AI News Daily · 2026-11-09

coding & agent

Coding-agent news in this window is about scale, what a model is allowed to change, and whether the interface should be rebuilt for agents. Bloomberg's figure for DeepSeek's open-source Harness is 1.54 million daily users on the official API, less than three months after launch, while OpenAI doubled Codex quotas on two paid plans and an apiiro count put 1,122 public GitHub files on Claude Code's "Don't Ask Again" settings. detailsdetailsdetailsOn the research side, Memorilla treats memory as a few tokens for a frozen model, and Meta's A/B agent is described as planning while deterministic code produces the numbers. detailsdetails

Platforms, quotas, and entry points

The 1.54 million count covers official API users only. After a desktop release, the seven-day average jumped from 271,000 on Sept 30. detailsDeepSeek frames Harness as an operating system for the agent era, and the project says built-in features can be replaced by plugins. detailsA 0.2.2 candidate adds experimental sub-agent plugins that hand standalone tasks to a local Codex or Claude CLI. The MIT-licensed repo is at about 276,000 stars and follows an "Everything is a Plugin" line. details

Thibault Sottiaux of the OpenAI Codex team confirmed that usage on Pro and Business 100 has been doubled. detailsSeparately, a Reddit user says an update allegedly set the default sub-agent to "Astra High," and that background usage then rose. detailsAmpersand, the collaborator model from Humans &, is on Vercel AI Gateway and free for a week. It has 744 billion parameters, 40 billion active, and the maker says it matches or beats frontier models such as Fable on its own long-horizon multi-user eval. detailstestingcatalog reportedly saw Google internally testing a next-generation Gemini agent for Gemini Enterprise, with custom MCP server connections in an internal UI. detailsPerplexity CEO Aravind Srinivas credits Teammates to a harness that runs agent traces largely in code, model neutrality, and search quality. detailsApple published an official Xcode plugin for Claude Code, installed with /plugin install xcode@claude-plugins-official on the latest Xcode. The visible capabilities are build, run, test, and SwiftUI previews. details

Evals, management, and orchestration

Garry Tan, president of Y Combinator, shared early open-source results on seven coding harnesses, including Claude Code, OpenClaw, Hermes, and QM on Opus, plus Codex, OpenClaw, and Hermes on a second model. On capability he says the setups still cannot be told apart, while cost differs by about 2x. detailsAt the AI Engineer event, McKinsey's Martin Harrysson and Prakhar Dixit said most large companies are seeing modest AI-coding gains, while the top 20-25% now report 2x end-to-end improvement. details

Addy Osmani, a Google VP of engineering, argues that a day with coding agents is already management: picking tasks, writing instructions, delegating, following up, and judging what can ship, and that this beats prompt engineering. detailsCoinbase CEO Brian Armstrong describes one always-on lead agent per focus area, a chief-of-staff agent above them, and sub-agents that write code, run tests, and chase reviews, with decisions collected in one prioritized queue. detailsesrtweet applies the bitter lesson in the other direction: no sub-agents, no accumulated skills, no unattended software factory, only a conversation that analyzes and writes code. detailsLangChain's "Agents Are Going Async" splits the loop into an orchestrator that stays available and background sub-agents that return when a long task finishes, instead of letting one long tool call block a ReAct loop. detailsGoogle open-sourced Elephant-Goldfish, a design-and-review plugin at about 14,700 stars. The line is to design with an elephant and review with a goldfish: design with context, review with fresh eyes. details

Memory and experiments

Memorilla, from Stanford, Toronto, Penn, MIT, and coauthors including Jure Leskovec, treats memory as a modality for a frozen LLM. A learned module reads a document collection and distills each query into memory for the model. The reported figures are 16 memory tokens, 4,212x fewer tokens, and 27x faster inference. detailsEvoMap's Evolver (arXiv 2611.07595) turns transient exploration into reusable "Gene" assets that are acquired, retrieved, and revised from feedback. The stated result is 4.4 percentage points higher success and 24% fewer tokens. detailsA Stanford paper, "Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies," runs 10 frontier LLMs through multi-resource trade, bilateral bargaining, and double auctions at different disclosure precision. The title result is that more information helps some trading agents and hurts others. detailsMeta's paper on AI-assisted A/B testing says the LLM plans while deterministic code computes every reported number, because experiments need figures that do not hallucinate. details

