Monday, November 9, 2026

Pulse (last 3d) · Research 44 · Agents 44 · Products 32 · Industry 27 · Legal 18 · Releases 16 · Open Source 14 · Policy 12 · Enterprise 11 · Models 10

Trending · Anthropic 17 · Claude Code 9 · Claude 7 · OpenAI 6 · The Verge 6 · Codex 5 · FDA 4 · GitHub 4 · Nvidia 4 · Cursor 3 · Latent Space 3 · YouTube 3

---

• In its first pass, agent compaction keeps all the rules but loses about 89% of numbers and 99% of exact commands. arxiv.org
• A 284B DeepSeek model runs on 64GB of system RAM with no GPU. github.com
• Thinking Machines' Inkling 2 Small is a 594B MoE with 17B active, fine-tunable on Tinker. producthunt.com
• FDA codified AI cervical cytology imaging as Class II with special controls under 21 CFR 864.3900. federalregister.gov
• Cheap checks found plainer explanations for several headline findings inside biology models. ai.rundatarun.io
• Longevity Bench, published in Cell, scores LLMs on working through omics data rather than recalling papers. x.com
• Goldman Sachs is pulling Claude Code from its bankers and traders (single source). thestateofai.com

• GPT-6 Cyber - OpenAI released a hacking-oriented model (single source). theneurondaily.com
• Anthropic power deal - $3.4bn agreement for gas-fired power for its data centres. x.com
• Claude Personal Finance - beta that connects bank and card accounts read-only for plain-language questions. claude.com
• Nvidia investigates Jensen Huang - Nvidia is probing his ties to an alleged Chinese spy; Sonnenfeld puts his job at 50-50. theverge.com | techcrunch.com | theverge.com | techcrunch.com | garymarcus.substack.com
• Artificial Analysis Decision Index v2 - adds a Decision Score that rewards calibration; 45% of its datasets are private. x.com
• OpenAI Navier-Stokes chart - an internal model shows no diminishing returns on test-time compute (single source). x.com
• Parallel Agent Teams - test-time compute scales better across parallel agent teams (preprint, unreplicated). arxiv.org
• Majority-Vote Pseudo-Labels in Test-Time RL - measures how reliable majority-vote labels are as a training signal. arxiv.org
• Monitor-Based Interventions for Reward Hacking - design choices for monitors that curb reward hacking during RL. arxiv.org
• EmbeddingGemma 2 GB10 rerun - final test against the production vault-search config; adoption requires r@1 at or above Qwen3. huggingface.co
• Siesta - pools a team's idle Macs into one private, OpenAI-compatible Ollama endpoint. producthunt.com
• Kept - private document Q&A that runs entirely on one Mac and cites exact passages. producthunt.com
• RootMind - local agentic document search without embeddings, with OCR, page citations and MCP. producthunt.com
• Silencio 2 - open-weight, directable text-to-speech with voice cloning that runs on a laptop CPU. producthunt.com
• Apple Core AI - open-weight models can run inside Apple's app framework. x.com
• Agentic Siri - Apple is racing to make Siri act on users' behalf. theverge.com
• What local models are useful for - one 4x RTX 6000 Blackwell DeepSeek rig runs web research for frontier models. reddit.com
• Six GPUs on one x16 slot - four Radeon AI PRO R9700s and two RTX 3090s share one slot through a PCIe switch. doug.sh
• RTX 5060 Ti 16GB - buyers are buying the card in bulk for local AI. en.gamegpu.com
• CPU-ARC - caches agent reasoning so CPU inference can reuse it. arxiv.org
• AgentBaton - lets several models share agentic tasks efficiently. huggingface.co
• P^2-MOPD - parameter priors speed up on-policy distillation from multiple teachers. huggingface.co
• Qwen3.5-0.8B English-to-bash - fine-tuned 0.8B model scores 79.4% on 471 held-out requests; stock models score 55-63%. reddit.com
• Codex tool-use trajectories - GPT-5.5 trajectories released from 25 billion tokens of one person's use. huggingface.co
• Code complexity benchmark - open-source measure of how much code models and harnesses add. reddit.com
