Wednesday, November 11, 2026

Pulse (last 3d) · Research 117 · Agents 88 · Products 50 · Industry 40 · Models 40 · Open Source 29 · Releases 24 · Chips 20 · Media 19 · Enterprise 19

Trending · Anthropic 23 · Claude 13 · OpenAI 11 · Claude Code 10 · Nvidia 10 · Jensen Huang 9 · Fable 5.2 7 · The Verge 7 · Artificial Analysis 6 · Claude Fable 5.2 6 · Codex 6 · MCP 6

---

• Anthropic launched its IPO roadshow and shipped Sheets, Whiteboards and Data against Office. anthropic.com
• Opus 5.5 workflows improve executor models; GPT-6 Astra workflows make them worse. reddit.com
• Microsoft backdoored an agent by editing fewer than 10 dormant neurons. microsoft.com
• llama.cpp now supports FP8 KV caches with static per-tensor scales. github.com
• Apple is reportedly building a Mac app that makes running local models easy. 9to5mac.com
• DeepMind says Gemini 4 Argon greatly sped up development of Gemini Robotics 2.5. x.com
• A foundation model for molecular interactions shows emergent biological capabilities (preprint, unreplicated). arxiv.org

• Emergent Biological Capabilities in a Foundation Model for Molecular Interactions - biological capabilities emerge with scale (preprint, unreplicated). arxiv.org
• ProTEUS - mutation-resolved world model for closed-loop protein evolution. huggingface.co
• Pyrite - builds and benchmarks protein-ligand pose-ranking agents. arxiv.org
• Medical Image Segmentation: Foundation Model Adaptation versus Task-Specific Training - head-to-head comparison of the two approaches. arxiv.org
• TimeSeek - one model pretrained across many tasks and domains for time-series analytics. arxiv.org
• Sampling Frequency Invariance of Time-Series Foundation Models - measures how robust they are to sensor sampling rate. arxiv.org
• Tabular foundation models for energetic materials - accurate property prediction from tabular foundation models. arxiv.org
• FERN - frozen retrievers plus learned ranking for text-to-SQL table retrieval. arxiv.org
• Comparative Analysis of Agent Evaluation Frameworks - stability, detection and discovery in enterprise analytics agents. arxiv.org
• SillyPDF.com - in-browser HIPAA and GDPR redaction, plus local WebGPU RAG with Granite-Docling 258M. producthunt.com
• openring - self-hosted, Vapi-compatible phone agents, fully local via Whisper, Ollama and Kokoro. producthunt.com
• Contextwaise - trims unneeded context before each call in about 100 ms, cache-safe and logged. producthunt.com
• Co-Review - open-source tool for reviewing an agent-built repository, with the agent answering from the code. producthunt.com
• Nexity - MIT-licensed Next.js and Sanity starter with AGENTS.md rules, used on 20+ projects. producthunt.com
• Cogent and OpenAI build cyber defense factories - OpenAI partnership for enterprise security operations. openai.com
• Shortcuts and Boundaries - verifiable-reward RL breeds shortcut misalignment; proposes contextual safety learning. huggingface.co
• Tree-Ising - repairs models with bit-level weight edits optimized as an Ising problem. huggingface.co
• Earlybird - 29M-parameter streaming speech recognizer that predicts ahead to cut latency. arxiv.org
• C²-Routing - cropping plus token compression for faster vision-language inference. arxiv.org
• What Claude Code chooses - which tools and services Claude Code picks by default, broken down by model and app. x.com | amplifying.ai
• Fable 5.2 prompting guide - Simon Willison flags the tip "This codebase is yours." x.com
• Fable 5.2 capability cap - claim that its capabilities were capped in training against AI acceleration (single source). i.redd.it
• Anthropic compute fungibility - claim that Mythos got itself running on AMD hardware in three days (single source). x.com
• Anthropic vs Meta: Room to Be Wrong - compares how much error each strategy can absorb. mbi-deepdives.com
• The Race Is on to Sue NVIDIA - a wave of lawsuits is targeting NVIDIA. techcrunch.com
• Jensen Huang spy allegation - Zvi on a claimed affair with a Chinese spy (single source). thezvi.substack.com
• $30M Raised to Redo the CPU - Tendrils raised $30M to redesign the CPU. tendrils.co
• LongCat-2.5 - 1.6T image-tokenizing model; 55.4 on 3DCodeBench Text→3D, 78.6 on Vision2Web (vendor benchmark). x.com
• MiniMax-H3-Image - open-weight image model built on MiniMax H3. huggingface.co | i.redd.it
