - A llama.cpp pull request adds support for MTP heads. Its author reports that Qwen3.6-27B reached 65 tokens per second on an RTX 3090, versus 38 with MTP enabled.
* [Post alleges court-record extraction and WhatsApp automation](https://ai.exponencialadm.net/okf/en/radar/post-relata-extracao-de-processos-e-3d5e12.md) - Published on May 17, 2026, the post attributes court-case extraction, data enrichment and WhatsApp messaging to a tool; it also mentions HWID licensing and periodic contact with a central server.
* [MTP Benchmarks on Three GTX 1080 Ti GPUs](https://ai.exponencialadm.net/okf/en/radar/benchmarks-mtp-em-tres-gtx-1080-2ec120.md) - A May 16, 2026 publication reports llama.cpp throughput tests with MTP for Qwen 3.6 models on three GTX 1080 Ti GPUs.
* [Radar: Algorithmica's computer architecture guide](https://ai.exponencialadm.net/okf/en/radar/guia-de-arquitetura-de-computadores-da-a01903.md) - Published on May 16, 2026, Algorithmica's guide covers computer architecture and hardware topics relevant to performance. The publication recommends reading it alongside the book *Algorithms for Modern Hardware*.
* [Georgia Tech series on high-performance computer architecture](https://ai.exponencialadm.net/okf/en/radar/serie-da-georgia-tech-ensina-arquitetura-4a90b3.md) - A May 15, 2026 publication recommends a six-part series on high-performance computer architecture and its connection to fast GPU computing.
* [SIMD in C++ for data processing](https://ai.exponencialadm.net/okf/en/radar/simd-e-c-26-para-acelerar-e1238c.md) - A talk on SIMD-accelerated JSON parsing, C++26 compile-time reflection, and UTF-8/16 and Base64 processing in browsers.
* [Tile Scheduling in CuTeDSL](https://ai.exponencialadm.net/okf/en/radar/tile-scheduling-no-cutedsl-6699c3.md) - A post introduces StaticPersistentTileScheduler in CuTeDSL and explains its usefulness when writing persistent kernels.
* [Radar: CUDA C++ programming guide](https://ai.exponencialadm.net/okf/en/radar/como-estudar-o-guia-de-programacao-e37ebb.md) - A primary reference for engineers developing with CUDA C++; the entry is dated May 14, 2026.
* [Class notes on efficient LLM serving](https://ai.exponencialadm.net/okf/en/radar/notas-de-aula-sobre-serving-eficiente-49a690.md) - Published on May 13, 2026, these notes cover techniques and system components intended to improve language-model inference and serving.
* [An introduction to what the processor is doing](https://ai.exponencialadm.net/okf/en/radar/uma-introducao-ao-que-o-processador-61ad53.md) - The briefing recommends Kay Lack’s “Just Enough Assembly to Blow Your Mind” as an introduction to understanding what a processor is doing. Reading it may help engineers build a clearer mental model.
* [Stanford’s CS336 covers modern language-model systems](https://ai.exponencialadm.net/okf/en/radar/cs336-de-stanford-explora-sistemas-modernos-34f954.md) - The 2026 edition of CS336 covers MoEs, GPU tiling, kernels, RLHF, and data. According to the listing, lectures appear on YouTube with a delay of about two weeks.
* [Firefly AIBOX-K3 pairs RISC-V with edge AI hardware](https://ai.exponencialadm.net/okf/en/radar/firefly-aibox-k3-leva-risc-v-755843.md) - Published on May 12, 2026, the item describes an industrial edge AI computer with a SpacemiT K3 RISC-V SoC, up to 32 GB of RAM, and up to 512 GB of UFS storage.
* [A minimal time-sharing kernel in Zig for RISC-V](https://ai.exponencialadm.net/okf/en/radar/kernel-minimo-de-time-sharing-em-468800.md) - An article published on May 12, 2026 presents a walkthrough for creating a minimal time-sharing kernel in Zig for RISC-V on OpenSBI.
* [C2IntelFeeds collects C2 infrastructure feeds](https://ai.exponencialadm.net/okf/en/radar/c2intelfeeds-reune-indicadores-de-infraestrutura-c2-c6b160.md) - Published on May 11, 2026, C2IntelFeeds offers feeds generated from large-scale internet scans, covering suspected and verified C2 infrastructure indicators.
* [Fluxion proposes hybrid sparse attention for long-context inference](https://ai.exponencialadm.net/okf/en/radar/fluxion-explora-atencao-esparsa-hibrida-para-1efd8f.md) - Published on May 11, 2026, the summary presents Fluxion, a hybrid sparse-attention system for long-context inference that uses CPU and GPU parallelism.
* [Radar: virtual memory and Linux internals](https://ai.exponencialadm.net/okf/en/radar/memoria-virtual-page-tables-tlbs-e-8a9cf8.md) - Published on May 11, 2026, this bookmark describes a practical guide to virtual memory, page faults, page tables, TLBs, NUMA topology, Linux internals, and performance in data-intensive systems.
* [The mathematics of Query, Key, and Value in attention](https://ai.exponencialadm.net/okf/en/radar/a-matematica-de-query-key-e-9305bf.md) - Published on May 10, 2026, the article presents the central attention formula in transformers and walks through a numerical example step by step.
* [CuBridge uses LLMs to reconstruct attention kernels](https://ai.exponencialadm.net/okf/en/radar/cubridge-explora-llms-para-reconstruir-kernels-c97c25.md) - Published on May 10, 2026, the item describes CuBridge, an LLM-based framework intended to understand and reconstruct high-performance attention kernels.
* [Video offers an overview of GPU architecture](https://ai.exponencialadm.net/okf/en/radar/video-apresenta-uma-visao-geral-da-2e0f28.md) - Published on May 10, 2026, the video gives an overview of graphics card internals and provides context for engineers working on GPU software.
* [KV cache reduces repeated computation during inference](https://ai.exponencialadm.net/okf/en/radar/kv-cache-menos-recalculo-na-inferencia-e56e87.md) - Published on May 9, 2026, the article explains how a KV cache stores Key and Value data for processed tokens so they need not be recalculated for each new token.
* [LeWorldModel: pixel-based JEPA trained on one GPU](https://ai.exponencialadm.net/okf/en/radar/leworldmodel-treina-um-jepa-com-pixels-440ee0.md) - A May 9, 2026 publication reports training a 15-million-parameter JEPA from raw pixels on one GPU and compares its planning speed on control tasks.
* [Luce PFlash reports 2.89× faster TTFT at 64K context](https://ai.exponencialadm.net/okf/en/radar/luce-pflash-relata-ttft-2-89-78cf6f.md) - A May 9, 2026 publication reports that Luce PFlash achieved 2.89× faster time to first token than Ollama in a benchmark with a 64K context.
* [Lucebox explores speculative inference on consumer GPUs](https://ai.exponencialadm.net/okf/en/radar/lucebox-explora-inferencia-especulativa-em-gpus-d4ae37.md) - Published on May 9, 2026, this entry presents Lucebox, a speculative inference server for heterogeneous hardware and consumer GPUs. The post claims Qwen3.6-27B reaches 120–200 tokens per second on an RTX 3090.
* [RTX 3090: power and concurrency tests for LLMs](https://ai.exponencialadm.net/okf/en/radar/rtx-3090-eficiencia-e-potencia-em-f9eecc.md) - Published May 9, 2026, the report compares tests of Qwen3.6 27B on an RTX 3090 at concurrency levels of 1, 4, 8, and 16. The author reports that 225 W was most efficient, while 250 W delivered higher throughput.
* [TC-JEPA uses captions to guide visual representations](https://ai.exponencialadm.net/okf/en/radar/tc-jepa-usa-legendas-no-aprendizado-09c80e.md) - Published on May 9, 2026, the post describes a self-supervised visual learning method that uses captions to guide predictions of masked image patches.
* [RTX 3090 tests compare LLM throughput and efficiency](https://ai.exponencialadm.net/okf/en/radar/rtx-3090-throughput-e-eficiencia-em-664623.md) - A briefing published on May 8, 2026 compares eight local LLMs on one RTX 3090, at power limits from 100 W to 450 W. Average throughput rose at the higher tested power, while energy efficiency fell.
* [SAM in mid-training and retention after fine-tuning](https://ai.exponencialadm.net/okf/en/radar/sam-no-mid-training-pode-reduzir-684768.md) - A post published on May 8, 2026 proposes using SAM during the final roughly 10% of mid-training to reduce forgetting after fine-tuning or quantization.
* [Study explores geometric limits of compression in RAG](https://ai.exponencialadm.net/okf/en/radar/estudo-investiga-limites-geometricos-da-compressao-58adfc.md) - An article presented as a NeurIPS 2026 paper connects compression in RAG with embedding geometry and claims that a spectral quantity establishes a minimum bound on compression.
* [Mamba-3 compresses context into a fixed-size state](https://ai.exponencialadm.net/okf/en/radar/mamba-3-comprime-o-contexto-em-6d3dcb.md) - Published on May 7, 2026, the excerpt describes Mamba-3 as a non-Transformer language-model architecture with a fixed-size state. It claims constant time per decoding step and faster inference than a Transformer beyond an unspecified threshold.
* [Open ASR Leaderboard adds private evaluation](https://ai.exponencialadm.net/okf/en/radar/leaderboard-de-asr-adiciona-avaliacoes-privadas-156d4c.md) - On May 6, 2026, the Open ASR Leaderboard announced private evaluation data intended to reduce test-set contamination and overfitting in speech recognition benchmarks.
* [APUE offers Unix portability tables](https://ai.exponencialadm.net/okf/en/radar/apue-e-as-tabelas-de-portabilidade-8f53d8.md) - Published on May 5, 2026, this bookmark describes APUE tables comparing system calls across Linux, BSD, macOS, and Solaris, and summarizes the book's edition history.
* [Selective attention and subquadratic limits](https://ai.exponencialadm.net/okf/en/radar/atencao-seletiva-e-limites-subquadraticos-825689.md) - A publication highlights two papers: one on selective attention and another on the limitations of subquadratic methods, as context for assessing claims about these approaches.
* [Qwen3-ASR-Enhanced v0.1 announced in beta](https://ai.exponencialadm.net/okf/en/radar/checkpoint-beta-do-qwen3-asr-enhanced-f71aea.md) - A May 5, 2026 announcement presents an initial beta checkpoint for automatic speech recognition and notes a known limitation: support for non-verbal tags is not yet stable.
* [Beej’s Guide recommended for learning sockets](https://ai.exponencialadm.net/okf/en/radar/guia-de-beej-e-recomendado-para-5661e9.md) - A post recommends Beej’s Guide for studying sockets and also mentions the author’s guides to C, Unix IPC, and Git.
* [GRPO fine-tuning tutorial for extracting receipt data](https://ai.exponencialadm.net/okf/en/radar/tutorial-explora-grpo-para-extrair-dados-e9db73.md) - Published on May 4, 2026, the tutorial introduces GRPO and a fine-tuning workflow to extract receipt data through OCR into JSON, using LFM2.5-1.2B-Instruct and Unsloth. It includes a Kaggle notebook.
* [dflash-mlx v0.1.5 expands runtime and serving](https://ai.exponencialadm.net/okf/en/radar/dflash-mlx-v0-1-5-amplia-82a78d.md) - Released on May 3, 2026, version 0.1.5 adds tools for serving and generation, benchmarking, diagnostics, profiling, and model listing, as well as caching and experimental long-context support.
* [Triton kernels and PyTorch ops for AttnRes](https://ai.exponencialadm.net/okf/en/radar/kernels-triton-e-operadores-pytorch-para-a66fad.md) - A May 3, 2026 bookmark about a repository with Triton kernels and PyTorch operators for Block Attention Residuals (AttnRes).
* [Luce PFlash adds prefix caching and cold-start tuning](https://ai.exponencialadm.net/okf/en/radar/luce-pflash-cache-de-prefixo-e-f9457d.md) - An update to a speculative inference server reports prefix caching, cold-start tuning, and a CUDA VMM fix. The author reports performance gains for Qwen3.6 27B.
* [Project implements microGPT in dependency-free C](https://ai.exponencialadm.net/okf/en/radar/microgpt-em-c-puro-explora-otimizacoes-830f52.md) - Published on May 3, 2026, the project presents a pure C implementation of Karpathy’s microGPT, without dependencies, using low-level optimization techniques.
* [A reference for Linux programming in C](https://ai.exponencialadm.net/okf/en/radar/uma-referencia-para-programacao-linux-em-fd8614.md) - The May 2, 2026 entry recommends The Linux Programming Interface as a reference on Linux interfaces and systems programming in C.
* [Recursive agents collaborate through latent states](https://ai.exponencialadm.net/okf/en/radar/agentes-recursivos-colaboram-trocando-estados-latentes-4867a4.md) - A paper described in a publication reports more precise collaboration, faster execution, and lower token use when agents refine shared latent states.
* [Lucebox explores speculative prefill for LLMs](https://ai.exponencialadm.net/okf/en/radar/lucebox-investiga-speculative-prefill-para-llms-8e7497.md) - Published on May 1, 2026, the post introduces Lucebox, a speculative inference server for heterogeneous hardware and consumer GPUs.
* [Coding-agent harnesses improve over ten iterations](https://ai.exponencialadm.net/okf/en/radar/como-evoluir-harnesses-de-coding-agents-65a1a9.md) - A paper describes a framework using observable components, condensed trajectory evidence, and decisions tested against task results. On Terminal-Bench 2, pass@1 rises from 69.7% to 77.0% after ten iterations.
* [Shared-KV attention: the architecture change described](https://ai.exponencialadm.net/okf/en/radar/mqa-de-alta-ordem-e-o-717168.md) - A publication dated April 29, 2026 compares an attention architecture with MQA and shared KV to MLA, noting a trade-off between cache efficiency and computation per token.
* [Post estimates hardware needs for a quantized model](https://ai.exponencialadm.net/okf/en/radar/post-estima-modelo-quantizado-em-duas-ee4fd7.md) - Published on April 29, 2026, the post claims a language model's experts are in 4-bit precision and that it could fit on two RTX Pro 6000 GPUs, while noting missing SM120 support in vLLM or SGLang.
* [An implementation of compressed sparse attention](https://ai.exponencialadm.net/okf/en/radar/implementacao-de-atencao-esparsa-comprimida-dff71b.md) - The page presents an implementation described as faithful to Compressed Sparse Attention (CSA), a sparse attention technique presented in a recent paper.
* [CUDA/PTX kernel beats cuBLAS on B200](https://ai.exponencialadm.net/okf/en/radar/kernel-cuda-ptx-relata-ganho-de-ab2780.md) - In a post dated April 27, 2026, Paul Chan says a custom B200 matrix multiplication kernel was 6% faster than cuBLAS in one specific configuration.
* [Guide surveys LoRA variants for efficient fine-tuning](https://ai.exponencialadm.net/okf/en/radar/como-comparar-variantes-de-lora-no-b05039.md) - Published on April 26, 2026, the guide presents more than 15 LoRA variants and outlines the evolution of parameter-efficient fine-tuning for large language models.
* [Ouroboros tests input-conditioned LoRA in recursive transformers](https://ai.exponencialadm.net/okf/en/radar/ouroboros-testa-lora-adaptada-a-cada-29630a.md) - Published on April 25, 2026, the work explores a compact controller that modulates weights at each step in a shared transformer block. On Qwen2.5-3B, the authors report lower training loss than a 17-layer baseline, with gains observed in-distribution.
* [PARE-Bench evaluates proactive agents in apps](https://ai.exponencialadm.net/okf/en/radar/pare-bench-avalia-agentes-proativos-em-d7f91d.md) - Published on April 25, 2026, PARE-Bench contains 143 tasks for testing agents in stateful app simulations, including goal inference, intervention timing, and orchestration across apps.
* [pyptx brings a Python DSL to NVIDIA PTX kernels](https://ai.exponencialadm.net/okf/en/radar/pyptx-leva-uma-dsl-python-aos-060c86.md) - Published on April 25, 2026, pyptx is described as a Python DSL for writing NVIDIA PTX kernels, with Hopper and Blackwell support and integration with JAX and PyTorch.
* [StructMem proposes temporal memory for LLM agents](https://ai.exponencialadm.net/okf/en/radar/structmem-propoe-memoria-temporal-para-agentes-9a8a58.md) - Published on April 24, 2026, the StructMem paper proposes hierarchical memory for long-term conversational agents, targeting event relations, temporal reasoning, and multi-hop questions.
* [DFlash and DDTree run Qwen3.6-27B on an RTX 3090](https://ai.exponencialadm.net/okf/en/radar/dflash-e-ddtree-levam-qwen3-6-6ec2d7.md) - A post published April 23, 2026 reports 73 tokens per second for Qwen3.6-27B on a single RTX 3090, using speculative decoding with DFlash and DDTree. The author says the stack loads the model because its architecture and layer and head dimensions match Qwen3.5, but throughput is lower.
* [GPU kernels for LLM operations](https://ai.exponencialadm.net/okf/en/radar/kernels-de-gpu-para-operacoes-de-bd0eb2.md) - On April 23, 2026, a company announced GPU kernels for language-model operations, developed with TileLang. The company says some are already used internally for training and inference.
* [Megatron-LM: model parallelism for large-model training](https://ai.exponencialadm.net/okf/en/radar/megatron-lm-e-o-paralelismo-de-9dcaa4.md) - A 2019 paper describes splitting the weights of language models with billions of parameters across GPUs, addressing models whose weights do not fit on a single GPU.
* [Fil-C aims to add memory safety to C/C++ builds](https://ai.exponencialadm.net/okf/en/radar/fil-c-busca-seguranca-de-memoria-8b8f91.md) - A report published on April 22, 2026 describes Fil-C as a compiler for C/C++ projects with memory safety and notes that direct assembly calls may need alternatives.
* [LRT explores reasoning with compact representations](https://ai.exponencialadm.net/okf/en/radar/lrt-busca-raciocinio-com-representacoes-compactas-55f23e.md) - Published on April 22, 2026, the post introduces Latent Reasoning Tuning (LRT), an approach that trains models to reason with compact internal representations rather than long text sequences.
* [What limits knowledge in attention and MLPs](https://ai.exponencialadm.net/okf/en/radar/o-que-limita-o-conhecimento-em-956b4b.md) - A study examines how attention matrices and MLPs store knowledge, highlighting limits tied to data, finite compute, and noise in long contexts.
* [TACO explores context compression for terminal agents](https://ai.exponencialadm.net/okf/en/radar/taco-aprende-regras-para-comprimir-contexto-288139.md) - A framework discovers and refines rules for compressing context from terminal-agent interactions. The paper reports tests on three benchmarks.
* [Loop transformers and energy-based inference](https://ai.exponencialadm.net/okf/en/radar/transformers-com-loop-e-a-interpretacao-081241.md) - Published on April 22, 2026, the text links residual updates in looped transformers to gradient descent on an energy function, subject to a specific condition on the block.
* [HybridGen coordinates CPU and GPU for long contexts](https://ai.exponencialadm.net/okf/en/radar/hybridgen-coordena-cpu-e-gpu-para-182d8e.md) - The proposed framework enables collaborative attention between CPU and GPU, using CXL memory to support long-context language-model inference.
* [Radar: weekly roundup features research on AI agents](https://ai.exponencialadm.net/okf/en/radar/pesquisa-recente-reune-estudos-sobre-agentes-26b6de.md) - Published on April 21, 2026, the roundup highlights work on agentic system design for Claude Code, long-term tasks, memory transfer in coding agents, distillation, and other topics.
* [CliffordNet explores Clifford algebra in neural networks](https://ai.exponencialadm.net/okf/en/radar/cliffordnet-explora-algebra-de-clifford-em-e78fd0.md) - The architecture combines inner-product and geometric interactions without attention or mixer layers; the report gives 77.82% accuracy on CIFAR-100 with 1.4 million parameters.
* [LACE coordinates parallel reasoning paths](https://ai.exponencialadm.net/okf/en/radar/lace-coordena-trajetorias-paralelas-de-raciocinio-523233.md) - Published on April 20, 2026, the brief describes LACE, a framework that lets parallel paths share intermediate ideas and correct one another through cross-thread attention.
* [GNU C Library Reference Manual](https://ai.exponencialadm.net/okf/en/radar/manual-de-referencia-da-gnu-c-1ced1f.md) - Published on April 20, 2026, this item points to a PDF copy of the GNU C Library Reference Manual, intended for engineers working with the library.
* [Skill-RAG adapts retrieval to knowledge gaps](https://ai.exponencialadm.net/okf/en/radar/skill-rag-direciona-buscas-conforme-lacunas-2c0a36.md) - Published on April 20, 2026, the paper presents a method that detects possible knowledge gaps and routes queries to specialized retrieval strategies. Its evaluation covers three question-answering datasets.
* [Protocol proposes auditable self-improvement for agents](https://ai.exponencialadm.net/okf/en/radar/agentes-que-melhoram-com-rastreabilidade-e-4879bc.md) - Published on April 19, 2026, this item summarizes a paper proposing audit and rollback mechanisms for agents that suggest, evaluate, and implement improvements.
* [AutoSOTA automates iterative AI research](https://ai.exponencialadm.net/okf/en/radar/autosota-automatiza-ciclos-de-pesquisa-em-891123.md) - Published on April 19, 2026, the report describes AutoSOTA, a system with eight specialized agents that reproduces and improves models from AI papers.
* [Guide compares attention mechanisms in LLMs](https://ai.exponencialadm.net/okf/en/radar/como-comparar-mecanismos-de-atencao-em-84e720.md) - An article published on April 19, 2026, presents more than 13 attention mechanisms and their role in how Transformers and LLMs process context and memory.
* [Compute and memory costs in loop models](https://ai.exponencialadm.net/okf/en/radar/como-estimar-custos-de-treino-de-e588ff.md) - Published on April 19, 2026, the article analyzes seven loop model variants and how gradient structure affects FLOPs and memory during training.
* [ML-Master 2.0 reports layered memory](https://ai.exponencialadm.net/okf/en/radar/ml-master-2-0-organiza-memoria-56878d.md) - A post published on April 19, 2026 describes Hierarchical Cognitive Caching, with short-, medium-, and long-term memory for research across experiments and sessions. The team reports a 56.44% medal rate on MLE-Bench after a 24-hour run.
