DeepMind reports weaknesses in legacy RSA

Meta announces specialized models trained using Muse Wonder for Muse and Meta AI, mentioning Active Memory. The available post cuts off during its remaining feature list, so a complete capability inventory cannot be inferred.

MiniMax says M3.1 Flash is available through Fireworks AI. The two companies also plan a San Francisco event about how it was built; this announcement contains no new comparative performance figures.

Mistral highlights Voxtral 3’s end-of-turn detection, which helps developers identify when someone has finished speaking and build more responsive voice interactions.

Tencent Hunyuan’s account reports first place among open-weight models on APEX-Agents. The short post does not supply a full score breakdown or evaluation configuration for comparison.

Clement Delangue shares a developer’s announcement of Qwen Concept 27B, built with direct weight-space edits inspired by recent research. This is a community project; its name does not establish an official Qwen release.

Soumith Chintala shares DesignArena’s report that Inkling 2 Small ranks fourth overall and first among open-weight models. The positions refer to that leaderboard at the time of the announcement.

Soumith Chintala says the team spent substantial effort studying customization and designed Inkling 2 Small to be competitive under fixed data and budget constraints. This describes the team’s training and customization priorities.

Thinking Machines describes proactive safety monitoring during training, internal and external testing, and work to understand attempts to evade safeguards and build appropriate protections.

Soumith Chintala shares an infrastructure-team account that Inkling 2 Small took slightly under three days to pretrain. That figure covers pretraining, rather than the entire research and post-training process.

Shunyu Yao shares OpenDesign results placing Hy4 0.1 points behind GPT-6.1 Sol at 41% lower estimated cost. Both comparisons belong to that design benchmark and should not be generalized across tasks.

A researcher introduces the open-source Drex family and technical report, reporting that Drex Max ties GPT-6 Sol on Decision Bench v4 and scores strongly on DecideBench. The available post does not provide the full evaluation configuration.

Anthropic’s science blog discusses AI’s changing role in explaining and predicting real-world systems, including the firing of individual nerve cells. The available announcement introduces the research direction without complete experimental results.

Cohere Labs reports testing 28 AI models with human-written cultural riddles and finding substantial gaps in cultural understanding. The available repost does not show every model’s scores.

The paper combines low-rank gating projections, state-row reweighting and matched-filter readout, reallocating saved parameters to wider feed-forward networks. Its authors report perplexity improvements over GDN-2 and Transformers at 166M–1.34B parameters.

Confucius4-R2T2 post-trains Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio. A longest-stable-prefix mechanism avoids rewriting earlier output; the paper reports low latency with strong recognition quality and releases weights and code.

The training framework keeps early denoising close to the base model and learns fidelity corrections at the trajectory’s tail without human-labeled editing pairs. The paper reports lower accumulated drift over fifteen to twenty edits with FLUX, SenseNova and Qwen models.

Reactive Optimizers introduces OPAL, which inspects a training run and rewrites executable optimizer code, including parameter-group rules, schedules and precision. The authors report matching or exceeding fixed recipes on selected untuned tasks and discuss additional compute costs.

Omar Khattab shares a technical account of a large model continually guiding reinforcement learning for a 0.4B-parameter retriever. The repost cuts off before the comparison target, so it does not support a sweeping claim of outperforming all larger models.

Arohan cites ByteDance experiments suggesting a Full Bandwidth Transformer variant may approach a model twice its size. This concerns a particular experiment, rather than demonstrating that every configuration halves compute needs.

A research post highlights MLIP-liquid-corrections on Hugging Face, providing BSSE and ATM correction labels. The truncated announcement does not establish the dataset’s scale or achievable accuracy.

LeRobot announces v0.7.0 and mentions workflows ranging from terabyte-scale data to prompting robots. The available repost does not include a full changelog.

A researcher introduces Möbius, a unified motion-language model connecting human-motion understanding and generation. The available repost does not provide complete task coverage or benchmark results.