Tactico's desktop app Momentum keeps one local memory for Claude, Codex, Grok, and more than 30 other coding tools. detailsOn Reddit, madsthines, who is building LoreKit, argues that each memory system keeps its own storage and that the field is still in silos. detailsA Hacker News thread on the same gap proposes an Open Agent Memory Format. details

Security, permissions, and leaks

apiiro found 1,122 public GitHub files exposing passwords through Claude Code's "Don't Ask Again" settings. Anthropic says .claude/settings.local.json is auto-gitignored and that the leaked files were manually created and checked in. The report disputes that auto-gitignore account. detailsA separate report says Claude Code auto-compacted twice in 24 hours on context overflow, on both Opus versions at max effort, even with global autoCompactEnabled set to false, and that the action was logged as manual. details

Nous Research shipped an official Security plugin for Hermes Agent. It builds a threat model of attacker-reachable surfaces and trusted inputs, then splits the code across parallel sub-agents that look for vulnerabilities and check them. detailsPawel Huryn released an MIT-licensed skill that embeds a Claude agent allowed to propose a Stripe change, such as a customer address, while the harness outside the model performs the write. detailsLimestoneHQ's workflow has Claude Code upgrade vulnerable dependencies one package at a time, keep only upgrades that pass the build and tests, and leave a draft pull request. detailsReporting dated Nov 3 on the roughly $6.34 billion Binance theft says the months-long operation against wallet-infrastructure staff began in June, when an engineer searching for an AI coding tool downloaded a malicious package. details

One production operator sends irreversible or external actions, including email, record edits, and spending, through human approval, and lets read-only calls run. detailsexedev quietly added ssh exedev new --sandbox for per-minute sandbox VMs, and users found the command before it was promoted. detailsThe Hermes Chronos plugin snapshots the working tree, including untracked files such as .env, into a shadow git repo before file edits, so a later step can roll back. details

Interfaces, inference, and bills

Vork is a 3D and VFX engine whose agents write shaders, materials, cameras, and physics. The core API is about 2,500 tokens. The author claims 2x better token use than Blender, Unreal, and Houdini, and up to 24x faster rendering. detailsA separate early note argues that APIs designed for LLMs rather than humans are not worth it: human interfaces are theoretically suboptimal, but similarity to the training distribution matters more. detailsInterpretability researcher sunshowers6 writes that an agent writing Rust pattern-matches on existing code, so type-system invariants and community norms still constrain the output. detailskitlangton's language Blood uses one comptime primitive, a function that runs during checking and returns a type. Regexes are parsed at compile time, and the sketches also turn SQL typos into compile-time errors. details

RustingStrata, a pure-Rust engine, decodes Qwen3.6-35B-A3B (Q2_K_XL, 11.7 GiB, some experts on CPU) on a 12 GB RTX 3060 at 118 tokens per second on short prompts and 113 after a 4K prompt, about 1.7x llama.cpp at 68 and 66. The same release is also quoted at up to 160 tokens per second. detailsBuoyant's Kinference sits in front of vLLM workers and routes on cached prefixes and queue length. On shared-prefix traffic, reported median time to first token fell from 6.88 seconds under plain Kubernetes routing to 0.31 seconds. detailsdbreunig says that, although he likes Opus 5.5, most of his tokens go to lighter work: a software-factory pipeline on ds4-flash and glm 5.3, plus a Hermes agent on meta spark, run through LunaRoute's concurrent-lane billing. detailsBaseten Loops is generally available with a serverless option used by Ambience AI, Datadog, and LangChain for SFT and RL. The serverless tier keeps infrastructure warm, so there is no cold start, and bills per token. detailsCodaLabs published an MIT-licensed JavaScript MCP template that pays for X402 APIs with gasless USDC transfers signed under EIP-712. detailsA Github_private connector adds OAuth access to private repositories and organization data on top of public GitHub MCP coverage. details