* [MoDA connects attention across layers](https://ai.exponencialadm.net/okf/en/radar/moda-conecta-a-atencao-entre-camadas-2fab2e.md) - An article published on April 19, 2026 presents Mixture-of-Depths Attention (MoDA), a mechanism that lets attention heads access key-value pairs from the current layer and depth key-value pairs from earlier layers.
* [Transformers in a loop as programmable computers](https://ai.exponencialadm.net/okf/en/radar/transformers-em-loop-podem-implementar-computacao-0f44e3.md) - Published on April 19, 2026, the article presents a construction in which transformers with specific weights, placed in a loop, can emulate basic computation blocks using a constant number of encoder layers.
* [Why RL scaling laws differ](https://ai.exponencialadm.net/okf/en/radar/por-que-rl-escala-de-forma-cd714f.md) - A publication dated April 18, 2026 compares pretraining and reinforcement learning (RL) scaling, highlighting compute costs and evaluation challenges.
* [A series on memory allocation strategies](https://ai.exponencialadm.net/okf/en/radar/uma-serie-sobre-estrategias-de-alocacao-2055f9.md) - gingerBill's series covers memory allocation strategies and can serve as a starting point for engineers comparing different approaches.
* [GP-SIMD and Processing-in-Memory](https://ai.exponencialadm.net/okf/en/radar/gp-simd-e-processamento-na-memoria-ae1f3f.md) - A post points to an ACM article about GP-SIMD, a topic of interest to engineers exploring GPU-style SIMD computing and processing-in-memory.
* [RoPE: an explanation of position in models](https://ai.exponencialadm.net/okf/en/radar/rope-uma-explicacao-sobre-posicao-em-0eb59c.md) - A post points to an explanatory article about rotary positional embeddings (RoPE), useful for engineers studying positional encoding in model design.
* [Ten C projects for practicing systems programming](https://ai.exponencialadm.net/okf/en/radar/dez-projetos-em-c-para-praticar-bd347c.md) - Published on April 11, 2026, the material lists ten C projects in increasing order of difficulty to explore systems programming and computing fundamentals.
* [Video covers pipeline and out-of-order execution](https://ai.exponencialadm.net/okf/en/radar/execucao-pipeline-e-fora-de-ordem-7191d4.md) - Dated April 11, 2026, this item points to a BitLemon video about pipeline execution, out-of-order execution, and Tomasulo's algorithm.
* [Interleaved Head Attention explores communication between heads](https://ai.exponencialadm.net/okf/en/radar/interleaved-head-attention-investiga-comunicacao-entre-8e9214.md) - Published on April 11, 2026, the post presents an architecture question: whether attention heads can benefit from communicating within a single layer.
* [Linux kernel documents x86 bus lock detection](https://ai.exponencialadm.net/okf/en/radar/linux-documenta-deteccao-de-bus-locks-a5d9f1.md) - Published on April 11, 2026, Linux kernel documentation explains detecting and handling bus locks on x86, including the role of split lock operations.
* [Floating-point arithmetic from scratch](https://ai.exponencialadm.net/okf/en/radar/aritmetica-de-ponto-flutuante-do-zero-c1bbdc.md) - An article presents implementing floating-point arithmetic from scratch as a challenging project and offers engineers a practical view of its mechanics.
* [Low-latency C++: patterns and performance](https://ai.exponencialadm.net/okf/en/radar/padroes-de-c-para-aplicacoes-de-4d88fa.md) - Published on April 7, 2026, this material introduces design patterns for low-latency C++ applications, including HFT, and mentions performance optimizations associated with advanced compiler and operating-system courses.
* [64-bit pointers and bindless access](https://ai.exponencialadm.net/okf/en/radar/acesso-bindless-e-ponteiros-de-64-0c0bfb.md) - An April 6, 2026 post compares buffer access approaches in GPU APIs and says 64-bit pointers avoid looking up descriptors in the heap.
* [Ultrafast trading in C++ starts with design](https://ai.exponencialadm.net/okf/en/radar/baixa-latencia-em-trading-comeca-no-c58923.md) - A CppCon 2024 talk by David Gross discusses ultrafast trading systems in C++. Its central point: early design and architecture decisions play an important role in achieving low latency.
* [SHINE maps context to LoRA adapters](https://ai.exponencialadm.net/okf/en/radar/shine-mapeia-contexto-para-adapters-lora-783d2c.md) - Published on April 6, 2026, the paper presents SHINE, a hypernetwork that maps varied contexts to LoRA adapters for LLMs. Its design reuses frozen base-model parameters and includes pretraining and instruction fine-tuning stages.
* [Playbook outlines large-scale LLM training practices](https://ai.exponencialadm.net/okf/en/radar/um-playbook-para-escalar-o-treinamento-5c6bf0.md) - Published on April 6, 2026, the material covers data, expert, tensor, pipeline, and context parallelism in GPU clusters, with findings from 4,000 scaling experiments using up to 512 GPUs.
* [FAQ on managing the C++ free store](https://ai.exponencialadm.net/okf/en/radar/faq-sobre-o-gerenciamento-do-free-81ad04.md) - A Standard C++ FAQ provides a reference on free store memory management for engineers working with memory allocation in C++.
* [A 2012 article discusses integer overflow in C and C++](https://ai.exponencialadm.net/okf/en/radar/integer-overflow-em-c-e-c-09fd4e.md) - A 2012 article, available as a PDF on the University of Utah website, covers integer overflow in C and C++, a relevant topic when writing and reviewing code in these languages.
* [Revised article explores fat pointers for C](https://ai.exponencialadm.net/okf/en/radar/exemplos-de-fat-pointers-em-c-60388b.md) - Published on April 4, 2026, the article attributed to Walter Bright presents C fat-pointer examples as an approach to addressing array-overflow bugs.
* [Jump threading: an accessible introduction](https://ai.exponencialadm.net/okf/en/radar/jump-threading-uma-introducao-acessivel-6c4cf5.md) - A Red Hat Developer article introduces jump threading optimizations in GCC and their goal: reducing dynamically executed jumps across different control-flow paths.
* [ELF: the Linux binary format](https://ai.exponencialadm.net/okf/en/radar/elf-entendendo-o-formato-dos-binarios-d7c70b.md) - A publication dated April 1, 2026 points to the ELF specification, which describes the format of Linux executables, object code, shared libraries, and core dumps.
* [View a Clang AST in Compiler Explorer](https://ai.exponencialadm.net/okf/en/radar/visualize-uma-ast-do-clang-no-11a788.md) - Published on April 1, 2026, the example shows a short C++ program compiled with Clang and its AST displayed in Compiler Explorer.
* [Dive into Systems: C, optimization and assembly](https://ai.exponencialadm.net/okf/en/radar/dive-into-systems-c-otimizacao-e-4ff9a1.md) - A free book on systems programming with C, code optimization and assembly across multiple architectures. A resource for studying how these topics connect.
* [WinDbg MCP exposes debugger functions to MCP clients](https://ai.exponencialadm.net/okf/en/radar/windbg-mcp-transforma-funcoes-de-depuracao-8c07a7.md) - A March 31, 2026 publication describes a server exposing Windows pybag debugger operations as tools for MCP-compatible clients, including process and crash-dump inspection.
* [ASM Visualizer: viewing assembly code](https://ai.exponencialadm.net/okf/en/radar/asm-visualizer-visualizar-codigo-assembly-b8f641.md) - ASM Visualizer is an assembly-code visualization tool that may help engineers inspect compiler output.
* [Compiling to Assembly from Scratch](https://ai.exponencialadm.net/okf/en/radar/compilando-para-assembly-do-zero-f4ea9e.md) - Published on March 30, 2026, Vladimir Keleshev’s article explores how a compiler can generate assembly code.
* [How to Design an ISA: Why It Matters](https://ai.exponencialadm.net/okf/en/radar/como-projetar-uma-isa-por-que-a40377.md) - An article by David Chisnall, published by ACM Queue, on instruction set architecture (ISA) design—a relevant topic for professionals working with compilers and performance.
* [Memory and performance: essential reading for programmers](https://ai.exponencialadm.net/okf/en/radar/memoria-e-desempenho-leitura-essencial-para-43d43f.md) - An article by Ulrich Drepper provides context for engineers working with memory behavior and software performance.
* [Radar: C++ multithreading mistakes](https://ai.exponencialadm.net/okf/en/radar/multithreading-em-c-erros-comuns-e-a604df.md) - Published on March 29, 2026, the material promises to cover 20 common C++ concurrency errors, with a focus on C++20/23.
* [Optimizing an SPSC lock-free ring buffer](https://ai.exponencialadm.net/okf/en/radar/como-otimizar-um-ring-buffer-spsc-58c2d4.md) - The article describes optimizing an SPSC ring buffer, from replacing mutexes with atomics to cache-focused adjustments, and reports performance at each stage.
* [DATA and Microwalk: a comparison for constant-time analysis](https://ai.exponencialadm.net/okf/en/radar/data-e-microwalk-comparacao-na-analise-9428b5.md) - An article compares DATA and Microwalk for detecting constant-time violations in cryptographic libraries. The March 28, 2026 entry does not report specific comparison results.
* [eBPF: introduction and learning resources](https://ai.exponencialadm.net/okf/en/radar/ebpf-no-linux-introducao-e-recursos-2191ba.md) - Published on March 28, 2026, the material presents eBPF as a way to run sandboxed programs in the Linux kernel without changing its source code or loading a module.
* [Introduction to GPU Architecture and Computing](https://ai.exponencialadm.net/okf/en/radar/introducao-a-arquitetura-e-a-computacao-c33507.md) - An introductory article provides context on GPU architecture and computing for engineers who develop or optimize software for this type of processor.
* [Assembly and intrinsics in C and C++](https://ai.exponencialadm.net/okf/en/radar/assembly-ou-intrinsics-escolhas-de-baixo-d28092.md) - Dan Luu’s article covers assembly and intrinsics, topics relevant to engineers assessing low-level implementation choices in C and C++.
* [Overlapping data transfers in CUDA C/C++](https://ai.exponencialadm.net/okf/en/radar/como-sobrepor-transferencias-e-computacao-em-31d6ed.md) - A Mark Harris post on the NVIDIA Technical Blog, published March 27, 2026, explains overlapping data transfers and computation in CUDA C/C++ and how this may help engineers make better use of CPU and GPU time.
* [CppCon 2025 explores JSON with C++26 reflection](https://ai.exponencialadm.net/okf/en/radar/cppcon-2025-explora-json-com-reflection-61f59c.md) - A CppCon 2025 talk presents JSON serialization and deserialization in C++ using reflection features planned for C++26 and the standard library, without macros or manual mapping.
* [Binary quantization may shrink embeddings to 128 bytes](https://ai.exponencialadm.net/okf/en/radar/quantizacao-binaria-pode-reduzir-embeddings-a-6fe3c9.md) - A post published on March 19, 2026 describes representing each embedding dimension with one bit instead of a 32-bit value. In its example, a vector shrinks from 4,096 bytes to 128 bytes.
* [Overview of RL methods for reasoning LLMs](https://ai.exponencialadm.net/okf/en/radar/como-comparar-metodos-de-rl-para-586c58.md) - Published on March 16, 2026, the article collects reinforcement learning methods used with language models and offers a starting point for engineers to compare them.
* [DeepCrossAttention combines layers based on input](https://ai.exponencialadm.net/okf/en/radar/deepcrossattention-combina-camadas-conforme-a-entrada-b97480.md) - Published on March 16, 2026, the brief describes DeepCrossAttention (DCA), an approach that learns input-dependent weights to combine Transformer layer outputs as an alternative to traditional residual addition.
* [Why low-precision Flash Attention training can fail](https://ai.exponencialadm.net/okf/en/radar/flash-attention-em-baixa-precisao-por-e729c1.md) - Published on March 15, 2026, this paper summary examines catastrophic loss spikes and identifies similar attention-data patterns and biased rounding errors as possible causes.
* [Article reports gains from exclusive attention](https://ai.exponencialadm.net/okf/en/radar/atencao-exclusiva-altera-o-calculo-entre-8481fe.md) - Published on March 12, 2026, the article presents an attention variant that excludes each token’s own value vector and reports better performance in language models of up to 2.7 billion parameters.
* [C++20: An Introduction to Ranges and Views](https://ai.exponencialadm.net/okf/en/radar/primeiros-passos-com-c-20-ranges-9c80b7.md) - Published on March 12, 2026, this item presents an introductory article on C++20 ranges and views for programmers beginning to learn these features.
* [Intel updates AVX10.2 to version 6.0](https://ai.exponencialadm.net/okf/en/radar/avx10-2-chega-a-versao-6-251daf.md) - Intel published version 6.0 of the AVX10.2 architecture specification and renamed AVX10_VNNI_INT to AVX10_V1_AUX; the change may affect compiler and hardware support.
* [ONNX: CPU and GPU performance compared](https://ai.exponencialadm.net/okf/en/radar/cpu-e-gpu-na-inferencia-com-e705c5.md) - An article published on March 11, 2026 compares ONNX Runtime inference on an NVIDIA L4 GPU and an AMD EPYC 9965 CPU; the author reports a CPU advantage in the described scenario.
* [FlashAttention-4 targets NVIDIA Blackwell GPUs](https://ai.exponencialadm.net/okf/en/radar/flashattention-4-mira-gpus-nvidia-blackwell-4d6da4.md) - Published on March 11, 2026, the note describes new pipelines, reformulated computations, and memory optimizations for FlashAttention-4 on Blackwell GPUs.
* [Review compares algorithms for float conversion](https://ai.exponencialadm.net/okf/en/radar/conversao-de-floats-em-decimais-curtos-6dfc05.md) - An article preview compares algorithms that convert binary floating-point values into short decimal strings. Schubfach and Dragonbox reach up to ten times the speed of Dragon4, according to the summary.
* [250+ notebooks for fine-tuning and RL](https://ai.exponencialadm.net/okf/en/radar/notebooks-para-fine-tuning-e-rl-26f31f.md) - Published on March 10, 2026, the repository collects more than 250 notebooks with fine-tuning and reinforcement-learning examples for models across several modalities.
* [FP4 training of MoE models on Hopper GPUs](https://ai.exponencialadm.net/okf/en/radar/treinamento-fp4-de-modelos-moe-em-b4b4ce.md) - A practical discussion of FP4 training for large-scale Mixture-of-Experts models on Hopper GPUs identifies activation memory and expert-parallel communication as bottlenecks.
* [Radar: designing a SIMD algorithm from scratch](https://ai.exponencialadm.net/okf/en/radar/como-projetar-um-algoritmo-simd-do-06c4a3.md) - Published on March 7, 2026, the article discusses designing a SIMD algorithm from scratch, focusing on Base64, and offers a concrete example for engineers.
* [FEVM prepares slim PC with Core Ultra 300 and 10GbE](https://ai.exponencialadm.net/okf/en/radar/fevm-prepara-pc-fino-com-core-530050.md) - A March 7, 2026 bulletin says FEVM is preparing an ultrathin PC with Intel Core Ultra 300, 10GbE networking, and three M.2 slots. The configuration may interest engineers evaluating compact systems with fast networking and multiple NVMe SSDs.
* [Practical bare-metal C++ guide for engineers](https://ai.exponencialadm.net/okf/en/radar/guia-pratico-de-c-bare-metal-a67e09.md) - A resource by Alex Robenko for engineers working with C++ in bare-metal environments.
* [Legal RAG Bench tests retrieval in criminal law](https://ai.exponencialadm.net/okf/en/radar/legal-rag-bench-mede-recuperacao-em-cd02d8.md) - A paper published on March 7, 2026 presents a benchmark with 4,876 criminal-law passages and 100 complex questions written by legal experts to assess retrieval in legal RAG systems.
* [C++ optimization techniques for modern CPUs](https://ai.exponencialadm.net/okf/en/radar/otimizacao-em-c-para-aproveitar-cpus-311cbe.md) - Published on March 7, 2026, the post highlights Fedor G. Pikus’s C++Now 2024 presentation on C++ optimization and points readers to the video and slides.
* [FireRedVAD detects voice and audio events in over 100 languages](https://ai.exponencialadm.net/okf/en/radar/fireredvad-detecta-voz-e-eventos-em-3429cc.md) - Published on March 5, 2026, the FireRedVAD GitHub project describes voice activity and audio event detection, with stated support for more than 100 languages.
* [LightRetriever speeds up LLM-based text retrieval](https://ai.exponencialadm.net/okf/en/radar/lightretriever-acelera-recuperacao-de-texto-com-afcc79.md) - An article published on March 5, 2026 reports that LightRetriever encodes queries 1,000 times faster and achieves 10 times the throughput while retaining 95% of benchmark performance.
* [OpenVul framework targets vulnerability detection with LLMs](https://ai.exponencialadm.net/okf/en/radar/openvul-framework-aberto-para-deteccao-de-4f01e3.md) - Published on March 5, 2026, OpenVul is described as an open-source post-training framework for vulnerability detection with LLMs. The description points to the project repository but reports no evaluation results.
* [TCL announces 32X3A OLED+ 4K 240 Hz monitor](https://ai.exponencialadm.net/okf/en/radar/tcl-anuncia-monitor-32x3a-oled-com-32b5b3.md) - On March 5, 2026, TCL announced the 32X3A, an OLED+ monitor with 4K resolution, a 240 Hz refresh rate, and an ultrathin design.
* [Linux kernel security collection receives updates](https://ai.exponencialadm.net/okf/en/radar/colecao-sobre-seguranca-do-kernel-linux-21cf30.md) - Published on March 4, 2026, the notice says a collection of resources on Linux kernel security and exploitation received updates for January and February.
* [CUDA warp primitives](https://ai.exponencialadm.net/okf/en/radar/como-primitivas-de-warp-ajudam-a-cb1195.md) - Published on March 4, 2026, the NVIDIA Technical Blog article introduces warp execution in the SIMT model and its use in CUDA programs.
* [Querying 3 Billion Vectors: requirements at scale](https://ai.exponencialadm.net/okf/en/radar/consultando-3-bilhoes-de-vetores-o-bbf169.md) - Published on March 4, 2026, the article discusses querying 3 billion vectors. Its preview says the requirements are difficult and presents the topic as a starting point for considering large-scale vector retrieval.
* [CUDA Agent applies reinforcement learning to kernel generation](https://ai.exponencialadm.net/okf/en/radar/cuda-agent-aplica-reinforcement-learning-a-0cd180.md) - A large-scale agentic reinforcement learning system aims to generate high-performance CUDA kernels, a task that requires specialized hardware expertise.
* [Fibers and runtime stack expansion](https://ai.exponencialadm.net/okf/en/radar/fibers-e-expansao-de-stack-em-17b63f.md) - Raymond Chen’s article discusses using fibers to expand a thread’s stack at runtime. Part 1 covers a simple return value.
* [LoRA as memory: auditing scale and saturation](https://ai.exponencialadm.net/okf/en/radar/lora-como-memoria-o-que-medir-8ea8e0.md) - A March 4, 2026 post summarizes a paper on how knowledge memory stored by LoRA scales and saturates as rank changes.
* [SIGReg chunking cuts LeJEPA VRAM use](https://ai.exponencialadm.net/okf/en/radar/sigreg-em-chunks-reduz-vram-no-43bd34.md) - A report published on March 4, 2026 says splitting 256 SIGReg slices into chunks of 32 used eight times less VRAM and finished faster on a 4090 MaxQ GPU.
* [Tutorial introduces subgroups in Vulkan 1.1](https://ai.exponencialadm.net/okf/en/radar/tutorial-apresenta-subgroups-no-vulkan-1-c5400d.md) - Khronos material explains how subgroups let parallel GPU tasks share and manipulate data, and invites comparison with shared-memory-based approaches.
* [Advent of Compiler Optimizations 2025 Archive](https://ai.exponencialadm.net/okf/en/radar/arquivo-do-advent-of-compiler-optimizations-df79bf.md) - Matt Godbolt’s blog collects an archive of Advent of Compiler Optimizations 2025, which may help C and C++ engineers explore compiler optimizations.
* [A fundamental introduction to x86 Assembly programming](https://ai.exponencialadm.net/okf/en/radar/assembly-x86-um-ponto-de-partida-62f6a5.md) - Published on March 3, 2026, Nayuki’s material introduces x86 Assembly as a learning resource and notes its relevance to C and C++ professionals analyzing low-level code and performance.
* [Vulkan memory barriers and image layouts](https://ai.exponencialadm.net/okf/en/radar/barreiras-de-memoria-e-layouts-de-39efcb.md) - A technical post covers memory barriers and image layouts in Vulkan, for engineers working on GPU memory synchronization.
* [Limitations of ultrasonic microphone blockers](https://ai.exponencialadm.net/okf/en/radar/bloqueadores-ultrassonicos-limites-a-avaliar-360c7a.md) - A March 3, 2026 publication questions claims about a portable audio blocker and highlights limits in device detection, range, and acoustic propagation.
* [LoRA as knowledge memory: an empirical analysis](https://ai.exponencialadm.net/okf/en/radar/lora-como-memoria-de-conhecimento-limites-68ab03.md) - Published on March 3, 2026, the article examines LoRA adapters as parametric, modular memory for updating pretrained language models, including capacity, scalability, and interference when merging modules.
* [Atomic operations and containment in coherent caches](https://ai.exponencialadm.net/okf/en/radar/operacoes-atomicas-e-contencao-em-caches-6ead51.md) - Published on March 3, 2026, Fabian Giesen’s article discusses primitives for building systems on coherent caches and follows an earlier text about how cache coherence works.
* [Inside NVIDIA’s Blackwell GPU](https://ai.exponencialadm.net/okf/en/radar/por-dentro-da-gpu-blackwell-da-faf4d0.md) - An article examines the Blackwell GPU and relates it to NVIDIA’s tradition of building large GPUs. The hardware context may help engineers understand the architecture behind their workloads.
* [Book examines inference engineering in production](https://ai.exponencialadm.net/okf/en/radar/um-guia-para-estudar-engenharia-de-9b4383.md) - Filed on March 3, 2026, this entry describes Philip Kiely’s book on the hardware, software, techniques, and infrastructure for running AI models in production, focused mainly on LLMs with some diffusion content.