Linus Ekenstam reports that Liber-0 faced 60 tasks across 600 trials without training on demonstrations collected specifically for those tasks. Reported success was 36% overall, 63.5% for pick-and-place and 34% for other manipulation; these remain attributed experimental figures.

A creator shares using RGBMocap to extract motion from video, mentioning a local MiniMax H3 workflow and a character-swap LoRA. This is an individual workflow example, rather than a guarantee across inputs.

OpenAI, Frontier Model Forum and member firms propose identifying when bypassing AI safeguards could enable serious cyber harm, then prioritizing stronger protections for those cases.

SGLang reports over 420 tokens per second per user at low concurrency and over 31,000 per second per GPU at high concurrency on GB300 NVL72. These figures describe different workloads, rather than a single operating point.

Woosuk Kwon shares Inferact’s statement that its platform serves more than seven trillion tokens per day. This is platform-disclosed traffic, rather than an independently audited total.

Inferact announces work with Google to provide day-zero Kimi K3.1 Pro support in vLLM on TPU v7x. The available repost does not supply throughput or cost comparisons.

CoffeeVectors shares LiquidEnergy’s description of NanoDC as a solar-powered autonomous data center for AI inference. The available post lacks verifiable capacity, cost and sustained-operation figures.

Dotey cites Ramp figures showing enterprise token spending rising 42% month over month in July, 1% in August and 0.2% in September. These concern spending in the stated sample, rather than token volume; the underlying statistics and valuation rumors were not independently verified from this item.

Jeff Dean announces work with Laude Institute and fellow computer scientists to fund exploratory open-science projects. The post does not state an aggregate funding amount.

The project describes Longevity Grok as a personal longevity research team running on Grok Bot, with health records, research briefs and appointment preparation. Clinic and laboratory editions require early-access applications.

Runway co-founder @c_valenzuelab observes that interfaces are moving from places where people directly make changes toward places where they supervise agents and interpret the resulting work.

Perplexity’s CEO expects agents to invoke other agents increasingly often. He argues that version-controlled code and execution traces can support collaboration around a compounding git record. This is his architectural perspective.

Arav says median weekly agent-hours received by employees have reached seven times September’s level, with one in five employees receiving more than forty hours weekly. Agent runtime does not directly measure human work saved or output gained.

Vercel’s CEO agrees that coding is solved while bugs remain, arguing that teams can spend tokens on endless refinement, including changes nobody needs. His point concerns judgment about value and problems.

Teortaxes argues that using stronger models to organize more Markdown files can work well while remaining limited by pretraining. This is a critique of agent memory, rather than an established capability ceiling.

Teortaxes cites Vera Rubin versus GB300 throughput ratios of 5.0x at 260 tokens per second per agent and 1.7x at 80, discussing interaction speed versus maximum throughput. Full measurement configurations are absent from the post, so the cited figures remain unverified here.

Dotey describes requiring screenshots of changed features with pull requests to encourage actual execution before review. He is applying the same practice to AI agents; this is a team-workflow observation.

Ethan Mollick argues that a potential pullback in AI stocks would not automatically change enterprise adoption or technical progress. This is his view of industry development, rather than evidence for a market forecast.

A creator lists Grok Bot assistants for an animated series, books, Japanese study and work. The example illustrates personal task decomposition without an independent efficiency evaluation.

Dotey shares an engineering view that spec and task files can remain internal to a Loop Agent while people provide discrete, rough inputs. This is a design position about dividing work between people and agents.

Clement Delangue argues that open source lets students, researchers and startups build on years of progress and contribute back, rather than concentrating robotics capability inside a few black-box systems.

Dotey notes that building a DeepSeek Harness plugin also required implementing plugin updates himself. It is a practical observation about maintaining developer tools.

Latent Space previews a cooking-show-format episode about Reducto AI’s PDF processing workflow. The fetched page contains only a channel description, so unseen video details are not reconstructed here.

CoffeeVectors shares Variety’s reporting on a Weta FX executive describing changing attitudes toward generative AI and the need for an AI strategy. This remains an attributed perspective, rather than proof that an entire industry has completed its transition.