• Frontier3DBench - tests 11 coding-agent configurations building 3D objects, scenes and worlds in Blender and Unreal. reddit.com | huggingface.co
• Video-Coder - uses RL to train coding agents that generate video. huggingface.co
• Computer-Use Agents Are Embodied Generalists - position paper arguing computer-use agents should be treated as embodied generalists. arxiv.org
• Agent-Native Primitives - reusable building blocks for agents doing real-world tasks. huggingface.co
• Beneath the Surface - names what discussions of agent self-evolution miss. huggingface.co
• Agent Clan - controls Claude Code and Codex agents by hand gesture and voice. producthunt.com
• outages.live - one feed of provider status pages, available to coding agents over MCP. producthunt.com
• DocsAlot - answers customer questions in Slack and Discord from an existing knowledge base. producthunt.com
• Hide slop PR descriptions - put verbose AI-written PR text in collapsible details blocks. simonwillison.net
• parse_number() as a SQLite extension - compares C and Rust implementations. simonwillison.net
• Stanford AI-native engineering - Mihail Eric on how Stanford's course teaches it. latent.space | latent.space
• Torchlens - inspect, visualize and intervene on any operation in a PyTorch network. github.com
• Generative Latent Priors for Protein LMs - latent priors used to interpret and steer protein language models. arxiv.org
• Agentic radiology - tests whether agents built on production AI tools close radiology's performance gap. arxiv.org
• Formally verified quantum programs - AI finds quantum programs that come with formal correctness guarantees. arxiv.org
• Machine learning for lattice QCD - review of ML methods for lattice QCD. arxiv.org
• LLM temporal self-knowledge - tests whether models know their knowledge cutoff, to prevent look-ahead leakage in backtests. huggingface.co
• Weight-Space Programming - capabilities and limits of programming directly in model weights. huggingface.co
• Maya a1 order sensitivity - tests whether option position drives a non-autoregressive decision model. arxiv.org
• Ensemble confidence for code summaries - estimates confidence in LLM code summaries without reference answers. arxiv.org
• RAG and algorithmic problem-solving - measures how retrieval affects LLM performance on algorithmic problems. arxiv.org
• Lexical conflicts in translation - how LLMs handle programming jargon that conflicts with everyday meaning. arxiv.org
• PRISM name collisions - the same name used across fields degrades scientific paper retrieval. huggingface.co
• Next-Concept Modeling - generates text in parallel, coarse to fine (preprint, unreplicated). arxiv.org
• GAP gated block - one gated block replaces attention plus MLP, with about 20% fewer parameters on WikiText-103 (single source). reddit.com
• Confucius4-R2T2 - streaming speech recognition that emits stable prefixes in real time. huggingface.co
• OneTrans-GSR - one autoregressive transformer handles both reranking and mix-ranking. huggingface.co
• Graph similarity with GNNs - scales graph similarity learning with graph neural networks. arxiv.org
• ReLU network closedness - theory on whether the set of ReLU network functions is closed. arxiv.org
• Frameboy by World Labs - browser demo streams views from the Atlas world model at up to 15 fps. producthunt.com
• Pokémon Red world model - playable diffusion world model with a System 1/2 agent harness. reddit.com
• EvoWorld - video models that simulate how parts of a scene out of view change. huggingface.co
• OmniLatent - one latent representation for modeling 3D parts. huggingface.co
• SplatAnything - normalizes Gaussian spacing so feed-forward 3DGS works at any scene scale. huggingface.co
• Declarative Dragging - specifies drag-and-drop interactions declaratively. declarative-dragging.github.io
• Suno lawsuit - rights holders for Elvis and Eminem are suing over AI music clones. theverge.com

Separate results put 284B to 2.8T parameter models on single machines, and new releases keep active parameters small.