* [AMD announces Ryzen AI PRO 400 desktop processors](https://ai.exponencialadm.net/okf/en/radar/amd-anuncia-ryzen-ai-pro-400-bc9679.md) - On March 2, 2026, AMD announced new Ryzen AI PRO 400 series products at Mobile World Congress, including desktop processors. The notice says the announcement is relevant to engineers tracking desktop processor options, but gives no specifications.
* [Radar: how libc++ implements std::string](https://ai.exponencialadm.net/okf/en/radar/como-o-libc-implementa-std-string-c30000.md) - Published on March 2, 2026, Joel Laity’s article examines the implementation of std::string in libc++. The topic may help engineers assess string behavior and performance.
* [Radar: x86-64 stack frame layout](https://ai.exponencialadm.net/okf/en/radar/entenda-o-layout-de-stack-frames-99e6a0.md) - An article by Eli Bendersky, published on March 2, 2026, about x86-64 stack frame layout and its relation to function calls.
* [Dynamic vs. static linking: costs and choices](https://ai.exponencialadm.net/okf/en/radar/linking-dinamico-ou-estatico-como-avaliar-a238f2.md) - Jordan Rose’s article examines the balance between startup time and disk space in dynamic and static linking, and notes that there is still room for improvement.
* [Parallel programming: challenges and approaches](https://ai.exponencialadm.net/okf/en/radar/programacao-paralela-desafios-e-caminhos-para-2fdd74.md) - Published on March 2, 2026, this entry presents a book by Paul McKenney on the challenges of parallel programming and what can be done about them. It is aimed at engineers seeking guidance on the subject.
* [RLHF with verl on AMD GPUs and ROCm 7.0](https://ai.exponencialadm.net/okf/en/radar/rlhf-com-verl-em-gpus-amd-59ad83.md) - In a post dated March 2, 2026, AMD describes deploying verl for RLHF training on AMD GPUs, with ROCm optimizations and Docker scripts. The article presents throughput and convergence results, but the available source summary gives no figures.
* [Training neural networks on Apple's Neural Engine](https://ai.exponencialadm.net/okf/en/radar/treinamento-de-redes-neurais-no-neural-0ef7da.md) - Published on March 2, 2026, this entry presents the ANE repository, which describes neural-network training through private, reverse-engineered APIs for Apple's Neural Engine.
* [Microsoft announces long-audio transcription with VibeVoice-ASR](https://ai.exponencialadm.net/okf/en/radar/vibevoice-asr-promete-transcrever-audio-de-60b7b5.md) - On March 2, 2026, Microsoft said VibeVoice-ASR transcribes audio up to 60 minutes in one pass and offers speaker diarization, timestamps, hotwords, and support for more than 50 languages.
* [DAPO reports AIME result rising from 30% to 50%](https://ai.exponencialadm.net/okf/en/radar/dapo-relata-grpo-de-30-para-94d03a.md) - The DAPO paper reports that standard GRPO scored 30% on AIME in its experiments; adding some techniques raised the reported result to 50%.
* [FlashAttention from scratch in CuPy and CUDA](https://ai.exponencialadm.net/okf/en/radar/flashattention-do-zero-em-cupy-e-2ca86f.md) - Published on March 1, 2026, this project implements FlashAttention with tiling and online softmax without materializing the NxN score matrix. The author reports mathematical verification, but no speedup over cuBLAS.
* [Fast LLM Inference From Scratch: a starting point](https://ai.exponencialadm.net/okf/en/radar/inferencia-rapida-de-llm-do-zero-2228e4.md) - Andrew Chan’s article, published on March 1, 2026, offers a practical starting point for engineers interested in implementing language-model inference.
* [Memory and branch prediction in software performance](https://ai.exponencialadm.net/okf/en/radar/memoria-e-previsao-de-desvios-no-a42036.md) - An article examines how the memory subsystem interacts with branch prediction and influences software performance.
* [Hardware memory models: an introduction](https://ai.exponencialadm.net/okf/en/radar/modelos-de-memoria-de-hardware-e-1906bb.md) - Russ Cox's article opens a series on memory models and explains why hardware models help engineers reason about concurrent software.
* [Memory models in programming languages](https://ai.exponencialadm.net/okf/en/radar/modelos-de-memoria-em-linguagens-de-81a770.md) - Published on March 1, 2026, Russ Cox’s article is Part 2 of a series on memory models and focuses on memory models used by programming languages.
* [Library order in static linking](https://ai.exponencialadm.net/okf/en/radar/ordem-de-bibliotecas-na-linkedicao-estatica-235289.md) - Published on March 1, 2026, Eli Bendersky’s article discusses library order in static linking and its usefulness in diagnosing static-linking problems.
* [V8: Pointer Compression Cuts Heap Size by up to 43%](https://ai.exponencialadm.net/okf/en/radar/pointer-compression-no-v8-menos-memoria-63a52c.md) - A V8 article presents pointer compression as a memory optimization and says it reduced heap size by up to 43%.
* [Value speculation and the challenge of surpassing L1 cache](https://ai.exponencialadm.net/okf/en/radar/value-speculation-e-o-desafio-de-8b5b1c.md) - An article by Francesco Mazzoli discusses value speculation as a performance technique for latency-sensitive code. The available summary does not detail implementation, results, or test conditions.
* [ELF format specification](https://ai.exponencialadm.net/okf/en/radar/formato-elf-guia-para-estudar-binarios-b16a63.md) - A February 28, 2026 entry highlights the specification for the Executable and Linkable Format (ELF), a reference for people working with executables, object files, and binary tools.
* [Comparing AI inference hardware](https://ai.exponencialadm.net/okf/en/radar/hardware-de-inferencia-em-2026-comparacao-822b26.md) - Published on February 28, 2026, the item outlines a comparison of the 2026 inference-chip landscape, focusing on cost per token and trade-offs against GPUs.
* [A step-by-step introduction to CUDA C++](https://ai.exponencialadm.net/okf/en/radar/introducao-passo-a-passo-a-programacao-9900e2.md) - A post published on February 28, 2026 introduces CUDA programming through a simple, step-by-step example of parallel programming.
* [The Lost Art of Structure Packing](https://ai.exponencialadm.net/okf/en/radar/organizacao-de-estruturas-em-c-um-9f1808.md) - Radar item: Eric S. Raymond published the article “The Lost Art of Structure Packing”, on structure packing, a topic tied to the C language and memory layout. The source gives no further details about the content.
* [Post compares ColBERTv2 and Qwen3-Embed-8B](https://ai.exponencialadm.net/okf/en/radar/colbertv2-qwen3-e-bm25-compare-o-7e44a0.md) - A post claims that 100-million-parameter ColBERTv2 outperforms Qwen3-Embed-8B and highlights BM25 with reranking as a competitive baseline for deep research.
* [Radar: Using modern C++ safely](https://ai.exponencialadm.net/okf/en/radar/como-adotar-c-moderno-com-seguranca-8494a3.md) - A February 27, 2026 publication recommends a book about when to use or avoid modern C++ features.
* [How to reduce synchronization in multithreaded systems](https://ai.exponencialadm.net/okf/en/radar/como-reduzir-a-sincronizacao-em-sistemas-a85bc0.md) - Good multithreaded design reduces unnecessary synchronization to allow more simultaneous execution, while handling the points where synchronization is essential for correctness.
* [Lepton XDS pairs thermal and RGB imaging](https://ai.exponencialadm.net/okf/en/radar/lepton-xds-reune-imagem-termica-e-14b4bf.md) - Published on February 27, 2026, the Lepton XDS combines a 160 × 120 radiometric thermal camera with a 5 MP visible camera. Its software offers measurement overlays and radiometric JPEG support; a development kit is available.
* [Why CPUs Have Multiple Cache Levels](https://ai.exponencialadm.net/okf/en/radar/por-que-cpus-tem-varios-niveis-a7854b.md) - An article by Fabian Giesen explores why CPUs have multiple cache levels and explains the topic’s relevance to performance analysis.
* [Talk explores memory-centric computing](https://ai.exponencialadm.net/okf/en/radar/computacao-centrada-na-memoria-palestra-de-3c9277.md) - A PDF of Onur Mutlu’s talk at EFCL Winter School 2026 presents an approach aimed at performance challenges related to memory.
* [Darkweb Scanner monitors keywords on .onion sites](https://ai.exponencialadm.net/okf/en/radar/darkweb-scanner-monitora-palavras-chave-em-a8e897.md) - Published on February 26, 2026, this entry presents an OSINT tool for monitoring keywords on .onion sites, with a stated focus on threat intelligence and brand monitoring.
* [OpenAnonymity proposes unlinkable AI inference](https://ai.exponencialadm.net/okf/en/radar/openanonymity-propoe-separar-identidade-e-inferencia-87a90f.md) - Published on February 26, 2026, the post presents an approach intended to prevent AI providers from linking inference calls to individual users. It says the approach was integrated into open-source infrastructure and a chat app.
* [A CUDA matrix optimization diary](https://ai.exponencialadm.net/okf/en/radar/otimizacao-iterativa-de-multiplicacao-de-matrizes-5e3406.md) - Published on February 26, 2026, the diary describes iterative optimization of a CUDA matrix multiplication kernel and decisions made while seeking performance close to cuBLAS.
* [Padding for reusing CUDA graphs](https://ai.exponencialadm.net/okf/en/radar/padding-de-inputs-para-reutilizar-cuda-3a3337.md) - Published on February 26, 2026, the summary describes a strategy for inference workloads with variable sequence lengths: set a maximum input size, pad inputs to the nearest bucket, and create a family of CUDA graphs. According to the author, this avoids recompilations.
* [Praetorian's multi-agent pipeline for CVE detection templates](https://ai.exponencialadm.net/okf/en/radar/pipeline-multiagente-transforma-pesquisa-de-cves-7c7e9a.md) - Praetorian describes a multi-agent pipeline that turns CVE research into validated Nuclei detection templates, using specialized agents, three model providers and an actor-critic refinement loop.
* [CUDA matrix transposition with shared memory](https://ai.exponencialadm.net/okf/en/radar/transposicao-de-matrizes-em-cuda-com-a68a37.md) - Published on February 26, 2026, Mark Harris’s article discusses how shared memory can improve matrix transposition performance in CUDA C/C++.
* [Tutorial covers techniques for dLLMs](https://ai.exponencialadm.net/okf/en/radar/como-estudar-tecnicas-de-inferencia-e-29a153.md) - Published on February 25, 2026, this roughly 22-minute tutorial discusses inference and training for diffusion language models (dLLMs).
* [Guide explores creating a virtual iPhone](https://ai.exponencialadm.net/okf/en/radar/guia-explora-a-criacao-de-um-ca2396.md) - A guide published in a GitHub repository may be a reference for engineers interested in exploring virtual iPhone configurations. The available description does not detail steps or technical requirements.
* [CUDA kernel patterns: four core operations](https://ai.exponencialadm.net/okf/en/radar/quatro-padroes-para-estudar-kernels-cuda-ffd171.md) - Published on February 25, 2026, the record highlights reduction, 1D convolution, histogram, and numerically stable softmax as key CUDA kernel patterns, with techniques for parallelization and memory use.
* [RL recipes for math, code, and tool use](https://ai.exponencialadm.net/okf/en/radar/receitas-de-rl-para-matematica-codigo-dbace5.md) - Prime Intellect’s documentation collects practical reinforcement-learning training recipes for mathematical reasoning, code generation, tool use, and other tasks in the Lab.
* [Technical analysis of Intel Management Engine](https://ai.exponencialadm.net/okf/en/radar/intel-me-vulnerabilidades-e-deteccao-4cc9f2.md) - The article examines Intel Management Engine architecture and firmware, exploitation of CVE-2017-5689, a TOCTOU race condition in CVE-2025-20037, PKfail, and detection methods.
* [QLoRA demonstration uses 50 model outputs](https://ai.exponencialadm.net/okf/en/radar/qlora-em-demonstracao-50-saidas-para-33e25d.md) - Published on February 24, 2026, the post describes a local distillation demonstration: 50 GPT-5.2 outputs were used to fine-tune Llama 3B with QLoRA. The author says the process transferred significant capabilities.
* [Secure Erasure in Magnetic and Solid-State Storage](https://ai.exponencialadm.net/okf/en/radar/exclusao-segura-em-midias-magneticas-e-05be89.md) - Peter Gutmann’s 1996 article discusses data remanence and the limits of a single overwrite on magnetic media.
* [HPC++ aims to parallelize C++ for CPUs and GPUs](https://ai.exponencialadm.net/okf/en/radar/hpc-leva-c-sequencial-a-cpus-2c94de.md) - A February 22, 2026 entry describes an LLVM-based framework that turns sequential C++ programs into parallel implementations for multicore CPUs and OpenCL-compatible GPUs.
* [ntransformer claims Llama 70B runs on an RTX 3090](https://ai.exponencialadm.net/okf/en/radar/ntransformer-executa-llama-70b-em-uma-da9fa3.md) - Published on February 22, 2026, the entry describes a C++/CUDA inference engine that claims to run Llama 70B on an RTX 3090, transferring data from NVMe to the GPU without passing through the CPU.
* [llmfit estimates model compatibility with local hardware](https://ai.exponencialadm.net/okf/en/radar/llmfit-ajuda-a-avaliar-modelos-para-48bd38.md) - A February 21, 2026 entry describes llmfit, a tool that analyzes hardware to identify compatible models and estimate performance before downloading weights.
* [VulnLLM-R-7B for vulnerability detection](https://ai.exponencialadm.net/okf/en/radar/vulnllm-r-7b-raciocinio-aplicado-a-5b13fd.md) - A February 21, 2026 entry presents VulnLLM-R-7B as a reasoning model specialized in vulnerability detection, without reporting benchmarks or quantitative results.
* [Booth compiles CUDA for AMD GPUs](https://ai.exponencialadm.net/okf/en/radar/booth-explora-compilacao-para-cuda-triton-8104fd.md) - Published on February 18, 2026, the bookmark presents Booth as an open-source compiler for CUDA, Triton, and HIP, targeting multiple GPU and CPU architectures.
* [Booth explores compilation for GPUs and CPUs](https://ai.exponencialadm.net/okf/en/radar/booth-explora-compilacao-para-gpus-e-79557e.md) - Published on February 18, 2026, the entry describes Booth as an open-source compiler for CUDA, Triton, and HIP, with support for multiple GPU and CPU architectures.
* [Guide surveys C++ standard algorithms](https://ai.exponencialadm.net/okf/en/radar/como-estudar-algoritmos-padrao-de-c-606e25.md) - Published on February 18, 2026, the entry points to a GitHub repository for “The Standard Algorithms in C++,” a book by Šimon Tóth.
* [Handling floating-point errors in C++](https://ai.exponencialadm.net/okf/en/radar/como-tratar-erros-de-ponto-flutuante-39c529.md) - An article published on February 18, 2026, presents techniques for detecting and handling floating-point errors in C++, considering possible effects on correctness and runtime performance.
* [Kascade reuses sparse attention across Transformer layers](https://ai.exponencialadm.net/okf/en/radar/kascade-reutiliza-atencao-esparsa-entre-camadas-f7ba16.md) - Published on February 18, 2026, the summary presents Kascade, an inference approach that computes full attention in selected reference layers and reuses those results in intermediate layers.
* [LLD linker documentation: design and performance](https://ai.exponencialadm.net/okf/en/radar/documentacao-do-linker-lld-design-e-9e5f6b.md) - Published on February 17, 2026, the LLVM documentation page for LLD describes a video covering the linker’s internals, design, data structures, and performance.
* [Intel guide documents instruction access through C functions](https://ai.exponencialadm.net/okf/en/radar/guia-da-intel-reune-intrinsics-para-c40c48.md) - Published on February 17, 2026, the Intel Intrinsics Guide lists C-style functions that provide access to processor instructions without requiring assembly code.
* [hf-mem estimates memory needs for new models](https://ai.exponencialadm.net/okf/en/radar/hf-mem-estima-memoria-necessaria-para-71f737.md) - Published on February 17, 2026, the report presents hf-mem, a command-line tool for estimating memory requirements as model architectures change.
* [CogRouter adjusts reasoning by agent step](https://ai.exponencialadm.net/okf/en/radar/cogrouter-ajusta-o-raciocinio-por-etapa-60abf6.md) - Published on February 16, 2026, the report describes four cognitive levels for adjusting reasoning depth in LLM agents and reports 82.3% benchmark success for a 7B model.
* [Radar: how to write shared libraries](https://ai.exponencialadm.net/okf/en/radar/como-estudar-e-escrever-shared-libraries-b4fcb5.md) - Published on February 16, 2026, this bookmark presents Ulrich Drepper’s guide to writing shared libraries, aimed at engineers working on their design and implementation.
* [Project Aura: ESP32-S3 air monitor](https://ai.exponencialadm.net/okf/en/radar/project-aura-monitor-diy-de-qualidade-ed4e74.md) - Published on February 16, 2026, Project Aura is a DIY air-quality monitor with a touchscreen, sensors, and Home Assistant support.
* [Self-distillation for continual learning without rewards](https://ai.exponencialadm.net/okf/en/radar/self-distillation-para-continual-learning-sem-e504de.md) - Published on February 16, 2026, the post describes a self-distillation technique for continual learning: a demonstration-conditioned model acts as teacher to train the same model without the demonstration, without defining a reward function.
* [Linkers: an introductory guide for C and C++](https://ai.exponencialadm.net/okf/en/radar/linkers-guia-inicial-para-entender-o-ba9635.md) - Published on February 15, 2026, this beginner guide highlights why understanding linkers helps engineers navigate the build process for C and C++ programs.
* [PyTorch notebook explores GRPO improvements](https://ai.exponencialadm.net/okf/en/radar/notebook-em-pytorch-explora-melhorias-no-b1d42f.md) - Published on February 15, 2026, chapter 7 of Reasoning From Scratch presents a GRPO implementation that adds and examines bounded policy ratios, a KL term, format rewards, and other improvements.
* [QED-Nano: a 4B model for olympiad proofs](https://ai.exponencialadm.net/okf/en/radar/qed-nano-um-modelo-4b-para-d7b5b9.md) - A project explores training a small model to prove difficult theorems and presents details about its training, datasets, and implementation.
* [Online RL tests HPC code on real machines](https://ai.exponencialadm.net/okf/en/radar/rl-online-avalia-codigo-hpc-em-cb7f5e.md) - An article published on February 15, 2026 describes online reinforcement learning with rewards based on benchmarks run on real machines.
* [TurningPoint-GRPO uses stage-level rewards](https://ai.exponencialadm.net/okf/en/radar/turningpoint-grpo-explora-recompensas-por-etapa-fe4736.md) - Published on February 15, 2026, TurningPoint-GRPO is described as a framework for sparse rewards in flow-matching models, identifying stages that reverse the local reward trend.
* [Exaviz CM5 board adds eight PoE ports to mini NAS/NVR](https://ai.exponencialadm.net/okf/en/radar/exaviz-cm5-leva-oito-portas-poe-bdf4f6.md) - The Raspberry Pi Compute Module 5 board combines eight PoE ports and a 2.5 Gbps uplink in a format that fits a 10-inch mini rack. The DeskPi enclosure shown is not yet the final version.
* [Training-free GRPO uses textual experience memory](https://ai.exponencialadm.net/okf/en/radar/grpo-sem-treinamento-usa-memoria-de-18c10d.md) - The method described keeps the model frozen and consults a text memory of successful and unsuccessful attempts. The report claims fine-tuning-level performance with 100 examples.
* [Qwen3 latency optimization on AMD MI300X](https://ai.exponencialadm.net/okf/en/radar/latencia-do-qwen3-em-gpus-amd-4b473b.md) - A February 13, 2026 bookmark about a LMSYS blog article on Qwen3 and Qwen3-VL latency on AMD MI300X GPUs.
* [GAT: attention highlights relevant graph neighbors](https://ai.exponencialadm.net/okf/en/radar/graph-attention-networks-aprendem-quais-vizinhos-cdf837.md) - A paper presented at ICLR 2018 describes Graph Attention Networks, which weight information from neighboring nodes using weights calculated for connected pairs. The record was published on February 12, 2026.
* [C inference for Qwen3-ASR transcription](https://ai.exponencialadm.net/okf/en/radar/inferencia-em-c-para-transcricao-qwen3-0d515e.md) - A repository offers C inference for the Qwen3-ASR 0.6B and 1.7B speech recognition models, with potential uses in transcription bots and voice-controlled interfaces.
* [SlimeVR Butterfly trackers use the nRF52833 MCU](https://ai.exponencialadm.net/okf/en/radar/trackers-slimevr-butterfly-usam-o-nrf52833-9d96bd.md) - Published on February 12, 2026, the entry describes open-hardware virtual-reality trackers combining IMUs, an nRF52833 wireless MCU, and a custom 2.4 GHz ESB protocol.
* [Three approaches to on-policy self-distillation](https://ai.exponencialadm.net/okf/en/radar/tres-abordagens-de-self-distillation-on-3dc1c7.md) - Published on February 12, 2026, the article examines OPSD, SDFT, and SDPO, on-policy self-distillation methods using denser feedback, as well as benchmarks and limitations.
* [vLLM announces streaming inputs and a WebSocket API](https://ai.exponencialadm.net/okf/en/radar/vllm-apresenta-entradas-em-streaming-e-00e07e.md) - Published on February 11, 2026, the vLLM blog describes streaming inputs and a real-time WebSocket API for applications that process audio, video, and other data incrementally.
* [Beelink ME Pro combines SATA, M.2 and 5GbE](https://ai.exponencialadm.net/okf/en/radar/beelink-me-pro-reune-sata-m-6852a4.md) - Announced on February 10, 2026, the Beelink ME Pro combines a compact NAS and mini PC form factor, two SATA bays, up to three M.2 NVMe sockets, and 5GbE networking.
* [ROME explores segment-level credit for agent training](https://ai.exponencialadm.net/okf/en/radar/rome-explora-credito-por-trechos-no-d4fae3.md) - A 3-billion-parameter model uses segment-level rewards, rather than assigning credit only by token or trajectory, to train agents that use tools.
* [BM25 ranks documents without embeddings](https://ai.exponencialadm.net/okf/en/radar/como-o-bm25-classifica-documentos-sem-9ff5e3.md) - Published on February 9, 2026, the post explains how BM25 scores search results using term rarity, diminishing returns for repeated terms, and document length normalization, without requiring training, embeddings, or fine-tuning.
