Friday, October 16, 2026

Pulse (last 3d) · Research 144 · Agents 94 · Products 72 · Industry 48 · Models 46 · Open Source 40 · Releases 30 · Media 24 · Enterprise 23 · Legal 22

Trending · OpenAI 16 · Claude 13 · Claude Code 12 · Google 12 · Anthropic 11 · Cursor 10 · FDA 10 · Codex 9 · ChatGPT 8 · Gemini 4 Argon 8 · Gemini 7 · GitHub 7

---

• Meta's Muse Spark Open is an 816B MoE with about 42B active parameters, co-optimized with Apple for MLX on Mac Studios. theverge.com
• Clinical LLM benchmarks cover 31% of real clinician use, and diagnosis is just 3.7% of queries. arxiv.org
• The best model scores 55.2% at acting on memory that a later observation silently invalidated. arxiv.org
• NVIDIA released Nemotron 3.5 Super VL 120B-A12B as open weights, including NVFP4 checkpoints. reddit.com
• OpenAI is opening GPT-Rosalind, its biology model, to academic research labs. openai.com
• Accenture cut employee Claude spend to $29 a month, down from as much as $1,000. 404media.co
• OpenAI is racing to fix safety issues before shipping its next flagship (single source). x.com

• GPT-Rosalind academic plan - OpenAI extends its biology model to academic research labs. openai.com
• OpenAI flagship safety scramble - Next flagship held for safety fixes amid fallout from three safety researchers' firings (single source). x.com | x.com
• Eight states sue OpenAI - The states allege that ChatGPT harmed users. nbcnews.com
• An update on independent evaluators - Anthropic updates its external evaluation arrangements. anthropic.com
• Mythos at Bank of America - The bank reportedly hired 1,000 contractors to fix flaws Mythos found (single source). i.redd.it
• Private RAG via random projection - Proposes random projection as a privacy mechanism for retrieval-augmented generation. arxiv.org
• AI Consilium - Benchmarks LLMs on clinical multidisciplinary second opinions. arxiv.org
• MindEval2 - Mental-health support eval with clinically grounded simulated patients and rubrics. arxiv.org
• Twice Upon a Time - Complementary representations improve LLM reasoning over time series. arxiv.org
• Evaluation awareness beyond English - Tests whether models behave differently when evaluated in Indic languages. arxiv.org
• Neural networks are clean - Goodfire argues networks decompose into clean, understandable structure. goodfire.com
• Mistral Large 4 preview - 1T parameters, 49B active, multimodal, weights promised by month end (single source). x.com
• Gemini 4 Argon agents beat Minecraft - Four agents sharing chat and memory defeat the Ender Dragon. arxiv.org
• AI search engines cite separate webs - Fudan study finds Qwen cited 96% Chinese sources even when half the questions were in English. reddit.com
• LLM-discovered quantum circuits - Lukin's Harvard group finds circuits that match or beat the state of the art. arxiv.org
• Self-assembling learning machinery - PNAS: random units with shared plasticity rules form learning circuits that perform comparably to backprop-trained networks. reddit.com
• Knowledge linear before fine-tuning - Knowledge that is represented linearly before fine-tuning is what generalizes after it. arxiv.org
• MLExpert - A generalist agentic model for autonomous ML engineering. huggingface.co | arxiv.org
• Reward-specific token credit assignment - Assigns each reward in multi-reward RL only to its relevant tokens. huggingface.co
• Waypoint-OPD - Transfers RL-induced policy shifts through gated on-policy distillation. huggingface.co
• MMSearch-OPD - On-policy distillation for long-horizon multimodal search agents. huggingface.co
• KnowGPO - Knowledge-grouped policy optimization for search agents. arxiv.org
• Cowork-Gym - Synthesizes interactive work environments at scale for agent training. huggingface.co
• Recursive Environment Scaling - Recursively scaled training environments for long-horizon agents. huggingface.co
• Mosaic - Multi-session agentic reinforcement learning. huggingface.co
• Self-orchestration in agent swarms - Decentralized LLM agent swarms learn to orchestrate themselves. huggingface.co | arxiv.org
• Learning to steer agents - Learned steering for LLM agents and multi-agent systems. arxiv.org
• FDE-Gym - An environment for last-mile, forward-deployed engineering work. arxiv.org