A 284B DeepSeek runs on 64GB of RAM, and a 2.8T Kimi fits on one Blackwell Llama-modchip serves a 284B DeepSeek model from 64GB of system RAM with no GPU. A compressed Kimi-K3.1-Pro-Coder shrinks from 1.77 TB to 775 GB, runs at 10.8 TPS on one Blackwell and reports 96.7% on AIME 2025 and LiveCodeBench (single source). Running frontier-class open weights inside a clinical firewall is now a hardware purchase, not a cluster build. github.com | i.redd.it

FP4 KV cache and reasoning-model quantization target the memory that remains BitDecoding-FP4 decodes directly from an FP4 KV cache by co-designing the algorithm and the system. CodeQuant2 adds outlier mitigation and adaptive clustering for low-precision reasoning models (preprints, unreplicated). With weights already shrunk, the long-context cache, which dominates memory on long patient records, is the next budget to cut. arxiv.org | arxiv.org

Thinking Machines ships a 594B MoE with 17B active, fine-tunable on Tinker Inkling 2 Small is built for long-horizon coding, visual tasks and agents, and Tinker offers it for both inference and fine-tuning. With 17B active parameters, serving costs land near a mid-size dense model, so a domain fine-tune on proprietary clinical text no longer means trading scale against serving cost. producthunt.com

Compaction, stale instruction files and unread errors share one failure: agents keep the gist and lose the exact details.

Compaction loses 89% of numbers and 99% of exact commands in its first pass Twelve real OpenClaw sessions were run through the product's own compaction mechanism. Rules survived all five rounds intact, while about 11% of numbers, 1 to 4% of audit pointers and about 1% of exact commands survived the first. The loss happens in round one rather than building up over rounds, so lab values and provenance pointers have to live outside the compacted context. arxiv.org

Stale instruction files and decaying memory carry measurable costs One paper puts a number on the hidden cost when coding agents follow outdated instruction files. Another compares a two-tier memory with compaction, decay and reinforcement directly against RAG (preprints, unreplicated). Both treat agent context as an artifact that drifts and needs maintenance, which is how auditors will treat the prompts and runbooks of a validated pipeline. arxiv.org | arxiv.org

LogBear stops agents from declaring "done" while errors sit unread LogBear is a free, MIT-licensed macOS wrapper that records every dev-server run locally. When an agent tries to finish while errors are still unread, LogBear blocks it and hands over the real error. It is a deterministic fix for the exact gap that causal agent evaluation frames as knowing when to intervene. producthunt.com | arxiv.org

Interpretability audits, an omics benchmark, usage data and a new FDA device class all test whether models do actual biological work.

Cheap checks explain away headline findings inside biology models A review of interpretability studies on biology foundation models finds that several headline numbers had simpler explanations, and the checks that found them were cheap. Any claim that a model has learned biology now needs baseline and confound controls before it reaches a scientific advisory board. ai.rundatarun.io

Longevity Bench in Cell tests omics analysis instead of literature recall The benchmark scores frontier LLMs on navigating omics data rather than retrieving published results (single source). The task matches real precision-medicine analysis, where literature-recall tests flatter models that cannot handle a raw expression matrix. x.com

Anthropic's 207,049 science conversations put biology near the top and lab work as the bottleneck Anthropic's study of how researchers use Claude finds biology is a major use and that physical experimentation, not analysis, is the main constraint. Faster analysis now just lengthens the wet-lab queue, so the return moves to tools for experiment design and prioritization. anthropic.com | x.com

FDA codifies AI cervical cytology as Class II with special controls The November 5 rule creates 21 CFR 864.3900 for digital cervical cytology imaging systems, codifying Hologic's Genius De Novo 33 months after it was granted. Later AI pathology devices now have a named classification and set of special controls to cite in a 510(k), instead of filing their own De Novo. federalregister.gov

Banks and big tech now set access limits and spending caps on agent tools by role.

Goldman pulls Claude Code from bankers and traders Goldman Sachs is removing Claude Code from front-office staff (single source), while one engineer's account of moving to the Pi harness shows developers shopping for alternatives. Regulated firms are drawing the line by role rather than by tool, the same approach compliance teams will take for users near PHI. thestateofai.com | caleb.software

Apple caps engineers at $200 a day in AI spend; Accenture caps staff at $29 a month Apple has set a $200 daily AI limit per engineer and may cut it to $100 (single source). At $200 a day, agent spend runs to roughly $50,000 a year per engineer, two orders of magnitude above Accenture's cap, so it is now budgeted per seat like cloud compute. stork.ai

15 items: opinion essays, open questions, tutorials, off-topic hardware and email news, or posts with no checkable claim.