* [LatentMoE distributes heads across GPUs](https://ai.exponencialadm.net/okf/en/radar/latentmoe-distribui-heads-entre-gpus-para-d880e4.md) - A February 9, 2026 post describes a LatentMoE design with head parallelism and reports up to 1.61× speedup over standard MoE with expert parallelism.
* [Local Studio PR proposes ROCm, speech, image, and video support](https://ai.exponencialadm.net/okf/en/radar/local-studio-propoe-suporte-a-rocm-7274eb.md) - Published on February 9, 2026, a Local Studio pull request describes proposed ROCm support and a speech mode with TTS/STT using ACE-Step and Whisper v3 Large. The proposal also mentions images and video.
* [MaxRL proposes likelihood training for binary-reward RL](https://ai.exponencialadm.net/okf/en/radar/maxrl-explora-verossimilhanca-em-rl-com-251ada.md) - Published on February 9, 2026, the post presents MaxRL: additional rollouts approximate maximum-likelihood training with non-differentiable sampling. The post says the method reduces underweighting of difficult prompts and achieves up to 20× greater test-time efficiency than GRPO.
* [Hex-Rays: a presentation on microcode](https://ai.exponencialadm.net/okf/en/radar/microcode-por-dentro-do-decompiler-hex-f63330.md) - Published on February 9, 2026, the entry describes a presentation on the internals of the Hex-Rays decompiler in IDA Pro, including microcode. The post points to a video and a PDF.
* [REMnux connects AI agents to more than 200 tools](https://ai.exponencialadm.net/okf/en/radar/remnux-integra-agentes-de-ia-a-86a03d.md) - Published on February 9, 2026, the report describes a REMnux MCP server connecting AI agents to more than 200 malware-analysis tools and offering guidance on selecting them and interpreting results.
* [Google Research presents Sequential Attention](https://ai.exponencialadm.net/okf/en/radar/sequential-attention-busca-modelos-mais-enxutos-9d44e9.md) - On February 9, 2026, Google Research presented an approach that selects features or layers during training to seek leaner, faster models without sacrificing accuracy.
* [VulnLLM-R-7B analyzes code vulnerabilities](https://ai.exponencialadm.net/okf/en/radar/vulnllm-r-7b-avaliando-um-modelo-142e3e.md) - Published on February 9, 2026, VulnLLM-R-7B is presented as a reasoning model that searches code for vulnerabilities by examining data and control flows.
* [VOLT: an open-source GPU compiler](https://ai.exponencialadm.net/okf/en/radar/volt-por-dentro-de-um-compilador-c8beb7.md) - An article published on February 8, 2026 introduces VOLT, an open-source GPU compiler, and discusses emerging open GPU architectures that define SIMT features.
* [Research explores generalizable activations](https://ai.exponencialadm.net/okf/en/radar/alphaevolve-pesquisa-funcoes-de-ativacao-generalizaveis-3cec2b.md) - A February 7, 2026 publication reports using AlphaEvolve to search for activation functions, as part of broader research into discovering AI architectures.
* [Open-source repository explores voice agents](https://ai.exponencialadm.net/okf/en/radar/repositorio-open-source-oferece-ponto-de-82a897.md) - Published on February 6, 2026, the repository offers engineers a starting point for creating voice agents with open-source models.
* [How GPUs Evolved from SIMD to SIMT](https://ai.exponencialadm.net/okf/en/radar/de-simd-a-simt-como-evoluiram-a0f576.md) - An ACM SIGGRAPH Blog article, published on February 5, 2026, describes how GPUs combined thread abstractions with SIMD hardware, becoming flexible processors for graphics, AI, and scientific computing.
* [Eight ways to implement defer in C](https://ai.exponencialadm.net/okf/en/radar/oito-maneiras-de-implementar-defer-em-b95b8e.md) - An article compares eight approaches to implementing defer in C and discusses portability and cleanup patterns.
* [Guide covers tool calling with local LLMs](https://ai.exponencialadm.net/okf/en/radar/tool-calling-com-modelos-locais-guia-412212.md) - Published on February 5, 2026, the guide presents function-calling examples with open models: stories, Python execution, terminal calls, and mathematics.
* [Veritensor 1.4.0 broadens the project's scope](https://ai.exponencialadm.net/okf/en/radar/veritensor-1-4-0-amplia-o-ed775f.md) - Announced on February 5, 2026, version 1.4.0 is described as a major update that expands a model scanner project into an AI security platform.
* [EB-JEPA brings modular JEPA implementations](https://ai.exponencialadm.net/okf/en/radar/eb-jepa-biblioteca-open-source-para-e36c1b.md) - Published on February 4, 2026, the article presents EB-JEPA, an open-source library for learning representations and world models with architectures that predict in representation space rather than pixel space.
* [Sysdig reports AWS intrusion reaching Bedrock](https://ai.exponencialadm.net/okf/en/radar/invasao-na-aws-comeca-com-credenciais-f02e37.md) - On February 4, 2026, Sysdig described an AWS intrusion that began with credentials exposed in S3 and escalated through Lambda to access Bedrock.
* [WeightWatch flags a late collapse in generalization](https://ai.exponencialadm.net/okf/en/radar/weightwatcher-sinaliza-anti-grokking-em-dois-0861c3.md) - A paper reports anti-grokking signals in weight matrices across two prolonged-training experiments: a three-layer MLP on a subset of MNIST and a transformer trained on modular addition.
* [Book covers CUDA for LLM workloads](https://ai.exponencialadm.net/okf/en/radar/cuda-para-llms-kernels-flash-attention-7a8798.md) - Published on February 3, 2026, the Manning book covers CUDA programming for NVIDIA GPUs, from initial kernels to Flash Attention, and mentions profiling with Nsight Compute.
* [CuTe Layouts: matrix shapes and strides](https://ai.exponencialadm.net/okf/en/radar/cute-layouts-ajudam-a-entender-shapes-6f50e9.md) - Published on February 3, 2026, this bookmark recommends the CUTLASS/CuTe Layout Algebra tutorials for understanding matrix shapes and strides, and points to CuTe documentation on layouts.
* [GLM-OCR announces a 0.9B multimodal OCR model](https://ai.exponencialadm.net/okf/en/radar/glm-ocr-modelo-multimodal-de-ocr-e09e1d.md) - Published on February 3, 2026, the announcement describes a 0.9B multimodal OCR model using the GLM-V architecture and an MIT license. The author says it ranks first on OmniDocBench v1.5, with a score of 94.62.
* [CoreRipperX 2.0 expands CPU stress testing](https://ai.exponencialadm.net/okf/en/radar/coreripperx-2-0-traz-interface-e-cf1f25.md) - The February 1, 2026 announcement describes a new interface, throttling and effective-clock checks, additional AVX tests, and load modes for the CPU stress-testing tool.
* [IPC and the microarchitecture behind CPU performance](https://ai.exponencialadm.net/okf/en/radar/ipc-e-microarquitetura-como-analisar-cpus-e3f060.md) - Published on February 1, 2026, this item offers a concise explanation of the nature of IPC and the CPU microarchitectural components that help achieve it.
* [Nanochat architecture and training choices](https://ai.exponencialadm.net/okf/en/radar/o-que-a-discussao-do-nanochat-d3e5ab.md) - A February 1, 2026 publication lists architecture and optimization choices discussed for nanochat, offering topics for engineers to investigate.
* [Scaled dot-product attention explained](https://ai.exponencialadm.net/okf/en/radar/scaled-dot-product-attention-o-mecanismo-516137.md) - Published on February 1, 2026, this bookmark introduces queries, keys and values, and points to a notebook implementing LLM and NLP systems from first principles.
* [AMD publishes specification for two instructions](https://ai.exponencialadm.net/okf/en/radar/amd-publica-especificacao-de-instrucoes-bit-5a7983.md) - On January 31, 2026, AMD published version 1.00 of document 69192, specifying Bit Matrix Multiply and Bit Reversal instructions for AMD64.
* [Exploration in RL for LLMs](https://ai.exponencialadm.net/okf/en/radar/exploracao-em-rl-para-llms-tres-ef5e52.md) - A CMU post describes three exploration regimes in reinforcement learning for language models and notes potential challenges in training on difficult tasks.
* [Chapter introduces GRPO with verifiable rewards](https://ai.exponencialadm.net/okf/en/radar/grpo-do-zero-para-recompensas-verificaveis-7d4f0e.md) - Published on January 31, 2026, the chapter presents GRPO from scratch for reinforcement learning with verifiable rewards. The author says future material will explore more runs, analyses, and algorithmic adjustments.
* [Using AI to extract TTPs from threat reports](https://ai.exponencialadm.net/okf/en/radar/ia-pode-revelar-lacunas-na-deteccao-c25d85.md) - A January 31, 2026 summary of a Microsoft-described workflow reports using AI to extract TTPs, map detection coverage, and flag possible gaps, with specialists reviewing and validating the results.
* [Modern C++ in embedded firmware](https://ai.exponencialadm.net/okf/en/radar/modern-c-em-firmware-embarcado-por-c041ac.md) - Published on January 31, 2026, the first article in a series discusses Modern C++ in embedded systems and challenges myths about C++ and microcontrollers.
* [VoxCPM proposes tokenizer-free speech synthesis](https://ai.exponencialadm.net/okf/en/radar/voxcpm-explora-sintese-de-fala-sem-209c13.md) - Published on January 31, 2026, the post presents VoxCPM, a text-to-speech system based on MiniCPM-4 and an end-to-end autoregressive diffusion architecture.
* [LMCache reuses KV states across storage tiers](https://ai.exponencialadm.net/okf/en/radar/lmcache-reaproveita-estados-kv-entre-camadas-4c2b69.md) - Published on January 30, 2026, the report describes LMCache, an open-source LLM serving extension that manages KV states across GPUs, CPUs, and local disks.
* [vLLM adds support for Qwen3-ASR](https://ai.exponencialadm.net/okf/en/radar/vllm-adiciona-suporte-ao-qwen3-asr-ddedc8.md) - vLLM has supported Qwen3-ASR since its launch. A linked guide gives engineers a starting point for running the model with vLLM.
* [djbsort adds AVX2 performance and support for more numeric types](https://ai.exponencialadm.net/okf/en/radar/djbsort-amplia-desempenho-e-tipos-numericos-b649a2.md) - A new djbsort release, a library for sorting numeric arrays, brings AVX2 performance improvements, support for six numeric types, and updated benchmarks.
* [ELF 101: A guide to Linux executables](https://ai.exponencialadm.net/okf/en/radar/elf-101-guia-para-entender-executaveis-2587f4.md) - Published on January 27, 2026, the material introduces ELF, a format used by Linux executables, and notes its value for inspecting and diagnosing these files.
* [MXFP8 quantizer exceeds 6 TB/s on B200](https://ai.exponencialadm.net/okf/en/radar/quantizer-mxfp8-em-cutedsl-supera-6-a24715.md) - Published on January 27, 2026, the item describes a CuTeDSL MXFP8 quantizer for B200 GPUs and reports throughput above 6 TB/s.
* [Tencent presents HPC-Ops operators for LLM inference](https://ai.exponencialadm.net/okf/en/radar/tencent-apresenta-operadores-hpc-ops-para-e7744d.md) - A briefing published on January 27, 2026 describes an open-source operator library for LLM inference and reports throughput gains and faster kernels, without giving figures.
* [UltraRAG 3.0 defines RAG workflows in YAML](https://ai.exponencialadm.net/okf/en/radar/ultrarag-3-0-organiza-workflows-de-dfb69c.md) - Published on January 27, 2026, UltraRAG 3.0 is described as a lightweight research framework for RAG systems, with MCP servers and configurable YAML workflows.
* [A series on writing a Linux debugger](https://ai.exponencialadm.net/okf/en/radar/uma-serie-sobre-escrever-um-debugger-e3f622.md) - Sy Brand published a series about writing a debugger for Linux, collected under the “Debuggers” tag on the author’s blog. The topic may interest engineers working on debugging tools.
* [INTELLECT-2 points to delta=4.0 for steadier GRPO](https://ai.exponencialadm.net/okf/en/radar/grpo-configuracao-de-estabilidade-com-delta-02bc5a.md) - Published on January 26, 2026, the item says the INTELLECT-2 technical report introduced a setting intended to improve GRPO training stability: use delta=4.0. It is a concrete parameter engineers can investigate.
* [Radar: selected resources on agentic reasoning](https://ai.exponencialadm.net/okf/en/radar/recursos-para-estudar-raciocinio-agentic-em-092159.md) - A GitHub list gathers papers and resources based on a survey about agentic reasoning for language models, aimed at engineers researching the topic.
* [Cornell introductory compiler notes](https://ai.exponencialadm.net/okf/en/radar/como-estudar-as-notas-de-compiladores-81cebb.md) - Cornell offers introductory compiler notes for its CS 4120/ENGRD 4120 course, taught in spring 2022.
* [CALM proposes predicting vector blocks](https://ai.exponencialadm.net/okf/en/radar/calm-propoe-prever-blocos-de-vetores-ba2b51.md) - Published on January 24, 2026, the item describes CALM models that predict vectors for meaning blocks rather than individual tokens. It claims 4× more information per step and a 44% reduction in training compute.
* [CLI estimates VRAM needed to load models](https://ai.exponencialadm.net/okf/en/radar/cli-estima-a-vram-necessaria-antes-bd0841.md) - Published January 23, 2026, this item describes a command-line tool that estimates the VRAM needed to load a model without loading it.
* [Unsloth guide to embedding fine-tuning](https://ai.exponencialadm.net/okf/en/radar/guia-de-fine-tuning-de-embeddings-8d3958.md) - Published on January 22, 2026, the material explains embedding model fine-tuning and includes a Colab notebook for EmbeddingGemma (300M).
* [Qwen3-TTS announced as open source](https://ai.exponencialadm.net/okf/en/radar/qwen3-tts-anuncio-de-modelo-open-8e6cc8.md) - On January 22, 2026, a post announced that Qwen3-TTS would be made available as open source and mentioned voice cloning and design. It provided no repository or other link.
* [Fine-tuning a 14B code model with SFT and DPO](https://ai.exponencialadm.net/okf/en/radar/ajuste-de-um-modelo-de-codigo-5dae03.md) - Published on January 20, 2026, the article reports fine-tuning a 14B model on code conversations with SFT and DPO, and discusses results, costs, and difficulties.
* [slime v0.2.2 adds Int4-QAT and R3 support](https://ai.exponencialadm.net/okf/en/radar/slime-v0-2-2-traz-int4-8ca9e5.md) - Published on January 18, 2026, the slime v0.2.2 entry reports memory and performance improvements, Int4-QAT training, and full R3 support with DeepEP and MTP.
* [Architecture influences retention and forgetting](https://ai.exponencialadm.net/okf/en/radar/arquitetura-influencia-o-aprendizado-continuo-24747e.md) - A January 17, 2026 note on a comparison of architectures for continual learning: wider networks and less global average pooling may favor retention, while ResNets and WideResNets learn new tasks quickly but forget more.
* [Study examines attacks on Federated Learning](https://ai.exponencialadm.net/okf/en/radar/inversao-de-gradientes-no-federated-learning-8f4682.md) - A study of gradient inversion attacks reports that the most practical attack method remains unreliable and proposes a three-stage defense.
* [SAR as a reward for LLM RL](https://ai.exponencialadm.net/okf/en/radar/sar-como-sinal-de-recompensa-em-3b22da.md) - A January 16, 2026 publication presents Self-Aligned Reward (SAR) as a reward signal compatible with PPO and GRPO, intended to improve reasoning and reduce tokens.
* [Step Audio R1.1: real-time spoken dialogue](https://ai.exponencialadm.net/okf/en/radar/step-audio-r1-1-dialogo-falado-718b6c.md) - A January 16, 2026 publication presents Step Audio R1.1 as an interactive real-time spoken-dialogue model and describes its license and architecture.
* [Unsloth describes GRPO with 380K context](https://ai.exponencialadm.net/okf/en/radar/grpo-com-contexto-longo-requisitos-e-f6ade4.md) - Published on January 15, 2026, the item summarizes documentation on long-context reinforcement learning fine-tuning with GRPO. Unsloth says its batching algorithms enable 380K context for gpt-oss on a 192 GB GPU.
* [Radar: controlled evolution for code optimization](https://ai.exponencialadm.net/okf/en/radar/evolucao-controlada-para-otimizar-codigo-algoritmico-b76f1e.md) - A January 14, 2026 article proposes an iterative method to improve algorithmic code using diverse strategies, feedback-guided mutations, and experience across tasks. The approach is evaluated on EffiBench-X.
* [Bayesian generalization in offline RL](https://ai.exponencialadm.net/okf/en/radar/generalizacao-bayesiana-em-rl-offline-baseado-5aca5a.md) - A January 14, 2026 post presents a paper on adaptive generalization in model-based offline reinforcement learning and long-horizon rollouts.
* [RAG: expand documents, not the model](https://ai.exponencialadm.net/okf/en/radar/rag-ampliar-documentos-antes-de-aumentar-27a693.md) - Published on January 14, 2026, this summary describes a CMU researchers’ study of expanding the retrieval corpus in RAG as an alternative to scaling up the language model.
* [UniversalRAG routes retrieval across modalities and granularities](https://ai.exponencialadm.net/okf/en/radar/universalrag-direciona-a-recuperacao-multimodal-26f154.md) - Published on January 14, 2026, the summary presents UniversalRAG, a framework that routes queries to a corpus for a specific modality before retrieving information from it.
* [MemRL learns to assess agent memories](https://ai.exponencialadm.net/okf/en/radar/memrl-aprende-a-priorizar-memorias-de-2b2d86.md) - Published on January 13, 2026, the summary presents MemRL, an approach that learns to rank episodic memories by utility without changing language-model weights.
* [Microsoft launches 14B FrogMini coding agent](https://ai.exponencialadm.net/okf/en/radar/microsoft-lanca-agente-de-codigo-frogmini-a0d7ad.md) - A January 13, 2026 publication reports the launch of FrogMini, a 14B coding agent, and gives a score of 45.3% on SWE-Bench Verified.
* [Pocket TTS: compact speech synthesis for laptops](https://ai.exponencialadm.net/okf/en/radar/pocket-tts-sintese-de-voz-de-0df406.md) - A January 13, 2026 publication presents Pocket TTS, a 100-million-parameter text-to-speech model described as running on a laptop without a GPU and offering voice cloning.
* [DroPE: extending context without positional embeddings](https://ai.exponencialadm.net/okf/en/radar/drope-propoe-estender-contexto-sem-embeddings-462e3c.md) - Published on January 12, 2026, the summary presents DroPE, a method that removes positional embeddings after training and uses brief calibration to extend context.
* [Interviews map the RL environments field](https://ai.exponencialadm.net/okf/en/radar/entrevistas-revelam-como-se-desenvolvem-ambientes-68b1d0.md) - In a post dated January 12, 2026, Chris Barber and JS Denain report interviews with 18 people from RL environment startups, neolabs, and frontier labs. The material offers engineers a view of how different organizations develop these environments.
* [Leaker: passive breach enumeration](https://ai.exponencialadm.net/okf/en/radar/leaker-enumeracao-passiva-de-vazamentos-de-6fa3d6.md) - Published on January 12, 2026, the listing describes Leaker as a tool that searches passive online sources for credential leaks associated with email addresses.
* [LEMAS announces speech dataset with 150,000 hours](https://ai.exponencialadm.net/okf/en/radar/lemas-apresenta-dataset-de-speech-com-be887f.md) - Published on January 12, 2026, the announcement describes a multilingual audio dataset with 150,000 hours, coverage of 10 languages and word-level timestamps, along with the generative models LEMAS-TTS and LEMAS-Edit.
* [LoPA explores lookahead in diffusion LLM decoding](https://ai.exponencialadm.net/okf/en/radar/lopa-explora-lookahead-para-acelerar-a-4c2e97.md) - Published on January 11, 2026, the summary describes LoPA, a training-free, plug-and-play algorithm that uses lookahead to choose token-filling orders in diffusion LLMs.
* [SFT and reinforcement fine-tuning in continual post-training](https://ai.exponencialadm.net/okf/en/radar/sft-e-reinforcement-fine-tuning-no-339740.md) - The article compares SFT and reinforcement fine-tuning for continual adaptation of foundation models; its title claims the latter naturally mitigates forgetting.
* [Report examines model tuning for the Q programming language](https://ai.exponencialadm.net/okf/en/radar/como-adaptar-modelos-a-linguagem-q-ef7fa8.md) - Published on January 9, 2026, the technical report presents an open-source approach to adapting language models to Q, a programming language used in quantitative finance.
* [SETA announces RL resources for terminal agents](https://ai.exponencialadm.net/okf/en/radar/seta-lanca-400-ambientes-de-rl-f60efc.md) - On January 9, 2026, SETA announced 400 training environments, a reinforcement-learning pipeline, and SETA-RL-Qwen3-8B weights for terminal-agent experiments.
* [OpenTinker proposes RL-as-a-Service infrastructure](https://ai.exponencialadm.net/okf/en/radar/opentinker-separa-programacao-e-execucao-de-b28cc8.md) - A Radar entry dated January 8, 2026 describes OpenTinker as an open-source reinforcement learning platform for agents, separating programming from execution.
* [ASUS and XREAL plan 240 Hz AR glasses for 2026](https://ai.exponencialadm.net/okf/en/radar/asus-e-xreal-planejam-oculos-ar-c3aef1.md) - The announcement plans augmented-reality glasses with 1080p micro-OLED panels, a USB-C connection, and a dock with HDMI and DisplayPort.
* [Dell announces a 52-inch 6K monitor](https://ai.exponencialadm.net/okf/en/radar/dell-anuncia-monitor-thunderbolt-6k-de-9a12a2.md) - On January 6, 2026, Dell announced a 52-inch Thunderbolt monitor with 6K resolution and inputs for connecting multiple computers.
* [Recursive models use code to process long prompts](https://ai.exponencialadm.net/okf/en/radar/modelos-recursivos-processam-prompts-longos-com-a99d5b.md) - A January 6, 2026 MIT post describes recursive language models that use a Python interface to search, filter and recursively process smaller sections of long inputs. The post says they handle inputs up to about 100 times larger than usual and outperform base models and retrieval configurations.