• MoE routing exploration - Revisits how routing exploration works in modern MoE training. arxiv.org
• Transformers for continuous diffusion LMs - Architecture design for continuous diffusion language models. arxiv.org
• Lean Graph Search - Graph-based search for formal theorem proving in Lean. arxiv.org
• Diverse scientific ideas - Trains reasoning models to generate more diverse scientific ideas. arxiv.org
• Linear recurrence vs attention at 15M - Hobbyist finds equal accuracy but different latency and training cost. reddit.com
• OpenMuse - An open-source, self-hostable personal agent with its own browser, terminal and files. producthunt.com
• Dreamscale Labs - YC F26 low-latency cloud inference for VLAs and world-action models. producthunt.com
• ChatGPT reads Chrome tabs - ChatGPT can now pull context from open Chrome tabs. theverge.com
• Windows for RTX Spark - Microsoft tunes Windows for Nvidia's RTX Spark laptops. theverge.com
• AI hardware reverse engineering - A company owner used AI to decompile a competitor's audio DSP firmware. blogs.gnome.org
• Sullivan & Cromwell - The law firm uses OpenAI models for agreement review. openai.com
• Krea 3 - Open-beta image model built for fast, cheap, precise regional editing. krea.ai
• Bluesky's AI tool - Users can now build shareable games and apps. theverge.com
• VLA-Δ - Trains vision-language-action models to learn physical state transitions. huggingface.co
• PINT - A unified robot manipulation model built on shared predictive intent. huggingface.co
• Dexterity from Supervised Play - Learns dexterous manipulation from supervised play data. arxiv.org
• T2-z - Test-time optimization of environment latents for robotic world models. arxiv.org
• CoDA - Co-evolves data engines and robot action policies. arxiv.org
• LIBERO-PeRM - A benchmark for personalized robotic manipulation. arxiv.org
• GeoState - View-decoupled geometry states for 3D world modeling. arxiv.org
• WorldCast - Applies visual generation models to world-state prediction. huggingface.co
• Grid-Bench - Tests whether LLMs can predict state changes in discrete grid worlds. arxiv.org
• OmniVerifier-C1 - A decomposed, graded verifier for rule-governed generative world modeling. arxiv.org
• 3DCodeVerse - Represents and generates 3D worlds as code. arxiv.org
• MJev-Spatial - Builds spatial intelligence through native attention matching. huggingface.co
• ACEdit - Asymmetric conditioning keeps multi-turn image editing stable. arxiv.org
• H-DDCM - Hadamard codebooks speed up generative image and video compression. huggingface.co
• AVC-Reward - Event-centric evaluation of audio-video consistency. arxiv.org
• Zero-bit audio watermarking - Studies backward compatibility in neural audio watermarks. arxiv.org
• PathFM - Fast generative forecasting of multivariate trajectories. arxiv.org
• Timing-aware power optimization - AI-driven power optimization for chip design. huggingface.co

Every item here moves a capable model closer to hardware that a regulated organization can buy, own and lock down.

Meta's 816B Muse Spark Open runs on Mac Studios Meta released Muse Spark Open, an 816B MoE with about 42B active parameters that it bills as the most capable US open-weight model, and worked with Apple to optimize it for MLX on Mac Studio setups. A US-origin frontier open model removes the provenance objection that has kept Chinese weights out of regulated on-prem stacks. reddit.com | theverge.com | x.com | x.com

NVIDIA ships a 120B multimodal MoE with FP4 weights Nemotron 3.5 Super VL 120B-A12B ships as base, BF16 and NVFP4 checkpoints plus a LiLiCorr speculative-decoding variant, alongside two smaller Bolt edge models. With 12B active parameters and native FP4, an open vision-language model becomes a single-node deployment on Blackwell. reddit.com | x.com

Ternary weights and diffusion decoding cut serving cost again Ternary Bonsai 2 (2B, 4B and 9B, built on Qwen 3.5) runs its 2B model entirely in the browser at up to 500 tok/s on an M4 Max, though users report looping. ALoDLM converts Qwen3-8B into a looped diffusion LM that is about 3x faster with better benchmark scores (preprint, unreplicated). reddit.com | reddit.com | reddit.com