Things shipped, and the review argument

A Reddit user built an MCP server, with Claude Code, that gives Claude a keyboard, mouse, gamepad, and screen capture inside a game while the desktop stays usable. The title demo is autonomous play in GTA IV and GTA V. detailsAnother project, dimension9.space, is a browser Backrooms driven by Claude Opus 5.5. It started from 15 hand-seeded rooms and the title count is 70 rooms. detailsBrowserBall is a browser 3v3 football game written almost entirely with Claude Code, with server-authoritative simulation at 120 Hz, client-side prediction, and WebTransport datagrams. detailsMall Bats, an isometric game set in a dying 2004 mall, is described as having nearly all of its code written by Claude Opus 5.5. detailsCinderscoop is a browser game, built with Claude Code, about flying an amphibious water bomber onto wildfires. detailsminchoi and his children shipped a playable Roblox game, "Capture A Beast," without writing code: Opus 5.5 for the concept, Roblox MCP for Studio, and Tripo for 3D beasts. details

vista8 spent three Codex reset cards and one Claude account over a weekend. AI added 64,300 lines and got the macOS recorder CapWise to a usable state. detailstetsuoai's music video "ATTENTION IS ALL YOU NEED" took 26 hours 51 minutes, 10.1 billion tokens, 131 agents, and 23,354 model calls from first lyrics to file. detailsThe MIT course "AI Engineering from Scratch" is listed at 503 lessons, 20 phases, about 320 hours, and about 48,000 GitHub stars, with implementations of backprop, a tokenizer, attention, and Transformers. detailsRohit Ghumare's writeup of the same curriculum says each algorithm is built from the math before a framework is introduced. details

YacineMTB's line on review latency is that a merge measured in weeks means the work is not important. The person he quoted said their median time to merge is hours, not weeks. detailsSeparately, Yuchenj_UW points to a blog that reportedly describes an AI lab firing an engineer in part for letting agents review and auto-merge non-production code. He says that if the account is true, the lab would rather leave pull requests sitting for weeks or a month. details

Apps

Application news in this window falls into three piles: subscription products folding generation into music, money, and hardware; indie software finished in a few days with coding models; and small tools that stay on the device. Spotify's AI remix is slated for Premium as early as Q1 2027, and Claude's Personal Finance beta is a US-only, read-only link to financial accounts. details details

Platform products

Spotify has shown an AI remix for songs from participating artists. The examples are folk turned into EDM, hip-hop into acoustic, and pop hits into lo-fi, with royalties for the original artists. It could reach Premium subscribers as early as Q1 2027. details OpenAI is teasing the Modretro Chromatic, a handheld that plays Game Boy and Game Boy Color games, at Supply Co. on November 9. It is a side project of Sam Altman. DevDay attendees were given a limited-edition transparent unit. details Separately, OpenAI is reportedly working on a gaming device rather than only a model for making games. That is unverified. Arabic commentator aziz4ai amplified that speculation. details

Anthropic's Personal Finance beta, limited to the United States, lets someone connect chosen financial accounts and ask questions in everyday language. Answers come from real balances and transactions, and the launch describes the link as read-only. details Peter Yang interviewed Ashley Alexander, VP of Product for OpenAI Health. Both of her parents were diagnosed with cancer within a week. She described using AI to prepare for doctor visits and to research treatment. details A Wirecutter piece says chatbot companies can read those private exchanges. A DuckDuckGo survey found that 32% of Americans had told a chatbot something withheld from their closest loved ones, and OpenAI says more than 230 million people a week ask ChatGPT health questions. details