* [Masters: mask-progressive distillation for VLMs](https://ai.exponencialadm.net/okf/en/radar/distilacao-progressiva-por-mascaras-para-vlms-72e358.md) - Published on January 4, 2026, the post describes the Masters framework and reports score gains on two VLMs, with the stated aim of avoiding online RL's computational cost.
* [JustRL tests simple RL on 1.5B models](https://ai.exponencialadm.net/okf/en/radar/justrl-testa-uma-receita-simples-de-c9f87f.md) - A paper preview describes single-stage training with fixed hyperparameters and reports average accuracies of 54.9% and 64.3% on two 1.5B reasoning models.
* [LoRA and RLVR: comparing strategies across domains](https://ai.exponencialadm.net/okf/en/radar/lora-para-rlvr-em-multiplos-dominios-3bca9a.md) - Published on January 3, 2026, this account compares practical PEFT strategies for multi-domain RLVR. In the reported experiments, averaging specialist LoRA weights and then continuing training outperformed the tested gating method.
* [GLM-4.7 185B W4A16 weights reported at 92 GB](https://ai.exponencialadm.net/okf/en/radar/pesos-w4a16-de-modelo-com-185-d6704b.md) - A January 3, 2026 post reports W4A16 weights for GLM-4.7 185B totaling 92 GB. The author says they ran 100 million tokens locally.
* [South Korean program sets three requirements for model training](https://ai.exponencialadm.net/okf/en/radar/programa-sul-coreano-define-tres-requisitos-f4a537.md) - The first round will include five teams, with four advancing. Projects must train models from scratch, be partly open for commercial use, and pursue ambitious goals such as large-scale omni models.
* [AMD B650 chipset expansion board for Intel systems](https://ai.exponencialadm.net/okf/en/radar/chipset-amd-b650-em-placa-para-28d702.md) - A video documents an open-source expansion board based on the AMD B650 chipset and explores whether it can be used with an Intel motherboard.
* [RK182X kit pairs RK3588 modules with AI accelerators](https://ai.exponencialadm.net/okf/en/radar/kit-rk182x-combina-modulos-rk3588-e-8ba402.md) - Published on December 30, 2025, the announcement describes a kit for evaluating LLM and VLM workloads on embedded systems, using RK3588 and RK182X modules.
* [Lisuan begins shipping 6 nm 7G100 GPUs](https://ai.exponencialadm.net/okf/en/radar/lisuan-inicia-envios-das-gpus-7g100-11b831.md) - Chinese company Lisuan has begun shipping its 7G100 GPUs, manufactured on a 6 nm process, to initial customers. The move marks the product’s initial deployment.
* [Mrav: custom CPU and browser playground](https://ai.exponencialadm.net/okf/en/radar/mrav-cpu-customizada-e-playground-de-325c1e.md) - Published on December 30, 2025, Mrav presents a custom CPU architecture and core, with a browser IDE for writing assembly and watching registers in real time.
* [Intel releases HPC tuning guide for Xeon 6](https://ai.exponencialadm.net/okf/en/radar/guia-de-tuning-hpc-para-intel-29aef3.md) - On December 29, 2025, Intel released version 1.2a of its guide to tuning and optimizing high-performance computing (HPC) applications for Xeon 6 systems.
* [Graphcore IPUs target small batches](https://ai.exponencialadm.net/okf/en/radar/ipus-da-graphcore-paralelismo-e-batches-e4916c.md) - A post dated December 29, 2025 describes Graphcore IPUs as massively parallel accelerators suited to small batches, but difficult to use.
* [Review of 10GbE PCIe Gen4 x1 NICs with RTL8127](https://ai.exponencialadm.net/okf/en/radar/nics-10gbe-pcie-gen4-x1-com-ce5e87.md) - Published on December 29, 2025, the article reviews low-power 10GbE network cards with a PCIe Gen4 x1 interface and the Realtek RTL8127 chip. The available summary gives no comparative results.
* [URM explores recurrence and nonlinearity](https://ai.exponencialadm.net/okf/en/radar/urm-explora-recorrencia-e-nao-linearidade-b53fd5.md) - A post dated December 29, 2025 describes research attributing Universal Transformers’ reasoning gains to recurrent inductive bias and strong nonlinearity. The proposed URM adds local token mixing and truncated backpropagation in recurrent loops.
* [Bit-Brick Cluster K1 supports up to four RISC-V modules](https://ai.exponencialadm.net/okf/en/radar/bit-brick-cluster-k1-reune-ate-a9c048.md) - The cluster accepts up to four SSOM-K1 modules, each with an octa-core RISC-V SpacemiT K1 processor. An integrated Gigabit Ethernet switch connects the modules, and listed operating systems include Bianbu OS, OpenHarmony, and OpenWrt.
* [Stop hooks for keeping agents running](https://ai.exponencialadm.net/okf/en/radar/stop-hooks-podem-prolongar-tarefas-de-7b9432.md) - Published on December 27, 2025, this bookmark describes a plugin using a stop hook to ask an agent to continue after it stops. The approach may help orchestrate longer agent tasks.
* [Self-play trains agents to insert and fix bugs](https://ai.exponencialadm.net/okf/en/radar/self-play-rl-treina-agentes-para-91eab0.md) - Published on December 26, 2025, the summary presents Self-play SWE-RL: a method for training a coding agent to insert and fix bugs in real repositories, without natural-language problem descriptions or human-labeled issues.
* [VL-JEPA predicts embeddings for vision and language](https://ai.exponencialadm.net/okf/en/radar/vl-jepa-preve-embeddings-de-texto-654af4.md) - Published on December 26, 2025, the summary introduces VL-JEPA, a model that predicts continuous embeddings of target text instead of generating tokens autoregressively.
* [Radar: Linux Kernel build and boot guide](https://ai.exponencialadm.net/okf/en/radar/do-codigo-ao-boot-guia-para-6aa541.md) - Published on December 25, 2025, this item presents an in-depth guide to Linux Kernel configuration, build systems, compilation, and boot.
* [Intel 64 and IA-32 manuals: architecture reference](https://ai.exponencialadm.net/okf/en/radar/manuais-intel-64-e-ia-32-97952b.md) - Published on December 25, 2025, this entry highlights Intel's official manuals on the Intel 64 and IA-32 architecture and programming environment, as a primary reference for engineers working on architecture and low-level programming.
* [MegaRAG proposes multimodal graphs for RAG](https://ai.exponencialadm.net/okf/en/radar/megarag-explora-knowledge-graphs-multimodais-para-ec6481.md) - Published on December 25, 2025, the MegaRAG abstract describes a method that automatically builds knowledge graphs from visual documents to support retrieval and cross-modal reasoning.
* [Agent-R1: modular training for LLM agents](https://ai.exponencialadm.net/okf/en/radar/agent-r1-treinamento-modular-de-agentes-00266a.md) - Published on December 24, 2025, this summary presents Agent-R1, an end-to-end reinforcement learning framework for training LLM agents in multi-turn interactions with tools and environments.
* [From C and C++ code to an executable: compilation steps](https://ai.exponencialadm.net/okf/en/radar/do-codigo-c-e-c-ao-af1fa0.md) - Learn the pipeline that turns C and C++ source code into an executable: preprocessing, compilation, assembly, and linking.
* [Kascade reuses key patterns in sparse attention](https://ai.exponencialadm.net/okf/en/radar/kascade-reutiliza-padroes-em-atencao-esparsa-b4dc6a.md) - Published on December 24, 2025, the report describes a training-free method that reuses key patterns across layers. It reports speed gains on H100 GPUs and little accuracy loss on long-context benchmarks.
* [Series introduces memory and process organization](https://ai.exponencialadm.net/okf/en/radar/memoria-e-processos-quatro-conceitos-para-0eeea5.md) - Published on December 24, 2025, Byte Lab’s series introduces virtual memory, address spaces, process anatomy, registers, stacks, threads, and allocation methods.
* [Research uses AI to optimize systems algorithms](https://ai.exponencialadm.net/okf/en/radar/pesquisa-de-desempenho-de-sistemas-orientada-837677.md) - The article presents ADRS, a framework in which language models generate, evaluate, and refine algorithms for systems performance problems. It reports applying three open-source frameworks to ten tasks in networking, databases, and distributed systems.
* [Transformer techniques in gpt-oss](https://ai.exponencialadm.net/okf/en/radar/tecnicas-para-otimizar-transformers-no-gpt-9139c8.md) - Published on December 24, 2025, the post lists optimizations used with gpt-oss in Transformers: MXFP4 quantization, tensor and expert parallelism, and dynamic sliding windows.
* [Collection catalogs prompt injection vectors](https://ai.exponencialadm.net/okf/en/radar/colecao-de-vetores-para-estudar-prompt-1be33c.md) - Published on December 23, 2025, the entry points to a collection of prompt injection vectors and strategies, with examples, that may help identify techniques during security evaluations.
* [Virtual memory: an introduction for beginners](https://ai.exponencialadm.net/okf/en/radar/memoria-virtual-fundamentos-para-entender-sistemas-d23d17.md) - An introductory article published on December 23, 2025, presenting the basics of virtual memory and its usefulness for understanding memory behavior in systems software.
* [Multimodal RL: advantage shaping in a single rollout](https://ai.exponencialadm.net/okf/en/radar/rl-multimodal-shaping-de-vantagem-em-ce326c.md) - A publication reports that single-rollout multimodal RL became stable with advantage shaping and an entropy bonus, a topic of interest to researchers studying policy-gradient methods.
* [SSR explores self-training for repository bugs](https://ai.exponencialadm.net/okf/en/radar/self-play-para-treinar-agentes-de-773b05.md) - Published on December 23, 2025, the post presents Self-play SWE-RL (SSR), a method for training a language agent to inject and fix bugs in real repositories without human-labeled issues or tests.
* [Squinch compresses gradients in pretraining](https://ai.exponencialadm.net/okf/en/radar/squinch-compressao-de-gradientes-no-pretraining-24e07c.md) - Published on December 23, 2025, the account credits the Squinch algorithm with using one quarter of IB bandwidth while maintaining SOTA MFU in pretraining on GCP H100-TCPX.
* [Iceberg broadens vector search evaluation](https://ai.exponencialadm.net/okf/en/radar/iceberg-amplia-a-avaliacao-de-busca-06a5c9.md) - A benchmark suite with eight datasets evaluates vector similarity search beyond recall and latency, examining its impact on downstream tasks such as RAG and image classification.
* [OpenTinker separates RL programming and execution](https://ai.exponencialadm.net/okf/en/radar/opentinker-separa-programacao-e-execucao-de-e26d76.md) - OpenTinker is an open-source RL-as-a-Service platform for foundation models. Its design separates programming from execution and describes developing environments locally while using remote compute.
* [Overview of RL training systems and pass@k results](https://ai.exponencialadm.net/okf/en/radar/panorama-sobre-sistemas-de-treinamento-rl-ed4a1c.md) - A December 21, 2025 publication covers on-policy rollout divergence, rollout-system design, and the effects of RL on pass@k performance.
* [SA-RAG applies spreading activation to knowledge graphs](https://ai.exponencialadm.net/okf/en/radar/sa-rag-amplia-a-busca-em-b11455.md) - Published on December 20, 2025, this item describes a retrieval approach for knowledge-graph RAG, intended to find connected evidence for multi-hop questions.
* [Why token order challenges masked models](https://ai.exponencialadm.net/okf/en/radar/por-que-a-ordem-dos-tokens-b01c33.md) - A paper recognized at ICML 2025 investigates masked models’ weaker performance on generation benchmarks and points to difficulties associated with training on arbitrary subsets of revealed tokens.
* [DEER proposes diffusion drafts to speed up LLMs](https://ai.exponencialadm.net/okf/en/radar/deer-usa-rascunhos-de-diffusion-para-34d0a4.md) - A publication dated December 18, 2025 describes DEER, which drafts tokens with diffusion models and verifies them with autoregressive models. It reports inference speedups of up to 5.54× with no loss of quality.
* [Radar: stabilizing reinforcement learning with LLMs](https://ai.exponencialadm.net/okf/en/radar/como-estabilizar-o-reinforcement-learning-com-85828e.md) - An article published on December 16, 2025 examines instability in reinforcement learning for large language models and highlights relevant differences and techniques.
* [CUDA-L2 optimizes HGEMM kernels with reinforcement learning](https://ai.exponencialadm.net/okf/en/radar/cuda-l2-aplica-aprendizado-por-reforco-3deb50.md) - Published on December 15, 2025, the paper presents an approach combining language models and reinforcement learning to optimize CUDA half-precision matrix multiplication kernels. It evaluates kernels across 1,000 configurations and reports results above matmul baselines.
* [Olmo 3: SFT, DPO and RLVR in training](https://ai.exponencialadm.net/okf/en/radar/olmo-3-sft-dpo-e-rlvr-467eeb.md) - Published on December 15, 2025, the post describes an Olmo 3 training pipeline: SFT on curated reasoning traces, followed by preference tuning with DPO and RLVR using Group Relative Policy Optimization.
* [PasoDoble explores dual-play for LLM reasoning](https://ai.exponencialadm.net/okf/en/radar/pasodoble-explora-o-dual-play-no-094231.md) - Published on December 15, 2025, the paper presents dual-play: two models take specialized roles—one poses questions and the other solves them—to reduce reliance on external supervision.
* [RouteRAG adapts retrieval across text and graphs](https://ai.exponencialadm.net/okf/en/radar/routerag-adapta-buscas-em-texto-e-1fe4e3.md) - The framework uses reinforcement learning to choose when to retrieve information from text and graphs during reasoning, with a reward that discourages unnecessary searches.
* [SWAA proposes more efficient inference with long context](https://ai.exponencialadm.net/okf/en/radar/swaa-e-inferencia-de-llms-com-778a49.md) - A December 15, 2025 post presents training-free recipes to adapt full-attention LLMs to linear scaling and recover performance.
* [Introductory C course uses a RISC-V approach](https://ai.exponencialadm.net/okf/en/radar/curso-introdutorio-de-programacao-em-c-e5e468.md) - Published on December 14, 2025, The C Language course says it teaches C from scratch and uses RISC-V programming as its learning approach. It may interest engineers looking for an introduction to the language.
* [ReCode proposes unifying planning and action in LLM agents](https://ai.exponencialadm.net/okf/en/radar/recode-aproxima-planejamento-e-acao-em-a3ea0e.md) - Published on December 13, 2025, the article presents ReCode, a paradigm using recursive code generation to let agents switch between strategic planning and detailed actions.
* [From Numba to Triton: a GPU programming path](https://ai.exponencialadm.net/okf/en/radar/de-numba-a-triton-um-caminho-978454.md) - Published on December 12, 2025, this practical Jupyter notebook introduces GPU programming with Numba and then moves to Triton, a Python-like language for writing kernels.
* [Analyzing the scaling of agent systems](https://ai.exponencialadm.net/okf/en/radar/como-prever-o-desempenho-de-sistemas-d60f01.md) - The article presents quantitative principles for modeling how coordination, model capacity, and system and task factors affect agent performance.
* [Mrav: CPU testing with Python and cocotb](https://ai.exponencialadm.net/okf/en/radar/testes-de-cpu-com-python-e-dc8e14.md) - Published on December 11, 2025, this walkthrough presents tests for a custom CPU using cocotb, Python scripts, and a class to model RAM.
* [BEAVER calculates safety bounds for LLMs](https://ai.exponencialadm.net/okf/en/radar/beaver-calcula-limites-de-seguranca-para-b5e7d1.md) - Published on December 10, 2025, this summary presents BEAVER, a framework that calculates deterministic, sound bounds on the probability that an LLM satisfies a safety property, in contrast with sampling-based estimates.
* [SEAL decomposes and verifies questions about knowledge graphs](https://ai.exponencialadm.net/okf/en/radar/seal-decompoe-perguntas-sobre-knowledge-graphs-ab927c.md) - Published on December 10, 2025, the post presents SEAL, a framework that extracts the minimal core of a question and validates it against a database to answer questions about knowledge graphs.
* [Sparse attention after post-training](https://ai.exponencialadm.net/okf/en/radar/pos-treinamento-reduz-a-atencao-a-72bfb5.md) - An article published on December 8, 2025 reports that models of up to 7 billion parameters retained their original pretraining loss with about 0.4% of attention connections.
* [Bazel toolchain for the custom Mrav CPU](https://ai.exponencialadm.net/okf/en/radar/uma-toolchain-bazel-para-uma-cpu-063f3f.md) - Published on December 4, 2025, the article presents a Bazel ecosystem for the custom Mrav CPU. One command prepares Go, builds a simulator, assembles and runs firmware, then compares CPU state with a reference file.
* [Segotep says reinforced pins reduce cable heating](https://ai.exponencialadm.net/okf/en/radar/cabos-titanload-12v-2x6-usam-pinos-112242.md) - Published on December 3, 2025, the report says Segotep’s Titanload 12V-2x6 cables use reinforced pins and that the design would reduce maximum temperatures by up to 72%. It notes that cable heating matters when engineers evaluate connectors under heavy loads.
* [Ente describes on-device OCR for Android](https://ai.exponencialadm.net/okf/en/radar/ocr-privado-no-dispositivo-o-exemplo-b4f799.md) - In a December 3, 2025 post, Ente describes bringing on-device text recognition to Ente Photos and says its OCR is fully open source.
* [Intel Arc Pro B60 Battlematrix preview reports 192 GB of GPU memory](https://ai.exponencialadm.net/okf/en/radar/intel-arc-pro-b60-battlematrix-preve-d1c3a4.md) - Published on December 1, 2025, StorageReview's preview presents Intel Project Battlematrix, a multi-GPU workstation design offering 192 GB of GPU memory for local AI.
* [Fuxi GPU with Imagination DXD architecture demonstrated](https://ai.exponencialadm.net/okf/en/radar/gpu-fuxi-com-arquitetura-imagination-dxd-1d87c7.md) - A Fuxi GPU with Imagination DXD architecture was demonstrated for Lenovo’s New Vision Z-Engine, of interest to engineers following GPU architectures and hardware platforms.
* [Open reference on data structures in C++](https://ai.exponencialadm.net/okf/en/radar/referencia-aberta-sobre-estruturas-de-dados-b0c503.md) - A November 29, 2025 publication highlights Open Data Structures as a resource focused on data structures in C++. It may be useful to engineers working with the language.
* [Video lessons on massively parallel programming](https://ai.exponencialadm.net/okf/en/radar/videoaulas-de-programacao-paralela-e-arquitetura-c36af9.md) - Published on November 29, 2025, this entry points to video lessons for “Programming Massively Parallel Processors” and highlights lesson 04 on GPU architecture.
* [SSSTC announces ER4 SATA SSDs up to 16 TB](https://ai.exponencialadm.net/okf/en/radar/ssstc-apresenta-ssds-sata-corporativos-er4-d19923.md) - In an announcement published on November 27, 2025, SSSTC introduced ER4 enterprise SATA SSDs with capacities up to 16 TB and performance of up to 98K IOPS.
* [Red Hat introduces GPU programming](https://ai.exponencialadm.net/okf/en/radar/introducao-da-red-hat-a-programacao-304df1.md) - Published on November 26, 2025, the first article in a four-part series gives a basic overview of the GPU programming model and points readers to a Cornell GPU architecture roadmap.
* [Tachyum announces open-source DDR5 TDIMM](https://ai.exponencialadm.net/okf/en/radar/tdimm-ddr5-aberto-promete-281-gb-6399f1.md) - In an announcement published on November 26, 2025, Tachyum introduced TDIMM technology compatible with DDR5 and declared memory bandwidth of 281 GB/s. The company also said modules could reach 1 TB and nodes 3 PB.
* [Tachyum announces open TDIMM DDR5 memory standard](https://ai.exponencialadm.net/okf/en/radar/tdimm-ddr5-propoe-ate-1-tb-eefb31.md) - Tachyum announced an open TDIMM DDR5 memory standard. The linked page title mentions up to 1 TB per DDR5 DIMM and 231 GB/s bandwidth; check those figures against the original source.
* [How to build a Linux container from scratch](https://ai.exponencialadm.net/okf/en/radar/como-construir-um-container-linux-do-14d2e1.md) - A tutorial published on November 25, 2025, introduces Linux primitives used to build containers, from namespaces and a root filesystem to cgroups, networking, and persistence.
* [SASS latency and instruction reordering on GPUs](https://ai.exponencialadm.net/okf/en/radar/latencia-sass-e-reordenacao-de-instrucoes-8af09a.md) - Published on November 25, 2025, the post questions the usefulness of SASS latency tables extracted with nvdisasm and says instruction reordering may improve performance by 3–4%, against a theoretical ceiling of 10%.
* [Tencent announces Hunyuan OCR with 1 billion parameters](https://ai.exponencialadm.net/okf/en/radar/tencent-anuncia-hunyuan-ocr-com-1-69315c.md) - A briefing published on November 25, 2025, describes a document-understanding model offering parsing, visual question answering, and translation. It also mentions an interactive demo and access to the model.
* [LeJEPA combines prediction and embedding regularization](https://ai.exponencialadm.net/okf/en/radar/lejepa-combina-predicao-jepa-e-sigreg-81a08b.md) - Published on November 22, 2025, the post presents LeJEPA as combining JEPA's latent-space prediction loss with SIGReg, a regularizer intended to shape embeddings into an isotropic Gaussian distribution.
* [Kneron invests in Innovedeus to bring on-device AI to education](https://ai.exponencialadm.net/okf/en/radar/kneron-investe-na-innovedeus-para-levar-0669cb.md) - Kneron’s investment in Innovedeus aims to bring schools a Raspberry Pi-like platform for machine learning and artificial intelligence running on the device itself.
* [Radar: a bare-metal C++ guide](https://ai.exponencialadm.net/okf/en/radar/guia-pratico-de-c-bare-metal-38be7d.md) - A reference by Alex Robenko on bare-metal C++ was listed on November 17, 2025. The available preview does not detail its contents.
* [Radar: introduction to CUDA kernels for array addition](https://ai.exponencialadm.net/okf/en/radar/primeiros-passos-com-kernels-cuda-para-b6dd73.md) - Published on November 16, 2025, this material introduces beginners to writing a simple CUDA kernel to add arrays. It is described as a starting point for engineers learning to create GPU kernels.