Commodity accelerators now serve large MoEs at usable speed Four Radeon AI PRO R9700 cards run DeepSeek-V4-Flash-0731 at about 79 tok/s, up from 36 on two cards. A £300, 35W Snapdragon mini PC runs a 35B coder at 30 tok/s and an 80B model that exceeds its RAM at 17 tok/s, which shows that active parameters, not total size, now set the hardware bill. reddit.com | reddit.com

Each item measures a gap between what benchmarks reward and what deployed systems actually face.

Clinical benchmarks cover 31% of what clinicians actually ask An 8-month audit of 127,833 queries from 6,342 clinicians found that real use is mostly documentation and administration (36.2%) and knowledge retrieval (28.9%), with diagnosis at 3.7% (preprint, unreplicated). Public benchmarks overlap with that use by a median of 31%, so diagnostic leaderboards say little about deployed value and deployment query logs make the better eval set. arxiv.org

Agents retrieve the corrected fact and act on the stale one HKUST's STALE benchmark finds the best frontier model scores 55.2% when a later observation silently invalidates an earlier memory, and dedicated memory frameworks also fail (preprint, unreplicated). Longitudinal patient records are full of exactly these implicit supersessions: revised diagnoses, discontinued drugs and amended results. arxiv.org

Legal RL triples a score that still fails four tasks in five Z.ai's GLM-Legal adds legal mid-training and RL over tens of thousands of task environments to GLM-5.3-Flash, lifting its pass rate on Harvey's Legal Agent Benchmark from 6.67% to 19.17% per Vals AI. The vertical-agent recipe is now public, along with a measure of how far domain RL still leaves a model from usable; API access is an application-only beta. producthunt.com

New agent tooling ships abstention and human approval as core features, and this week's incidents show the cost of running without them.

Approval gates and disposable machines become products withHuman intercepts tool calls from Claude Code, Codex, LangGraph and Pydantic AI agents, approves routine calls by rule, and routes risky ones to a named reviewer over an open protocol (AAP). Buddy Sandboxes give each task, branch or PR its own isolated Linux machine with root, Docker and a persistent disk, which keeps agent writes off shared systems. producthunt.com | producthunt.com

Retrieval tools now return "unresolved" instead of guessing resolve() maps each entity mention to a stable ID from Wikidata or a private vocabulary and returns "unresolved" with a reason when evidence is thin, under Apache-2.0. SuperLocalMemory relaunched as local-first agent memory whose recall answers supported, related or no answer. Explicit abstention is the property that ontology-linking in clinical pipelines has lacked. producthunt.com | producthunt.com

Broad write access keeps producing the incidents An unsupervised agent with production-server access deleted a TV show's finished season weeks before release, and a backup failure left it unrecoverable (single source). AI coding agents separately exposed about 13,000 internal images, including billing records, on GitHub. tomshardware.com | thehackernews.com

Buyers are capping budgets and routing work to cheaper models while vendors hold or cut prices, so revenue now trails usage.

Accenture caps Claude at $29 a month per employee Accenture cut many employees' Claude allowance to $29 a month, down from as much as $1,000 for some. When the largest AI integrator rations seats, per-user frontier access becomes a managed cost line rather than a productivity bet. 404media.co | linkedin.com

Grok Voice 3.0 tops speech-to-speech at the old price xAI's Grok Voice 3.0 ranks first on the Artificial Analysis Speech-to-Speech Index at 86.9, with sub-second latency (vendor claim), and stays at $0.08 per minute. Agents on grok-voice-latest were upgraded silently, which breaks any validated workflow that does not pin its model version. producthunt.com

Routing caps lab revenue even as usage climbs Google's Gemini 4 Argon launched September 30 at half price until December 1 (single source). One analysis argues that frontier revenue growth will slow because developers route easy tasks to cheaper models, and the Guardian reports that some companies that rushed into AI are now pulling back. x.com | reddit.com | theguardian.com

12 items: roundups, teasers, opinion pieces and discussion prompts with no new factual claim.