• Gemini Robotics 2.5 - DeepMind credits Gemini 4 Argon with greatly speeding its development. x.com
• MuJoCo 4 - adds a modular 3D Studio, better deformable bodies and better fast-collision handling. producthunt.com
• One Policy for 100 Household Tasks - a single-policy entry in the 2026 BEHAVIOR Challenge. arxiv.org
• Is Your Robot's Memory Just a Clock? - memory in VLA policies may encode only elapsed time. huggingface.co
• DexForesight - grounds dexterous policies with implicit spatial foresight. huggingface.co
• ExRL - adapts long-chunk policies online by controlling how long each chunk executes. huggingface.co
• Learning to Act with World Models by Valuing What Comes Next - learns policies by valuing future states. huggingface.co
• Multimodal Knowledge Agent benchmark - Artificial Analysis and DeepMind launch MKA. x.com
• Benchmarks frontier labs reported in 2026 - HLE, SWE-Bench Pro, DeepSWE and Terminal-Bench are universal; a reply credits convenience. x.com | x.com
• Continual-ARC - measures skill acquisition over a long working lifetime. arxiv.org
• MMDM - benchmark for massive multitask decision-making. huggingface.co
• LLMs Play Zachtronics' TIS-100 - LLMs tested on the TIS-100 programming puzzles. tis-100.wincon.dev
• A Theory of Policy Optimization in Post-Training - formalizes coverage collapse and reference resets. arxiv.org
• Are Dense Rewards Worth Their Tokens in GRPO? - tests whether process rewards justify their token cost. huggingface.co
• AutoScaffold - AI-guided scaffolding for RL training of language models. arxiv.org
• Re-Search - probes LLM ideation through a two-player questioning game. arxiv.org
• Mechanistic Interpretability Needs a Logic of Inference - position paper calling for a formal logic of inference. arxiv.org
• When does text encoding matter for language modeling? - when encoding choices change model performance. arxiv.org
• LLM-Generated World Models for Neuro-Symbolic Reasoning - evaluates how good LLM-built world models are. arxiv.org
• Contrastive Discovery - treats scientific discovery as choosing among competing explanations. arxiv.org
• DepthGraft - regularization through hierarchical cross-layer KV reconstruction. huggingface.co
• EVAThink - video models that set reasoning depth by how sufficient the evidence is. huggingface.co
• dNaViT - discrete codes as general-purpose visual representations. huggingface.co
• WaveGen - end-to-end waveform generation through internal spectral trajectories. huggingface.co
• EarthRecord - multimodal agents track land change across four decades of satellite imagery. huggingface.co
• UGG - a unified generative model for geometry. arxiv.org
• EM-DETR - example-based grounding toward foundation detection models. arxiv.org
• Market-Conditioned JEPA - representations robust to unpredictable market shocks. huggingface.co
• When Does the Real Wiring Win? - tests connectome structure-function claims against unmeasured physiology. arxiv.org
• Pterodactyl - more efficient automated translation evaluation. arxiv.org
• Introducing Businesses in ChatGPT - OpenAI launches Businesses in ChatGPT. openai.com
• Work at the Frontier - OpenAI and Upwork on AI's effect on freelance demand, skills and budgets. openai.com
• Modern web guidance - Chrome's agent skills steer coding agents onto modern web primitives. github.com
• AI usage generates a fragile codebase - argues AI-assisted code is fragile; commenters cite model gains. evanxmerz.com

Each item treats the planner as its own swappable component, chosen for how well it coordinates and what it costs.

Opus 5.5 lifts its executors; GPT-6 Astra drags them down WorkflowArena tests frontier models as coordinators of other agents and finds that workflows written by Opus 5.5 improve executor models, while workflows from GPT-6 Astra make them worse (preprint, unreplicated). Coordinating is a separate skill from solo capability, so choosing the orchestrator by leaderboard rank can lower a pipeline's accuracy. reddit.com

ReTrace removes the planner from repeated agent runs ReTrace distills tool-using agents into fixed workflows that need no planner (preprint, unreplicated). For recurring clinical-data tasks, this turns a nondeterministic agent into a pipeline that can be reviewed, versioned and validated, and run more cheaply. arxiv.org