* [RobustMerge combines specialist models without retraining](https://ai.exponencialadm.net/okf/en/radar/robustmerge-reune-modelos-especialistas-sem-retreinamento-b910fe.md) - Published on November 15, 2025, the post presents RobustMerge, a method for combining task-specific fine-tuned multimodal models into a multitask model without retraining.
* [Series covers C/C++ compilers and program loading](https://ai.exponencialadm.net/okf/en/radar/serie-pratica-sobre-compiladores-e-carregamento-e66209.md) - Published on November 13, 2025, the series covers C and C++ compilers, linking, object files, shared libraries, loaders, and related topics.
* [Banana Pi BPI-R4Pro: ports and storage](https://ai.exponencialadm.net/okf/en/radar/banana-pi-bpi-r4pro-portas-e-dddd38.md) - The BPI-R4Pro router board combines a MediaTek MT7988A, four 2.5G ports, two 10G SFP+/RJ45 ports, and storage and expansion options for evaluating network appliances.
* [OCR: comparing Qwen3-VL 8B and 30B](https://ai.exponencialadm.net/okf/en/radar/qwen3-vl-8b-e-30b-trade-89c253.md) - Published on November 11, 2025, the comparison reports that the quantized 30B version was faster and more accurate at structured data extraction, but used more memory.
* [Banana Pi BPI-6202: industrial board](https://ai.exponencialadm.net/okf/en/radar/banana-pi-bpi-6202-placa-para-87f48f.md) - Published on November 10, 2025, the entry describes the BPI-6202 and its specifications for possible industrial gateway and embedded system projects.
* [Waveshare CM5 carrier board offers dual Ethernet and optional 4G/5G](https://ai.exponencialadm.net/okf/en/radar/carrier-waveshare-cm5-combina-ethernet-dupla-43b41e.md) - The CM5-DUAL-ETH-4G/5G-BASE combines network connections, video output, and expansion for storage or an AI accelerator in embedded and edge-computing projects.
* [Meta presents ASR for more than 1,600 languages](https://ai.exponencialadm.net/okf/en/radar/meta-apresenta-asr-para-mais-de-4b2c0d.md) - Published on November 10, 2025, this briefing describes Omnilingual ASR as an open-source set of speech-recognition models and a dataset for more than 1,600 languages, including 500 described as previously unsupported.
* [A suggested path to learning CUDA programming](https://ai.exponencialadm.net/okf/en/radar/um-caminho-para-aprender-programacao-cuda-3db649.md) - Recommendation published on November 9, 2025: study C/C++ and linear algebra before moving on to CUDA, using books and a course as learning resources.
* [Linux NVIDIA GPU module bugs](https://ai.exponencialadm.net/okf/en/radar/falhas-em-modulos-linux-para-gpus-5b717d.md) - Published on November 8, 2025, Quarkslab's report describes two bugs in NVIDIA Linux Open kernel modules, triggerable by an unprivileged local process. A proof of concept demonstrates kernel read and write primitives.
* [Step-Audio-EditX: prompt-based audio editing](https://ai.exponencialadm.net/okf/en/radar/modelo-propoe-edicao-de-audio-guiada-634008.md) - Announced on November 8, 2025, the model is described as a 3-billion-parameter system for audio editing, zero-shot multilingual speech synthesis, and prompt-based vocal control.
* [Gigabyte announces X870 AORUS XTREME X3D AI TOP motherboard](https://ai.exponencialadm.net/okf/en/radar/gigabyte-anuncia-placa-mae-x870-aorus-8d4c52.md) - In an announcement published on November 7, 2025, Gigabyte introduced the X870 AORUS XTREME X3D AI TOP motherboard, which includes a 5-inch LCD screen.
* [AAEON introduces a Pico-ITX board with Core Ultra](https://ai.exponencialadm.net/okf/en/radar/pico-aru4-reune-core-ultra-em-37cf18.md) - Published on November 7, 2025, the entry describes the PICO-ARU4, a compact computing board with Intel Core Ultra 5 or 7 processors and up to 32 GB of soldered LPDDR5 memory.
* [RLoop alternates RL and policy consolidation](https://ai.exponencialadm.net/okf/en/radar/rloop-combina-exploracao-e-consolidacao-no-e32671.md) - Published on November 7, 2025, the account describes an iterative training strategy for RLVR and reported gains in a math evaluation.
* [Intel Arc Pro B60 24GB arrives in Europe](https://ai.exponencialadm.net/okf/en/radar/intel-arc-pro-b60-de-24-41c914.md) - Published on November 6, 2025, the item expects the Intel Arc Pro B60 workstation GPU, with 24GB of memory, to arrive in Europe between mid- and late November, starting at €769.
* [MoE kernels explore AWS EFA](https://ai.exponencialadm.net/okf/en/radar/kernels-moe-exploram-o-aws-efa-6bfa23.md) - On November 6, 2025, Perplexity described expert-parallelism kernels for serving MoE models across multiple AWS GPUs with EFA, without GPUDirect Async.
* [Log points to Intel Xe3P variants for discrete GPUs](https://ai.exponencialadm.net/okf/en/radar/log-vazado-aponta-variantes-intel-xe3p-0e1c76.md) - A Wccftech report on a leaked log indicates that Intel Xe3P HPM may power discrete GPUs and mentions an LPM variant, probably intended for Nova Lake.
* [Radar: NeoEyes NE301 Edge AI camera](https://ai.exponencialadm.net/okf/en/radar/neoeyes-ne301-camera-edge-ai-com-9a4f35.md) - Published on November 6, 2025, the announcement describes an Edge AI camera with an STM32N6 MCU, Neural-ART NPU, 4 MP sensor, Wi-Fi 6, and Bluetooth 5.4.
* [Inworld TTS 1 Max tops speech ranking](https://ai.exponencialadm.net/okf/en/radar/tts-max-lidera-ranking-de-preferencia-7de1f5.md) - Published on November 6, 2025, the item reports Inworld TTS 1 Max leading Artificial Analysis's Speech Arena, a human-preference evaluation spanning four real-world speech categories.
* [Inworld TTS 1 Max tops speech ranking](https://ai.exponencialadm.net/okf/en/radar/tts-ranking-e-recursos-para-avaliar-31fa4e.md) - A November 6, 2025 publication reports that Inworld TTS 1 Max ranks first in Speech Arena, a blind comparison of text-to-speech models judged by people.
* [FFmpeg Assembly Lessons](https://ai.exponencialadm.net/okf/en/radar/licoes-de-assembly-no-ffmpeg-e46123.md) - Published on November 5, 2025, this entry describes lessons for learning Assembly programming in the context of FFmpeg, which says it uses hand-written Assembly to improve performance.
* [Why RL training may be more fragile than SFT](https://ai.exponencialadm.net/okf/en/radar/por-que-o-treinamento-com-rl-e9b423.md) - Published on November 5, 2025, the article explains how noisy feedback and complex infrastructure can destabilize reinforcement learning (RL) training.
* [RLAC uses an adversarial critic for free generation](https://ai.exponencialadm.net/okf/en/radar/rlac-aposta-em-critico-adversarial-para-f5fde4.md) - Published on November 5, 2025, the report describes a generator and an adversarial critic with an external validator. It reports greater factuality and better code scores with fewer calls and less data.
* [Video lessons on massively parallel programming](https://ai.exponencialadm.net/okf/en/radar/videoaulas-de-programacao-paralela-massiva-112545.md) - Published on November 4, 2025, the post presents a YouTube playlist accompanying Programming Massively Parallel Processors and highlights lesson 04, on GPU architecture.
* [Learning GPU programming with a compute-first approach](https://ai.exponencialadm.net/okf/en/radar/aprender-programacao-de-gpu-comecando-pelo-2e117c.md) - An article advocates a compute-first approach to learning GPU programming and offers engineers a perspective on how to approach that learning process.
* [CUDA C++: a step-by-step introduction](https://ai.exponencialadm.net/okf/en/radar/cuda-c-uma-introducao-passo-a-628bdf.md) - A NVIDIA post introduces parallel programming in CUDA C++ through a simple example, offering a concise starting point for engineers learning the subject.
* [Expert trajectories for model reasoning](https://ai.exponencialadm.net/okf/en/radar/trajetorias-de-especialistas-no-raciocinio-de-ca4263.md) - A framework uses expert trajectories to teach small language models to reason about difficult problems.
* [CE-GPPO and gradients in LLM RL](https://ai.exponencialadm.net/okf/en/radar/ce-gppo-busca-preservar-gradientes-no-176422.md) - Published on October 30, 2025, the summary introduces CE-GPPO, a reinforcement learning method for LLMs addressing low-probability token gradient signals discarded by PPO clipping.
* [Latent reasoning in looped language models](https://ai.exponencialadm.net/okf/en/radar/modelos-de-linguagem-em-loop-chegam-f5041b.md) - An article reports looped models with 2.6 billion parameters, trained on more than 7 trillion tokens, achieving performance equivalent to state-of-the-art language models two to three times larger.
* [AI explores autonomous discovery of reinforcement learning algorithms](https://ai.exponencialadm.net/okf/en/radar/ia-pode-descobrir-algoritmos-de-reinforcement-8444be.md) - A Nature paper reports a study led by David Silver on the autonomous discovery of reinforcement learning algorithms and the possibility of finding better methods.
* [C practice projects, from basic to advanced](https://ai.exponencialadm.net/okf/en/radar/projetos-de-pratica-em-c-do-3e63dd.md) - Published on October 28, 2025, this briefing describes a repository of C practice projects for beginner through advanced levels, without listing specific projects.
* [RE-Searcher uses goals and self-reflection for evidence-based search](https://ai.exponencialadm.net/okf/en/radar/re-searcher-objetivos-e-autorreflexao-para-ede07d.md) - Published on October 28, 2025, the summary presents RE-Searcher, a search agent that sets goals and checks whether retrieved evidence meets them, aiming to resist misleading cues in noisy environments.
* [Socratic-Zero co-evolves agents to generate reasoning data](https://ai.exponencialadm.net/okf/en/radar/socratic-zero-coevolui-agentes-para-gerar-1bdf4d.md) - The approach uses three agents and 100 initial questions to produce reasoning data; the paper reports a 20.2-point advantage across seven mathematical reasoning benchmarks.
* [KArAt proposes learnable attention without softmax](https://ai.exponencialadm.net/okf/en/radar/karat-propoe-attention-learnable-como-alternativa-23290d.md) - Published on October 27, 2025, the account presents Kolmogorov-Arnold Attention (KArAt), an alternative to attention with fixed softmax. Results in vision transformers are mixed, and low-rank variants aim to reduce memory use.
* [Ant Group presents hybrid Ring-linear 2.0 models](https://ai.exponencialadm.net/okf/en/radar/ring-linear-2-0-explora-atencao-cc8c8e.md) - A publication dated October 27, 2025 describes two models combining linear and softmax attention and claims efficiency gains from a proprietary FP8 operator library.
* [TPA proposes reducing KV cache memory with low-rank components](https://ai.exponencialadm.net/okf/en/radar/tensor-product-attention-reduz-cache-kv-9f6cdd.md) - An article published on October 27, 2025 proposes Tensor Product Attention, which represents queries, keys, and values with contextual low-rank components to seek lower KV cache memory use during inference.
* [Penny: a hand-built GPU communication library](https://ai.exponencialadm.net/okf/en/radar/penny-biblioteca-open-source-de-comunicacao-e8bfaf.md) - Published on October 26, 2025, the entry describes Penny as a hand-built, open-source GPU communication library whose implementation can be examined and whose development can receive contributions.
* [EditScore evaluates complex image edits](https://ai.exponencialadm.net/okf/en/radar/editscore-aplica-reward-models-a-edicao-8d91e5.md) - A family of 7B–72B reward models based on EditReward-Bench is presented as support for RL training and, according to the post, outperforms GPT-5 on that benchmark.
* [Meta-learning proposes update rules for RL](https://ai.exponencialadm.net/okf/en/radar/meta-aprendizado-busca-regras-para-atualizar-cf095c.md) - An article published on October 24, 2025 presents DiscoRL, a method using meta-learning across agents and environments to discover reinforcement learning update rules.
* [KL regularization in RL and mode collapse](https://ai.exponencialadm.net/okf/en/radar/rl-com-regularizacao-kl-alem-do-b5720b.md) - An article dated October 24, 2025, examines reverse KL and forward KL in reinforcement learning, questioning whether the familiar intuition about mode-seeking and mass-covering applies directly.
* [BAPO proposes stabilizing off-policy RL for LLMs](https://ai.exponencialadm.net/okf/en/radar/bapo-busca-estabilizar-o-rl-off-2ed5f9.md) - Published on October 23, 2025, the work presents BAPO, a method using balanced policy optimization and adaptive clipping for off-policy training scenarios such as partial rollout and experience reuse.
* [SFT: lower learning rates may preserve general capabilities](https://ai.exponencialadm.net/okf/en/radar/sft-e-taxas-de-aprendizado-menores-3c15cf.md) - A study published on October 23, 2025, revisits domain-specific fine-tuning of language models. Its summary says lower learning rates may preserve general capabilities while maintaining performance in the target domain.
* [Intel Nova Lake linked to a 6th-generation NPU](https://ai.exponencialadm.net/okf/en/radar/intel-nova-lake-deve-trazer-npu-b13397.md) - Published on October 22, 2025, this Radar item records a TechPowerUp report on the NPU in Intel’s future Nova Lake desktop processors.
* [MoGA proposes group-based attention for long videos](https://ai.exponencialadm.net/okf/en/radar/moga-propoe-atencao-por-grupos-para-163974.md) - The publication presents MoGA, a Mixture-of-Groups Attention approach for end-to-end long-video generation. The brief provides no implementation details or performance results.
* [olmOCR 2 uses verifiable rewards for OCR](https://ai.exponencialadm.net/okf/en/radar/olmocr-2-aposta-em-recompensas-verificaveis-4794fa.md) - Published on October 22, 2025, the post presents an OCR model for PDFs and scans, supporting tables, equations, and handwriting. Its approach combines synthetic data and unit tests as verifiable rewards.
* [Ring-1T: a one-trillion-parameter MoE model](https://ai.exponencialadm.net/okf/en/radar/ring-1t-combina-moe-de-1-2381a6.md) - Published on October 22, 2025, the summary presents Ring-1T, a reasoning MoE model with one trillion parameters and about 50 billion active per token, and describes its three-stage training.
* [How to choose and deploy open OCR models](https://ai.exponencialadm.net/okf/en/radar/como-escolher-e-implantar-modelos-open-6ed45c.md) - Published on October 21, 2025, the post covers choosing OCR models, comparing leading options, deploying them locally or remotely, and exploring uses beyond OCR.
* [Radar: TX9190 CPO switch with liquid cooling at OCP 2025](https://ai.exponencialadm.net/okf/en/radar/switch-cpo-tx9190-com-refrigeracao-liquida-f6a2e4.md) - Published on October 21, 2025, ServeTheHome reports seeing Marvell's TX9190 CPO switch with liquid cooling on display at OCP 2025.
* [Text-to-LoRA creates adapters from descriptions](https://ai.exponencialadm.net/okf/en/radar/text-to-lora-cria-adapters-a-e70a6a.md) - Published on October 21, 2025: Text-to-LoRA creates task-specific LoRA adapters for Mistral-7B-Instruct from text descriptions. Across 479 tasks, it reached 67.7% average accuracy.
* [Leak points to Ryzen 9 9950X3D2 with 192 MB of L3 cache](https://ai.exponencialadm.net/okf/en/radar/vazamento-aponta-ryzen-9-9950x3d2-com-7e0b18.md) - A report published on October 21, 2025, describes rumors of a 16-core AMD processor with 192 MB of L3 cache and a 200 W TDP; the specifications are not official confirmation.
* [WD Black SN8100 adds an 8 TB option](https://ai.exponencialadm.net/okf/en/radar/wd-black-sn8100-passa-a-ter-6c631b.md) - Published October 21, 2025: Western Digital’s WD Black SN8100 SSD is now available in an 8 TB version.
* [Banana Pi BPI-R4 Pro adds 10GbE and a Wi-Fi 7 option](https://ai.exponencialadm.net/okf/en/radar/banana-pi-bpi-r4-pro-reune-e55228.md) - Published on October 20, 2025, the announcement describes a router board with a MediaTek MT7988A SoC, 8 GB of RAM, multiple Ethernet interfaces, and a Wi-Fi 7 option.
* [FinePdfs releases resources for PDF OCR](https://ai.exponencialadm.net/okf/en/radar/finepdfs-abre-recursos-para-ocr-de-c2a42f.md) - On October 20, 2025, FinePdfs announced source code, labeled datasets, and a classifier intended for PDF OCR workflows.
* [MTEB v2 updates embedding evaluation](https://ai.exponencialadm.net/okf/en/radar/mteb-v2-atualiza-avaliacao-de-embeddings-5bba0a.md) - Published on October 20, 2025, version 2 of the MTEB suite introduces simplified evaluation, multimodal support, rerankers, new interfaces, documentation, dataset statistics, and migration guidance.
* [LaSeR explores last-token rewards for LLMs](https://ai.exponencialadm.net/okf/en/radar/laser-explora-recompensas-pelo-ultimo-token-9b651a.md) - In a post dated October 19, 2025, Tencent presents LaSeR, a reinforcement learning algorithm that aligns last-token scores with real rewards for reasoning and LLM self-rewarding. The post mentions one extra token inference.
* [Study compares RL design choices for agentic reasoning](https://ai.exponencialadm.net/okf/en/radar/o-que-comparar-no-treinamento-de-f8548a.md) - Published on October 19, 2025, the study examines data, algorithms, and reasoning modes for training agentic LLMs. Its preview highlights a comparison between concatenated synthetic trajectories and real end-to-end tool-use trajectories.
* [Router-R1 coordinates multiple models with RL](https://ai.exponencialadm.net/okf/en/radar/router-r1-coordena-varios-llms-com-e80fcc.md) - Published on October 18, 2025, the account presents Router-R1 as a reinforcement learning approach that alternates between reasoning and calls to multiple LLMs. It reports results on seven question-answering benchmarks.
* [Course explores prompt optimization with DSPy GEPA](https://ai.exponencialadm.net/okf/en/radar/curso-de-otimizacao-de-prompts-com-0fe271.md) - Published on October 17, 2025, this item shares a course on optimizing language-model prompts with DSPy GEPA, aimed at engineers. The description provides no detailed curriculum or evaluation results.
* [DiaMoE-TTS combines IPA and MoE for dialects](https://ai.exponencialadm.net/okf/en/radar/diamoe-tts-combina-ipa-e-especialistas-24c696.md) - Published on October 17, 2025, the report describes a speech synthesis framework combining IPA representation, dialect-aware experts, and adaptation methods.
* [LongCat-Audio-Codec for Speech LLMs](https://ai.exponencialadm.net/okf/en/radar/longcat-audio-codec-explora-audio-para-49231d.md) - On October 17, 2025, Meituan released an open-source tokenizer and detokenizer optimized for Speech LLMs, featuring parallel tokens and a low-latency streaming decoder.
* [Open-source RLM for long contexts](https://ai.exponencialadm.net/okf/en/radar/rlm-em-python-analisa-contextos-longos-aeeb06.md) - A Python implementation of Recursive Language Models keeps long inputs in a Python REPL for analysis and reports support for more than 100 LLM providers through LiteLLM.
* [Aalto offers a free course on parallel programming](https://ai.exponencialadm.net/okf/en/radar/curso-pratico-da-aalto-sobre-computacao-195211.md) - Aalto University offers a free, practical course on algorithm engineering for modern parallel computers, including multicore CPUs and GPUs. Its materials cover designing programs to make use of these resources.
* [Definitions and AGI benchmarks draw criticism](https://ai.exponencialadm.net/okf/en/radar/definicoes-e-benchmarks-de-agi-sob-214ef1.md) - A commentary on an AGI paper praises its discussion of capability variation but questions its benchmarks and narrow view of artificial intelligence.
* [DSPy GEPA uses reflection to optimize prompts](https://ai.exponencialadm.net/okf/en/radar/dspy-gepa-otimiza-prompts-por-reflexao-864faa.md) - A cookbook reports that reflective prompt optimization with DSPy GEPA raised accuracy on NuminaMath-1.5 by 11%, for a total cost of less than US$0.50.
* [Parallel Algorithms and Memory Hierarchy](https://ai.exponencialadm.net/okf/en/radar/algoritmos-paralelos-e-hierarquia-de-memoria-0e5238.md) - Published on October 15, 2025, the entry recommends Guy Blelloch’s book as a companion to an MIT course and highlights memory hierarchy in parallel computing.
* [RLMs explore long prompts through a REPL](https://ai.exponencialadm.net/okf/en/radar/modelos-de-linguagem-abordam-prompts-longos-0b0fc0.md) - A publication dated October 15, 2025 describes Recursive Language Models, which interact recursively with very long prompts through a REPL, and reports results on the OOLONG and BrowseComp-Plus benchmarks.
* [Study examines data and reasoning in agentic RL](https://ai.exponencialadm.net/okf/en/radar/como-avaliar-dados-e-raciocinio-em-3f09c8.md) - Published on October 14, 2025, the study examines data, algorithms, and reasoning modes in reinforcement learning for agents, including trajectories and tool use.
* [Radxa Orion O6N: Nano-ITX board with a 12-core SoC](https://ai.exponencialadm.net/okf/en/radar/radxa-orion-o6n-computador-nano-itx-586d61.md) - Published on October 14, 2025, this entry describes the Radxa Orion O6N, a Nano-ITX board with a 12-core CIX P1 SoC, up to 64 GB of LPDDR5, and a 30/45 TOPS AI accelerator.
* [SkyRL tx v0.0.2 adds multi-LoRA training](https://ai.exponencialadm.net/okf/en/radar/skyrl-tx-v0-0-2-habilita-68a3b4.md) - Published on October 14, 2025, the announcement says SkyRL tx version v0.0.2 enables multi-LoRA training through the Tinker API.
* [ACE accumulates structured context with feedback](https://ai.exponencialadm.net/okf/en/radar/ace-contexto-estruturado-com-feedback-de-4b6125.md) - Published on October 13, 2025, the summary presents Agentic Context Engineering (ACE), an approach that incrementally accumulates and refines structured context items based on performance.
* [Benchmark compares audio reasoning and latency](https://ai.exponencialadm.net/okf/en/radar/benchmark-compara-raciocinio-e-latencia-em-3b4f2c.md) - The report gives an audio reasoning model a 92% score on Big Bench Audio and compares time to first token with and without its reasoning step.
* [InfLLM-V2 switches between dense and sparse attention](https://ai.exponencialadm.net/okf/en/radar/infllm-v2-alterna-entre-atencao-densa-2b1832.md) - Published on October 13, 2025, the article presents a system that switches between dense and sparse attention to adapt processing from short to long sequences.
* [Circuits help predict reasoning errors](https://ai.exponencialadm.net/okf/en/radar/verificacao-por-circuitos-preve-erros-de-44c7ea.md) - Published on October 13, 2025, the report presents an approach that analyzes models’ internal circuits to predict incorrect reasoning and reports results on arithmetic tasks.
* [Study compares attention levels in 13 LLMs](https://ai.exponencialadm.net/okf/en/radar/estudo-testa-niveis-de-attention-em-ce50f3.md) - Experiments reported with 13 LLMs found the best result at 17% attention, equivalent to two of 12 layers. The comparison may inform architecture choices, but does not establish a universal rule.
* [Guide surveys LLM post-training methods](https://ai.exponencialadm.net/okf/en/radar/guia-de-post-training-de-llms-de11c7.md) - Published on October 12, 2025, the guide covers the shift from next-token prediction to instruction following, along with methods and evaluation.
* [SanDisk proposes high-bandwidth flash memory for AI](https://ai.exponencialadm.net/okf/en/radar/high-bandwidth-flash-uma-proposta-para-a31f68.md) - In a publication dated October 12, 2025, SanDisk proposed High Bandwidth Flash, a NAND design aiming for far greater capacity than HBM with similar read bandwidth and price.
* [SwiReasoning switches reasoning modes](https://ai.exponencialadm.net/okf/en/radar/swireasoning-alterna-raciocinio-latente-e-explicito-8dae28.md) - A publication dated October 12, 2025 describes a method that switches between latent and explicit reasoning and reports token-efficiency gains on Qwen3-8B models.
* [BPI-AIM7: an AI module with RK3588](https://ai.exponencialadm.net/okf/en/radar/banana-pi-bpi-aim7-modulo-de-2e9926.md) - Published on October 11, 2025, the item describes the BPI-AIM7 as a low-power AI module based on RK3588 and compatible with the Jetson Nano ecosystem.
* [TUMIX coordinates agents to refine responses](https://ai.exponencialadm.net/okf/en/radar/tumix-coordena-agentes-para-refinar-respostas-9044f7.md) - Published on October 11, 2025, the post presents TUMIX: agents with different tools solve problems independently, share responses, and refine them over several rounds.
* [Agent learning without reward signals](https://ai.exponencialadm.net/okf/en/radar/aprendizado-de-agentes-com-estados-futuros-7aaf9b.md) - Published on October 10, 2025, the post describes using future states from an agent’s own interactions as supervision, without reward signals, and examines two strategies: implicit world modeling and self-reflection.
* [AHN proposes continuous memory for LLMs](https://ai.exponencialadm.net/okf/en/radar/ahn-propoe-memoria-para-llms-de-b99fe9.md) - A publication dated October 9, 2025, reports that ByteDance introduced Artificial Hippocampus Networks (AHN), an architecture for long-context LLMs that continuously compresses information outside the context window.
* [Study links attention sinks and compression valleys in LLMs](https://ai.exponencialadm.net/okf/en/radar/attention-sinks-e-vales-de-compressao-f9c320.md) - Experiments on models ranging from 410 million to 120 billion parameters associate extreme activation norms of the beginning-of-sequence token in intermediate layers with compression valleys and attention sinks.
* [A theory proposes three stages for information flow in LLMs](https://ai.exponencialadm.net/okf/en/radar/tres-etapas-para-pensar-o-fluxo-e4e2d9.md) - A publication dated October 9, 2025 links massive activations, attention sinks, and compression valleys, and proposes three stages for information flow in language models.
* [JEPA-SCORE estimates density using JEPA encoders](https://ai.exponencialadm.net/okf/en/radar/jepa-score-estima-densidade-sem-retreinar-8f6556.md) - A study published on October 8, 2025 reports that an anti-collapse term in JEPAs can implicitly estimate data density and presents JEPA-SCORE, which does so without retraining.
* [UserRL trains agents with simulated users](https://ai.exponencialadm.net/okf/en/radar/userrl-treina-agentes-com-usuarios-simulados-f30170.md) - Published on October 8, 2025, Salesforce’s post presents UserRL, a framework for training agents with standardized environments and simulated users.
* [CCA reduces the space used by attention](https://ai.exponencialadm.net/okf/en/radar/cca-leva-a-attention-para-um-d4b9af.md) - Published on October 7, 2025, the post describes Compressed Contextual Attention (CCA), which computes attention in a smaller latent space and applies RoPE there.
* [CCA proposes compressed attention for long contexts](https://ai.exponencialadm.net/okf/en/radar/cca-propoe-atencao-comprimida-para-transformers-4f9c56.md) - Published on October 7, 2025, the article presents Compressed Convolutional Attention (CCA), which projects queries, keys, and values to smaller dimensions to perform attention in a compressed latent space.
* [FireRedChat explores full-duplex voice interaction](https://ai.exponencialadm.net/okf/en/radar/fireredchat-explora-conversa-por-voz-em-ffa51d.md) - Published on October 7, 2025, this bookmark describes an open-source modular voice interaction system, with cascade and semi-cascade implementations and features for real-time conversation.
* [GPU MODE analyzes the AF3 trimul kernel competition](https://ai.exponencialadm.net/okf/en/radar/gpu-mode-analisa-a-competicao-do-cf643c.md) - A GPU MODE analysis covers the optimization challenges of the AF3 trimul kernel, how the competition was planned, and what participants implemented.
* [CyberBlue bundles blue-team tools in containers](https://ai.exponencialadm.net/okf/en/radar/cyberblue-reune-ferramentas-blue-team-em-23163b.md) - Published on October 6, 2025, the entry describes CyberBlue as a container-based platform bundling open-source tools for SIEM, DFIR, CTI, SOAR, and network analysis.
* [Index catalogs public proof-of-concept exploits for CVEs](https://ai.exponencialadm.net/okf/en/radar/indice-reune-exploits-publicos-de-prova-2afe2c.md) - Published on October 6, 2025, the briefing describes a GitHub project indexing public proof-of-concept exploits from five sources to support vulnerability research and security testing.
* [Model explores neural dynamics and synchronization](https://ai.exponencialadm.net/okf/en/radar/modelo-explora-sincronizacao-em-nivel-de-5bb98a.md) - Published on October 6, 2025, the article presents the Continuous Thought Machine, which uses neuron-level processing and synchronization as core representations inspired by biological neural dynamics.
* [MoE-CL for continual instruction tuning](https://ai.exponencialadm.net/okf/en/radar/moe-cl-adaptacao-continua-com-experts-0c48ae.md) - A paper published on October 6, 2025 proposes the MoE-CL framework for continual instruction tuning and reports evaluations on the MTL5 and Tencent3 benchmarks, plus an A/B test on Tencent. The available summary gives no quantitative results.
* [Study compares in-context and in-weight learning](https://ai.exponencialadm.net/okf/en/radar/estudo-explora-aprendizagem-por-exemplos-e-51db3e.md) - Published on October 5, 2025, the study reports that a model trained on 12,000 small tasks combined new ideas and found a balance between flexibility and memory.
* [GEPA and detecting AI-generated malicious code](https://ai.exponencialadm.net/okf/en/radar/gepa-testa-prompts-para-detectar-codigo-bc6274.md) - A publication dated October 5, 2025 reports that prompts found with GEPA detected malicious behavior in AI-generated code, blocking 90% of malicious submissions with 1% of the audit budget.
* [Selection features papers on search and generative embeddings](https://ai.exponencialadm.net/okf/en/radar/pesquisa-em-recuperacao-de-informacao-embeddings-438426.md) - Published on October 5, 2025, volume 124 of a weekly paper selection covers generative embeddings with test-time scaling and efficient retrievers for visual documents.
* [RLAD tests short hints to guide LLM reasoning](https://ai.exponencialadm.net/okf/en/radar/rlad-testa-dicas-curtas-para-orientar-fa31a6.md) - A study trains a hint generator and a hint-conditioned solver with reinforcement learning. The report says it achieved 44% higher math accuracy on AIME 2025 than RL with long chains.
* [Grokking in modular addition: a three-stage explanation](https://ai.exponencialadm.net/okf/en/radar/grokking-em-adicao-modular-tres-etapas-0a45e2.md) - A post published on October 4, 2025 describes a paper that presents grokking in modular addition as three stages and proposes an example threshold for generalization.
* [Slides link kernel regression to attention](https://ai.exponencialadm.net/okf/en/radar/regressao-por-kernel-oferece-uma-lente-834441.md) - Published on October 4, 2025, a ten-slide summary examines the relationship between kernel regression and the attention mechanism.
* [Maxsun and Abee announce a four-GPU workstation](https://ai.exponencialadm.net/okf/en/radar/workstation-reune-quatro-gpus-arc-pro-5a4318.md) - Announcement dated October 4, 2025: Maxsun and Abee presented a workstation with four Intel Arc Pro B60 Dual 48G GPUs, for a total of 192 GB of VRAM.
* [Dragon Hatchling explores local neural memory in language models](https://ai.exponencialadm.net/okf/en/radar/dragon-hatchling-explora-regras-neuronais-locais-ed0cfe.md) - Published on October 3, 2025, the post describes Dragon Hatchling, a model whose connections store short-term memory through Hebbian learning, and its GPU-optimized version.
* [RLAD explores procedural hints for reasoning in LLMs](https://ai.exponencialadm.net/okf/en/radar/rlad-explora-abstracoes-de-raciocinio-com-5d66c5.md) - Published on October 3, 2025, this record presents RLAD, a two-player reinforcement learning framework for discovering natural-language hints that structure reasoning exploration.
* [ckit.h offers utilities for C projects](https://ai.exponencialadm.net/okf/en/radar/ckit-h-reune-utilitarios-para-projetos-25d3e5.md) - Published on October 2, 2025, ckit.h is described as a small kit of essentials for C projects that engineers can assess as a source of reusable utilities.
* [SDL3 can fill the browser window in Emscripten](https://ai.exponencialadm.net/okf/en/radar/sdl3-pode-preencher-a-janela-do-dc478c.md) - A publication dated October 2, 2025, shows an option for SDL3 games compiled with Emscripten to fill the browser window.
* [Inside NVIDIA GPUs: matmul kernels](https://ai.exponencialadm.net/okf/en/radar/como-kernels-de-matmul-aproveitam-gpus-abfec7.md) - An article published October 1, 2025, covering GPU architecture and kernel techniques for high-performance matrix multiplication.
* [Dragon Hatchling explores locally interacting neural particles](https://ai.exponencialadm.net/okf/en/radar/dragon-hatchling-explora-particulas-neuronais-de-da9463.md) - The Dragon Hatchling architecture combines locally interacting neural particles in a biologically inspired, scale-free network, with an emphasis on interpretability.
* [GPU kernels, graphs and profiling](https://ai.exponencialadm.net/okf/en/radar/kernels-graficos-e-profiling-em-gpus-975f99.md) - A blog post covers kernels, graphs, profiling and related topics, potentially offering engineers an overview of GPU programming.
* [ToolUniverse connects LLMs to research tools](https://ai.exponencialadm.net/okf/en/radar/tooluniverse-conecta-llms-a-ferramentas-de-2cb75c.md) - Published on September 30, 2025, the post presents a shared interface for discovering and calling local and remote research tools via MCP. The library includes more than 600 tools and a drug-discovery demonstration.
* [Guide covers high-performance CUDA matmul kernels](https://ai.exponencialadm.net/okf/en/radar/guia-de-kernels-cuda-para-matmul-e39331.md) - Published on September 29, 2025, the material covers GPU memory hierarchy, the CUDA programming model, and PTX/SASS inspection to guide compiler output for matmul kernels.
* [LoRA recipes for reinforcement learning fine-tuning](https://ai.exponencialadm.net/okf/en/radar/receitas-de-lora-para-fine-tuning-2e1365.md) - A Thinky Machines post dated September 29, 2025 highlights LoRA configuration choices for reinforcement learning fine-tuning, including 10× higher learning rates and LoRA in every layer.
* [Energy-based transformers refine next-token prediction](https://ai.exponencialadm.net/okf/en/radar/energy-based-transformers-refinam-previsoes-do-a71479.md) - Published on September 27, 2025, the report describes an approach that iteratively reduces the energy assigned to next-token candidates. In tests with 44-million-parameter models on RedPajama-Data-v2, it outperformed same-sized vanilla transformers on three of four benchmarks.
* [FlashAttention-4: kernel optimization](https://ai.exponencialadm.net/okf/en/radar/flashattention-4-engenharia-reversa-e-ganho-26a23d.md) - Published on September 26, 2025, Modal’s post describes reverse engineering FlashAttention-4 and attributes a 20% performance gain to changes in computation and execution.
* [RLPT applies reinforcement learning to pre-training data](https://ai.exponencialadm.net/okf/en/radar/rlpt-leva-aprendizado-por-reforco-aos-b52897.md) - A September 25, 2025 publication presents RLPT, which uses next-segment rewards to train models on pre-training data without additional human annotations. It reports benchmark gains for Qwen3-4B-Base.
* [Stanford course surveys self-improving agents](https://ai.exponencialadm.net/okf/en/radar/stanford-reune-estudos-sobre-agentes-que-7b8c54.md) - Published on September 25, 2025, the briefing introduces Stanford’s CS329A course and names three works on self-improving AI agents.
* [Linux boot chapter updated](https://ai.exponencialadm.net/okf/en/radar/capitulo-atualizado-sobre-a-inicializacao-do-de70f5.md) - Update dated September 24, 2025: the fourth linux-insides chapter on Linux kernel boot has been revised to cover modern versions.
* [Meta announces open 32-billion-parameter model for code](https://ai.exponencialadm.net/okf/en/radar/meta-apresenta-modelo-de-codigo-com-288aa5.md) - On September 24, 2025, Meta introduced Code World Model, a 32-billion-parameter language model with open weights, aimed at research on code generation with world models.
* [RF-DETR applied to basketball detection](https://ai.exponencialadm.net/okf/en/radar/rf-detr-ajustado-para-detectar-elementos-941cf4.md) - A September 24, 2025 publication describes a fine-tuned RF-DETR model for detecting players, jersey numbers, referees, the ball, and the hoop in basketball scenes.
* [Radar: a unified view of Attention-MoE and FFN-MoE](https://ai.exponencialadm.net/okf/en/radar/uma-visao-unificada-de-attention-moe-f75bd5.md) - A September 24, 2025 publication describes treating pre-mixed tokens and (V·Wo) as an FFN expert, unifying Attention-MoE and FFN-MoE with shared experts.
* [Radar: cache, prefetch, false sharing and branches](https://ai.exponencialadm.net/okf/en/radar/cache-prefetch-e-branches-hardware-que-489e05.md) - A light reading, published on September 23, 2025, introduces hardware fundamentals for programmers: cache, prefetch, false sharing, and branches.
* [LIMI tests data curation for agents](https://ai.exponencialadm.net/okf/en/radar/curadoria-de-dados-e-autonomia-de-89e450.md) - Published on September 23, 2025, the LIMI report says 78 selected demonstrations achieved 73.5% on AgencyBench, outperforming models trained on 10,000 samples.
* [MIT report describes magnetic-semiconductor transistor](https://ai.exponencialadm.net/okf/en/radar/transistor-magnetico-pode-integrar-controle-e-686522.md) - Published on September 23, 2025, the report attributes a transistor made with a magnetic semiconductor to MIT researchers. It says magnetism improves electrical control and could integrate memory into the device.
* [Triton: understanding `tl.make_block_ptr`](https://ai.exponencialadm.net/okf/en/radar/triton-entendendo-tl-make-block-ptr-8afde8.md) - A post introduces how tensors are arranged in GPU memory and explains strides, offsets, and `tl.make_block_ptr`, using visual resources to help reason about data access in Triton kernels.
* [CUDA: 256-bit operations on the GPU](https://ai.exponencialadm.net/okf/en/radar/cuda-aritmetica-de-256-bits-na-87d203.md) - A tutorial published on September 22, 2025 covers implementing 256-bit multiplication and addition on a GPU with CUDA; the full code is said to be in the comments.
* [Mini-Omni-Reasoner interleaves reasoning and speech](https://ai.exponencialadm.net/okf/en/radar/mini-omni-reasoner-intercala-raciocinio-e-0fbb51.md) - Published on September 22, 2025, the post describes an architecture alternating silent reasoning and spoken tokens, a new dataset, and reported gains in arithmetic reasoning and contextual understanding.
* [MIT study tests theory-based prompts for social agents](https://ai.exponencialadm.net/okf/en/radar/prompts-teoricos-para-agentes-sociais-generalizaveis-4a19af.md) - Published on September 22, 2025, the post describes a study using behavioral theory and small human datasets to create prompts. Agents were evaluated in related but distinct scenarios and reportedly had lower prediction error on new game variants than language models and baselines.
* [Qwen3-Omni combines speech, text, images and video](https://ai.exponencialadm.net/okf/en/radar/qwen3-omni-reune-fala-texto-imagem-766469.md) - On September 22, 2025, Qwen announced the release of three multimodal models that can follow instructions, reason and generate captions, with speech input and output.
* [Penny: a development diary for GPUs](https://ai.exponencialadm.net/okf/en/radar/penny-um-diario-de-desenvolvimento-para-3d1b49.md) - The first installment in a series follows the manual creation of Penny, a GPU communication library, offering context for engineers interested in this kind of development.
* [Video guide to practicing CUDA in C on Colab](https://ai.exponencialadm.net/okf/en/radar/pratique-cuda-em-c-gratuitamente-no-2dff60.md) - Published on September 20, 2025, the item presents a video guide created by its author for free practice in GPU programming with CUDA in C on Google Colab.
* [Brook and stream computing for GPUs](https://ai.exponencialadm.net/okf/en/radar/brook-e-o-inicio-da-computacao-174ffe.md) - Published on September 19, 2025, the entry highlights a 2004 paper on kernels processing elements in streams as a GPU programming model.
* [GPU and AI performance reading list](https://ai.exponencialadm.net/okf/en/radar/como-estudar-performance-de-gpus-e-7cef4b.md) - Published on September 18, 2025, this bookmark collects material on CUDA, H100 and cuBLAS matmul optimization, nondeterminism in LLM inference, transformers, scaling, and hardware-model co-design.
* [Graphcore IPU: architecture and parallelism](https://ai.exponencialadm.net/okf/en/radar/ipu-da-graphcore-nucleos-threads-e-675408.md) - A post published on September 18, 2025 describes an IPU with 1,472 processing cores, nearly 9,000 parallel threads, and 900 MB of in-processor memory, aimed at graph-based computing and irregular, sparse workloads.
* [Tool transcribes long recordings](https://ai.exponencialadm.net/okf/en/radar/ferramenta-em-python-transcreve-audio-longo-1d2f1f.md) - Published on September 17, 2025, a Python command-line tool transcribes long audio and video through a speech-recognition API.
* [FireRedTTS 2: an open text-to-speech model](https://ai.exponencialadm.net/okf/en/radar/fireredtts-2-amplia-opcoes-de-sintese-272cb1.md) - Published on September 17, 2025, the announcement describes an open text-to-speech model supporting long conversational speech, multilingual and cross-lingual voice cloning, low latency, and random timbre generation.
* [PHLoRA extracts LoRA adapters from fine-tuned models](https://ai.exponencialadm.net/okf/en/radar/phlora-extrai-adapters-lora-de-checkpoints-bc24b7.md) - Published on September 17, 2025, the post describes a method for extracting low-rank adapters from fine-tuned checkpoints without training data or gradients.
* [GPU programming resources and an introduction to GPU systems](https://ai.exponencialadm.net/okf/en/radar/recursos-de-programacao-para-gpus-e-634bf9.md) - A bookmark published on September 17, 2025, describes a collection of GPU programming resources, including a chapter on architecture and interconnects in the context of LLM workloads.
* [REFRAG proposes faster RAG with reusable embeddings](https://ai.exponencialadm.net/okf/en/radar/refrag-acelera-rag-com-embeddings-reutilizaveis-093a48.md) - Published on September 15, 2025, the report describes REFRAG, a method that replaces most retrieved tokens with precomputed, reusable chunk embeddings. The post reports speed and context gains.
* [FireRedTTS-2 explores streaming speech with multiple speakers](https://ai.exponencialadm.net/okf/en/radar/tokens-de-streaming-para-fala-com-63bf89.md) - Published on September 15, 2025, the summary describes a streaming speech synthesis system for long, contextual dialogue, using 12.5 Hz speech tokens and a dual-transformer architecture.
* [Cornell workshop introduces GPU architecture](https://ai.exponencialadm.net/okf/en/radar/workshop-da-cornell-introduz-a-arquitetura-3a8e23.md) - Published on September 13, 2025, this entry presents a Cornell GPU architecture workshop as a starting point for learning about GPUs and highlights architectural differences from CPUs.
* [Research agents trained with reinforcement learning](https://ai.exponencialadm.net/okf/en/radar/agentes-autonomos-para-pesquisa-em-tarefas-3b823d.md) - A paper dated September 9, 2025 describes autonomous agents for deep research using web search, browsing, Python, and summaries of earlier results when context is limited.
* [Deep research strategies compiled as code](https://ai.exponencialadm.net/okf/en/radar/universal-deep-research-transforma-estrategias-em-7cfee2.md) - A September 7, 2025 entry describes Universal Deep Research as a system that turns natural-language research strategies into executable functions using search and an interchangeable language model.
* [Radar: designing a SIMD algorithm from scratch](https://ai.exponencialadm.net/okf/en/radar/como-projetar-um-algoritmo-simd-do-dc05ab.md) - Published on August 17, 2025, the article presents an example of developing a SIMD algorithm from scratch, focused on Base64.