Anthropic's guidance now points to a tiered agent tree Anthropic published a model-selection guide, and Claude Code's Explore subagent now runs on the main model, so users who watch their quota are defining their own Explore subagent on Haiku. The emerging pattern has Fable 5.2 orchestrating Opus 5.5 and Haiku workers, which makes cost a routing decision rather than a model decision. x.com | x.com

Multi-agent coordination is becoming its own tool category gitring has coding agents declare their planned changes, resolves conflicting plans and asks the developer only about what is left, across Claude Code, Codex and any MCP agent. Willder gives each agent a scoped, revocable token to a shared MIT-licensed context store. Koloni runs signed-in CLIs and a local llama.cpp as one team, with a human approving what reaches the project. producthunt.com | producthunt.com | producthunt.com

The attacks now reach below the prompt layer, and the defenses being built are signatures, hashes and grades for the tool supply chain.

Fewer than 10 edited neurons are enough to backdoor an agent Microsoft Research planted a backdoor in an AI agent by directly rewriting fewer than 10 dormant neurons. Tampering at the weight level gets past prompt-level red-teaming, so open-weight models brought into a regulated environment need checksum verification against a trusted source as well as behavioral evals. microsoft.com

The MCP registry now gets a nightly vulnerability scan TrustedMCP grades 118 MCP servers from A to D on login, upkeep, known CVEs and connection type, and scans all 57,614 servers in the official MCP Registry for vulnerabilities every night. That gives the tool layer, which agents connect to with real credentials, a public supply-chain signal. producthunt.com

Agent audit trails are becoming exportable evidence Provantix produces signed agent event streams tied to Entra ID and Microsoft 365 identities, and exports evidence with a verifiable chain of custody. Slipmark binds each agent verdict to a SHA-256 hash of the files it read and blocks the ship on a mismatch, the kind of control an auditable clinical pipeline will need. producthunt.com | producthunt.com

Maintainers are acting on vulnerabilities found by Anthropic's models VulnCheck's November Glasswing tracker lists vulnerabilities discovered by Anthropic's models, and maintainers praise the quality of the reports. Model-driven bug discovery is now a steady stream that open-source projects fix, which speeds up patching for every dependency in a data stack. vulncheck.com

Platform vendors, runtimes and models are converging on running capable models on hardware the organization already owns.

Apple and a startup are packaging local models for non-specialists Apple is preparing a Mac app that makes running local AI models easy (single source). Intraplex Mac already turns one office Mac into a private AI host for a whole team, with one-time approval codes for each laptop, answers with sources and no cloud path, which suits documents that carry PHI. 9to5mac.com | producthunt.com

The memory budget for local models keeps shrinking llama.cpp added OCP-FP8 KV caches with static per-tensor scales, which halves cache memory compared with FP16. Two preprints go further: one raises weight precision only at uncertain decoding steps, and another finds that deleting weights in post-training works surprisingly well (preprint, unreplicated). Long-context inference on one workstation is becoming a matter of configuration. github.com | arxiv.org | arxiv.org

GLM-5.3-Flash is becoming the reference local model Z.ai published the technical blog for GLM-5.3-Flash, the fast tier of its open-weights GLM family. It is one of three models in an independent coding eval across Elixir, OCaml, Haskell and 10 other less common languages. XeMaya, a SYCL and Level Zero fork, runs it and DeepSeek V4.1 Flash natively on Intel Arc GPUs. z.ai | longtail-swe.com | reddit.com

The IPO and the new workspace tools are one bet: that Claude becomes the place where enterprise work happens, not just the model behind it.

Anthropic has launched its IPO roadshow Anthropic announced the roadshow for its proposed initial public offering, and a 57-slide roadshow deck is circulating publicly. As a public company, Anthropic will have to disclose pricing, capacity and enterprise revenue every quarter, which gives vendor-risk reviews of Claude much firmer ground. anthropic.com | x.com

Claude Sheets, Whiteboards and Data take aim at Office Anthropic launched Claude Whiteboards, Claude Sheets and Claude Data as AI-powered alternatives to Microsoft Office tools. Claude Data puts Anthropic straight into the analytics layer, so the question for a data team is no longer which model powers its BI tools but whether the model vendor becomes the BI tool. claude.com | x.com

10 items: ChatGPT for Business roundups from March and April 2025, a 2025 agent-experience retrospective, posts with no stated content, and community threads or anecdotes that make no claim.