* [RLVMR combines rewards with step supervision](https://ai.exponencialadm.net/okf/en/radar/rlvmr-combina-recompensas-finais-e-supervisao-b7612b.md) - In a publication dated August 15, 2025, Tencent researchers proposed RLVMR, a method combining outcome rewards with supervision of agent steps. The text reports 83.6% on the hardest unseen tasks in ALFWorld and ScienceWorld for a 7B agent.
* [Selected resources for exploring GraphRAG](https://ai.exponencialadm.net/okf/en/radar/recursos-selecionados-para-explorar-graphrag-01eae9.md) - Published on August 14, 2025, this briefing presents a curated list of surveys, papers, benchmarks, and open-source projects on graph-based retrieval-augmented generation.
* [LogicRAG plans reasoning at query time](https://ai.exponencialadm.net/okf/en/radar/logicrag-monta-um-grafo-de-raciocinio-69b95d.md) - A post published on August 12, 2025 describes a method that plans a DAG for each query, retrieves evidence in sequence, and updates an incremental summary. It says the method outperformed GraphRAG systems on three multi-hop QA datasets.
* [SWE-Swiss: a training recipe for software engineering models](https://ai.exponencialadm.net/okf/en/radar/swe-swiss-uma-receita-para-treinar-4944e8.md) - Researchers from Peking University, ByteDance Seed, and HKU present a training approach for software engineering tasks. The SWE-Swiss-32B model, with 32 billion parameters, scored 60.2% on SWE-bench Verified, according to the authors.
* [Sipeed Maix4-HAT brings an NPU to Raspberry Pi](https://ai.exponencialadm.net/okf/en/radar/maix4-hat-leva-aceleracao-transformer-a-5b02e3.md) - Published on August 11, 2025, the Maix4-HAT for Raspberry Pi 5 combines an AXera AX650N NPU and 8 GB of LPDDR4x RAM to accelerate Transformer models at the edge.
* [Survey examines memory in LLM-based agents](https://ai.exponencialadm.net/okf/en/radar/survey-examina-memoria-em-agentes-baseados-492b3a.md) - A publication from Renmin University of China and Huawei covers memory design, evaluation, applications, limitations, and future directions for LLM-based agents.
* [Zerox extracts structured fields from documents](https://ai.exponencialadm.net/okf/en/radar/zerox-extrai-campos-estruturados-de-documentos-358ee1.md) - Published on August 10, 2025, the post presents Zerox, an open-source OCR and document extraction project that uses a schema to retrieve specific information without converting the entire document to Markdown.
* [XIAOML Kit brings ML to embedded hardware](https://ai.exponencialadm.net/okf/en/radar/xiaoml-kit-leva-estudos-de-ml-2e49f3.md) - Published on August 7, 2025, this item describes a kit with ESP32-S3 boards and sensors paired with a free machine learning systems book and hands-on lessons.
* [Modal’s streaming STT example](https://ai.exponencialadm.net/okf/en/radar/transcricao-de-fala-em-streaming-exemplo-fedbfa.md) - Published on August 6, 2025, Modal’s example presents a deployment of a streaming audio transcription service using Kyutai STT.
* [Mirage compiles LLMs into persistent megakernels](https://ai.exponencialadm.net/okf/en/radar/mirage-compila-llms-em-megakernels-persistentes-58a52e.md) - The Mirage Persistent Kernel compiler turns LLMs into optimized megakernels. The publication says the approach reduces latency by 1.2–6.7×.
* [Dream Router 7 review: WiFi 7 and 10GbE](https://ai.exponencialadm.net/okf/en/radar/dream-router-7-wifi-7-e-608074.md) - Published on June 4, 2025, StorageReview’s review examines the Ubiquiti Dream Router 7 gateway and its features for homes and small and midsize businesses.
* [DocEx simulates document exfiltration](https://ai.exponencialadm.net/okf/en/radar/docex-simula-exfiltracao-de-documentos-sensiveis-b19527.md) - Published on April 20, 2025, the entry describes DocEx, an APT emulation tool for exfiltrating office files and PDFs to assess security defenses.
* [Chrome: V8 and JavaScript internals](https://ai.exponencialadm.net/okf/en/radar/exploracao-do-chrome-uma-introducao-aos-ad10a2.md) - A three-part series on Chrome internals and browser exploitation; part one introduces V8 and JavaScript internals.
* [Series covers antivirus evasion with syscalls and XOR](https://ai.exponencialadm.net/okf/en/radar/syscalls-diretas-e-xor-contexto-defensivo-b9949b.md) - Published on April 19, 2025, the material describes a two-part series on evading Windows Defender, including lab setup, direct syscalls, and XOR. Part one provides basic code for experimentation.
* [olmOCR proposes plain-text PDF extraction](https://ai.exponencialadm.net/okf/en/radar/olmocr-extrai-texto-simples-de-pdfs-2e6467.md) - In a February 25, 2025 post, AllenAI introduced olmOCR, an open-source tool for extracting plain text from PDFs, described as handling various document types and running on a GPU.
* [Expressive TTS beta and cross-platform local launcher](https://ai.exponencialadm.net/okf/en/radar/beta-de-tts-com-clonagem-de-b6fd2f.md) - On February 15, 2025, a post announced a beta for expressive speech synthesis with voice cloning, Transformer and SSM-hybrid models under Apache 2.0, and a one-click launcher for Mac, Windows, and Linux.
* [Hibiki: streaming speech translation](https://ai.exponencialadm.net/okf/en/radar/hibiki-traducao-de-fala-em-streaming-5c8d95.md) - Published on February 6, 2025, Hibiki is described as a decoder-only model for simultaneous speech translation between French and English, trained on synthetic data.
* [Moonshine integrated into Transformers](https://ai.exponencialadm.net/okf/en/radar/moonshine-chega-ao-transformers-para-transcricao-92f135.md) - A January 20, 2025 post announces integration of the Moonshine automatic speech recognition model with Transformers and highlights support for audio of any duration.
* [HPAC-IDS applies hierarchical attention to intrusion detection](https://ai.exponencialadm.net/okf/en/radar/hpac-ids-aplica-atencao-hierarquica-a-b216c8.md) - Published on January 14, 2025, the summary describes a method that segments raw network packets and uses hierarchical structures and self-attention. Experiments on CIC-IDS2017 report high accuracy, few false positives, and adversarial resilience.
* [VideoRAG retrieves videos for multimodal answers](https://ai.exponencialadm.net/okf/en/radar/videorag-amplia-a-recuperacao-multimodal-de-0f2864.md) - Published on January 14, 2025, this brief describes VideoRAG, a KAIST research framework that retrieves videos relevant to queries and combines visual and textual information in generated answers.
* [Piper: locally run neural speech synthesis](https://ai.exponencialadm.net/okf/en/radar/piper-sintese-de-fala-neural-em-c3905d.md) - Piper converts text to speech with multiple voices, phonemization libraries, multilingual support, and customizable options. Engineers can evaluate it as a local text-to-speech alternative.
* [TheAgentCompany assesses agents on workplace tasks](https://ai.exponencialadm.net/okf/en/radar/theagentcompany-avalia-agentes-em-tarefas-de-34b142.md) - Published on December 19, 2024, TheAgentCompany is presented as a benchmark for assessing AI agents on software development, project management, administration, and data science tasks. The material provides no results or scores.
* [Graphiti explores temporal knowledge graphs](https://ai.exponencialadm.net/okf/en/radar/graphiti-explora-grafos-temporais-de-conhecimento-15e926.md) - Published on December 10, 2024, this entry presents Graphiti as a tool for creating and querying dynamic graphs that represent relationships between entities over time.
* [Auto-RAG automates iterative retrieval in RAG systems](https://ai.exponencialadm.net/okf/en/radar/auto-rag-refina-buscas-de-forma-795ec5.md) - Published on December 2, 2024, the post describes a method that plans searches and refines queries across multiple interactions, adjusting the number of iterations to the question's difficulty.
* [HijackRAG explores injected text attacks on RAG systems](https://ai.exponencialadm.net/okf/en/radar/hijackrag-expoe-riscos-de-textos-injetados-c81029.md) - Published on December 1, 2024, the post presents HijackRAG, an attack that manipulates RAG systems through retrieved content, attention diversion, and instructions.
* [RARe adapts retrieval models with retrieved examples](https://ai.exponencialadm.net/okf/en/radar/rare-adapta-modelos-de-retrieval-com-226d57.md) - Published on December 1, 2024, the report presents RARe, which adds instructions and semantically similar query-document pairs retrieved with BM25 to retrieval models. The post says similar examples outperformed random ones, with gains increasing as the number of examples rose, up to the 10 tested.
* [AssistRAG separates memory from generation](https://ai.exponencialadm.net/okf/en/radar/assistrag-separa-memoria-da-geracao-de-fcdd76.md) - Published on November 30, 2024, AssistRAG proposes managing memory and external knowledge without retraining the main model that generates answers.
* [LazyGraphRAG combines vector and graph retrieval](https://ai.exponencialadm.net/okf/en/radar/lazygraphrag-combina-busca-vetorial-e-em-55978d.md) - A post published on November 29, 2024 describes an approach combining vector RAG with iterative breadth-first and best-first search, without index summarization. It identifies possible uses and trade-offs for assessing alternatives to full GraphRAG.
* [Notebook demos RAG with PDFs](https://ai.exponencialadm.net/okf/en/radar/rag-com-pdfs-docling-e-weaviate-b805ce.md) - Published on November 28, 2024, the notebook presents a recipe for retrieval-augmented generation over PDF files using Docling and Weaviate. Docling is described as an open-source Python package that converts documents to Markdown or JSON.
* [DeepSpeed Domino for distributed LLM training](https://ai.exponencialadm.net/okf/en/radar/deepspeed-domino-paralelismo-para-treinar-llms-e516c6.md) - A post published on November 26, 2024 presents Domino as a tensor-parallelism mechanism intended to reduce communication overhead in LLM training.
* [Autoflow combines Graph RAG and TiDB Serverless](https://ai.exponencialadm.net/okf/en/radar/autoflow-explora-graph-rag-com-tidb-e6ed34.md) - A post dated November 22, 2024 describes Autoflow as a conversational knowledge base built on Graph RAG, using TiDB Serverless for vector storage.
* [OpenScholar releases code, models, and research corpus](https://ai.exponencialadm.net/okf/en/radar/openscholar-abre-codigo-modelos-e-acervo-d6b03f.md) - On November 19, 2024, the OpenScholar project announced code, model checkpoints, a corpus of over 45 million papers, and the ScholarQABench benchmark.
* [Ultravox: a multimodal model for real-time voice](https://ai.exponencialadm.net/okf/en/radar/ultravox-um-llm-multimodal-para-voz-1eb5e0.md) - Published on November 18, 2024, this entry presents Ultravox as a fast multimodal LLM for real-time voice. It says the source code is available on GitHub for examining and assessing its approach.
* [KAR expands queries with knowledge graphs](https://ai.exponencialadm.net/okf/en/radar/kar-amplia-queries-de-retrieval-com-e780a8.md) - Published on November 13, 2024, this entry describes a workflow using model-extracted entities, relevant documents, and knowledge-graph relations to generate context-aware query expansions.
* [Gradio app adds support for multilingual voice fine-tunes](https://ai.exponencialadm.net/okf/en/radar/app-gradio-amplia-suporte-a-ajustes-c89666.md) - An update to the e2-f5-tts Gradio app enables testing custom F5-TTS fine-tunes for Spanish, Hungarian, German, Portuguese, and Greek.
* [FACT rewrites context to retrieve multiple facts](https://ai.exponencialadm.net/okf/en/radar/fact-reescreve-o-contexto-para-recuperar-f80338.md) - Published on October 30, 2024, the paper presents FACT, an iterative context-rewriting approach for retrieving multiple facts. It examines how important information can be lost in long contexts and lead to incomplete answers.
* [SimpleQA evaluates model factuality](https://ai.exponencialadm.net/okf/en/radar/simpleqa-como-avaliar-factualidade-e-alucinacoes-f8e34a.md) - Published on October 30, 2024, SimpleQA contains 4,000 human-written factual questions, each with a single indisputable answer. An autograder labels answers correct, incorrect, or not attempted; two annotators verified the reference answers.
* [MaskGCT: zero-shot TTS with masked codec](https://ai.exponencialadm.net/okf/en/radar/maskgct-e-clonagem-de-voz-com-a22836.md) - A paper published on October 24, 2024 presents MaskGCT, a zero-shot text-to-speech model. The publication says it can clone a voice from five seconds of speech.
* [E2-F5-TTS adds a fine-tuning interface](https://ai.exponencialadm.net/okf/en/radar/e2-f5-tts-ganha-interface-para-eef1e8.md) - A post dated October 21, 2024, says the E2-F5-TTS project added a Gradio app for fine-tuning the model with user-provided audio, with automatic transcription of audio clips.
* [White-hat script aims to recover funds from compromised wallets](https://ai.exponencialadm.net/okf/en/radar/script-white-hat-tenta-recuperar-fundos-8162df.md) - An open-source Bash script combines Linux tools with Foundry’s cast and chisel to try to move funds before sweep bots do. Security professionals should assess the approach and its dependencies.
* [RAG technique tutorials in notebooks](https://ai.exponencialadm.net/okf/en/radar/tutoriais-praticos-exploram-tecnicas-de-rag-3c9693.md) - Published on October 15, 2024, the RAG_Techniques repository collects notebooks on retrieval-augmented generation, including reranking, chunking, filtering, and Graph RAG.
* [LightRAG combines graphs with two-level retrieval](https://ai.exponencialadm.net/okf/en/radar/lightrag-combina-grafos-e-recuperacao-em-5b554e.md) - Published on October 13, 2024, this bookmark describes LightRAG, a RAG framework using graph-based text indexing and low- and high-level retrieval.
* [Speech-to-speech toolkit for voice agents](https://ai.exponencialadm.net/okf/en/radar/toolkit-de-fala-para-criar-agentes-48b376.md) - A project published on GitHub on October 1, 2024, presents a toolkit for building voice agents with open-source models.
* [AirLLM describes layer-wise inference for large models](https://ai.exponencialadm.net/okf/en/radar/airllm-executa-modelos-grandes-por-camadas-eb19ac.md) - Published on September 25, 2024, the text describes loading Transformer layers from disk to avoid keeping the entire model in GPU memory.
* [Radar: French Parler-TTS Mini model](https://ai.exponencialadm.net/okf/en/radar/french-parler-tts-mini-referencia-para-696072.md) - A September 25, 2024 post highlights the French Parler-TTS Mini model by PHBJT, hosted on a model platform. It also mentions XTTS and FishAudio and speculates about a future multilingual version.
* [Moshi: an open speech-to-speech model](https://ai.exponencialadm.net/okf/en/radar/moshi-modelo-open-source-de-speech-39aede.md) - Published on September 18, 2024, Moshi is described as a speech-to-speech model with 7.6 billion parameters, accompanied by Mimi, a streaming audio codec.
* [Open-source tools for studying network filtering](https://ai.exponencialadm.net/okf/en/radar/ferramentas-open-source-para-estudar-filtragem-fbea71.md) - A 2024 publication highlights GoodbyeDPI for Windows, ByeDPIAndroid, and zapret as open-source tools related to bypassing deep packet inspection. Engineers can evaluate them in the context of network filtering and censorship.
* [ADAS explores search for agent design](https://ai.exponencialadm.net/okf/en/radar/adas-usa-busca-para-projetar-agentes-6c0b76.md) - Published on August 26, 2024, the entry describes an iterative search approach to proposing and evaluating agentic systems. The excerpt does not report the results of the tests it mentions.
* [Five recent RAG papers in a curated selection](https://ai.exponencialadm.net/okf/en/radar/cinco-artigos-recentes-sobre-rag-em-cbeffe.md) - Published on August 26, 2024, this briefing lists five papers on retrieval-augmented generation, covering metadata and synthetic QA, modular research frameworks, graph RAG, fact-checking, and dense retrieval with weak supervision.
* [VoiceCraft: speech editing and zero-shot TTS](https://ai.exponencialadm.net/okf/en/radar/voicecraft-edicao-de-fala-e-tts-616fc0.md) - Published on August 25, 2024, VoiceCraft is described as a neural codec language model for speech editing and zero-shot text-to-speech. The publication says it can clone or edit an unseen voice using a few seconds of reference audio.
* [AutoToS refines search components with tests and BFS](https://ai.exponencialadm.net/okf/en/radar/autotos-refina-componentes-de-busca-com-6a19ba.md) - Published on August 23, 2024, the account describes iterative refinement of successor and goal-test functions, guided by unit-test feedback and breadth-first search (BFS) checks.
* [AI21 announces Jamba 1.5 Mini and Large](https://ai.exponencialadm.net/okf/en/radar/jamba-1-5-modelos-mini-e-ee6c72.md) - On August 22, 2024, AI21 Labs announced two MoE models combining Mamba and Joint Attention, with a 256K context window, ExpertInt8 quantization, JSON mode, tool use, and Transformers integration.
* [Radar: EfficientRAG for multi-hop questions](https://ai.exponencialadm.net/okf/en/radar/efficientrag-reduz-chamadas-em-perguntas-multi-1faa9d.md) - Published on August 17, 2024, the article presents a retrieval method that iteratively generates queries for multi-hop questions without requiring a language-model call at every iteration.
* [Llama 3.1 8B on an RTX 3090 under load](https://ai.exponencialadm.net/okf/en/radar/llama-3-1-8b-em-uma-9fcd60.md) - Published on August 16, 2024, the summary reports a benchmark of Llama 3.1 8B in fp16 on an RTX 3090: token-per-second performance was described as reasonable with more than 100 simultaneous requests.
* [LLM Compressor gathers compression techniques for vLLM](https://ai.exponencialadm.net/okf/en/radar/llm-compressor-reune-tecnicas-de-compressao-322419.md) - A briefing published on August 16, 2024, presents a library for applying GPTQ, SmoothQuant, and SparseGPT to models used with vLLM.
* [Parler-TTS Mini and Large v1 are released](https://ai.exponencialadm.net/okf/en/radar/parler-tts-mini-e-large-v1-9e9092.md) - Announced on August 8, 2024, the speech synthesis models were trained on 45,000 hours of speech and let users control voice characteristics with text prompts.
* [SGLang and vLLM: throughput on A6000 GPUs](https://ai.exponencialadm.net/okf/en/radar/sglang-e-vllm-throughput-em-gpus-45befa.md) - Published on August 4, 2024, the report compares two serving engines with Mistral Large 2 on four A6000 GPUs: about 1,100 tokens/s with SGLang and under 750 with vLLM; performance was similar at smaller batch sizes.
* [sqlite-vec brings vector search to SQLite](https://ai.exponencialadm.net/okf/en/radar/sqlite-vec-leva-busca-vetorial-ao-8e3784.md) - A publication dated August 2, 2024 describes a dependency-free C extension that adds vector search to SQLite and lists distance metrics, quantization, integrations, and SDKs.
* [PEER retrieves over one million experts](https://ai.exponencialadm.net/okf/en/radar/peer-recupera-mais-de-um-milhao-b121c9.md) - Published on August 1, 2024, this summary describes a Mixture of Experts architecture that uses product keys for sparse expert retrieval.
* [LoRA example for a 5-million-parameter MLP](https://ai.exponencialadm.net/okf/en/radar/lora-em-um-mlp-um-exemplo-488e6f.md) - Published on July 30, 2024, this example applies LoRA fine-tuning to a simple classification MLP and compares performance with adapters enabled and disabled.
* [A self-reasoning framework for RAG](https://ai.exponencialadm.net/okf/en/radar/rag-com-autoavaliacao-relevancia-e-evidencias-0d9645.md) - Published on July 30, 2024, the summary describes a framework that evaluates retrieved documents, selects evidence, and examines reasoning steps before answering.
* [PyTorch presents torchchat for local inference](https://ai.exponencialadm.net/okf/en/radar/torchchat-explora-inferencia-local-de-llms-30b324.md) - Published on July 30, 2024, the post introduces torchchat, a PyTorch library for running Llama 3, 3.1 and other language models on laptops, desktops and mobile devices.
* [Notebook transfers domain knowledge to Llama 3 8B](https://ai.exponencialadm.net/okf/en/radar/como-usar-conhecimento-gerado-para-ajustar-b8d5dc.md) - Published on July 29, 2024, an open-source Colab notebook fine-tunes Llama 3 8B with domain knowledge generated by Llama 3 405B.
* [Agentic components for Llama Stack APIs](https://ai.exponencialadm.net/okf/en/radar/componentes-agentic-para-apis-do-llama-ae9fde.md) - A July 24, 2024 publication presents a GitHub repository of agentic components for Llama Stack APIs and highlights using tools and functions with Llama 3.1 models.
* [Project explores a 64-bit RISC-V computer on FPGA](https://ai.exponencialadm.net/okf/en/radar/explorando-um-computador-risc-v-de-d56d47.md) - Published on July 10, 2024, the entry describes an open-source project aiming to implement a 64-bit RISC-V computer on an FPGA, able to run Linux and host its own development.
* [ESFT selects task-relevant experts in sparse LLMs](https://ai.exponencialadm.net/okf/en/radar/esft-direciona-o-ajuste-de-experts-841657.md) - Published on July 5, 2024, the paper examines parameter-efficient fine-tuning for Mixture-of-Experts models. ESFT trains selected experts based on their relevance to customization tasks.
* [XEUS: multilingual speech encoder announced](https://ai.exponencialadm.net/okf/en/radar/xeus-encoder-de-fala-multilingue-lancado-668bbb.md) - On July 1, 2024, WavLab announced XEUS, a self-supervised speech encoder trained on more than 1 million hours of audio and covering more than 4,000 languages.
* [Radar: comparing speech-to-text models](https://ai.exponencialadm.net/okf/en/radar/como-comparar-modelos-de-transcricao-de-0d1d8d.md) - Published on June 24, 2024, this item describes a comparison of speech-to-text models by word error rate, speed, and price.
* [Gradient descent as attention over training examples](https://ai.exponencialadm.net/okf/en/radar/gradiente-descendente-visto-como-atencao-aos-d078ed.md) - A February 6, 2023 article interprets gradient-trained linear layers as memories of examples and initial weights, relating predictions to a form of attention.

# Follow

* [RSS](https://ai.exponencialadm.net/en/radar/feed.xml) - Radar
* [HTML](https://ai.exponencialadm.net/en/radar/) - Radar