Developer nima_owji posted a screenshot of Grok testing "Create a teammate." One click creates a bot that skips the setup form and configures itself in the chat, beside Create a Bot and Browse shared Bots. It has not been announced officially. details Rudrank Riyam says he now runs two xAI Ultra accounts and found a Grok Bot referral: the friend gets two free weeks, and the referrer gets a usage reset after that friend subscribes. details Marco updated Post Performance Bot, a third-party Grok Bot template that reviews an X account each night, flags which posts did well, helps revise drafts, looks at new followers, and tracks lost followers. details

Per ghacks, Microsoft removed PC-to-PC migration from Windows 11 and now points people to OneDrive-based Windows Backup. details Google is rolling out Data Manager (Beta) inside GA4, with a user-provided-data setup and a map of tags, SDKs, data streams, and connected products, aimed at centralized first-party data. details

Software finished in days

Over one weekend, vista8 used three Codex reset cards and lost one Claude account. AI added 64,300 lines and brought his macOS screen recorder, CapWise, to a usable state. The screenshot shows a recording and editing tool. details lavel0rz built BrowserBall, a browser 3v3 football game, writing almost none of the code by hand and using Claude Code instead. The match is a server-authoritative simulation at 120 Hz with client-side prediction, and it sends WebTransport datagrams. details Matthew spent four days with Claude Max on Kinoklip, an open-source C++ clone of DaVinci Resolve, aimed at Linux install friction, missing h264/h265/AAC codecs, and weak performance off NVIDIA. The write-up calls the result about 10x faster on Linux. details

zylcut.com is a local-first video editor for Mac, Windows, and Linux. Basic edits are free, and a project is treated like a VS Code tree. The app was built in two weeks with Claude Code as the main coder. details YANGPA is an open-source serverless messenger on Tor, with all of the code written by Claude Code. Traffic is end-to-end encrypted with post-quantum hybrid cryptography. Most of it is Rust, and the client is a Tauri/React app. The account of the build puts the size at about 265,000 lines across about three weeks. details dimension9.space is a browser Backrooms driven by Claude Opus 5.5. Fifteen rooms were hand-seeded before further distillation, the creator does not know what will appear next, and the count given is 70 rooms. details

historai.ca takes a historical topic or person and, about a minute later, returns a two-host episode with artwork. The listener can interrupt, then get a quiz and a suggestion to continue. On scriptwriting, Opus 4.8 beat 5.5. details aziz4ai shipped a free browser RPG set in AlUla: at night the player crosses the Hegra tombs against a serpent and a monster that erases memory. Programming and story came from Claude, audio from Halai, and the build took a little over two days. details jasonkneen's non-commercial fan game Slop Eisley has an AI-generated look, solo play, a multiplayer preview, and modes for flying, racing, and AT-AT runs. Version 0.6.104 joins a shared Tatooine capped at a configurable 50 players, refreshes the menu, and adds TRM-800 freighter naming. details details

Midnight Bell is a gothic card autobattler planned for Steam in Q2 2027. Claude runs C# inside the Unity editor through the Unity CLI and starts test fights on its own. The post also credits that workflow with balance, art, and a trailer. details builtbydimi rebuilt the Sao Paulo Grand Prix circuit in a day from city elevation maps and aerial photos. Nearly 3,500 people raced it on the first day, and the game lets you use your phone. details

On-device tools

NIDO is a phone assistant on llama.cpp: no account, no cloud, no subscription, and it still works in airplane mode, with offline chat and local memory. details DeepTutor, the HKUDS tutoring project, shipped version 1.7 as Mac and Windows desktop apps. The repository shows 56.5k stars and 6.9k forks, and the new version can generate narrated material from a topic. details Phosphene 5.0, under an MIT license, runs video, image, speech, and music models locally on Apple Silicon through MLX. The Mac installer does not need a terminal and can reuse Pinokio models. Video with sound includes LTX-2.5. details

Tactico's Momentum is a local desktop memory layer. Claude, Codex, Grok, and more than 30 other agent tools can share one persistent context instead of being rebriefed after every switch. details SnapJudge, a free open-source detector for scientific text by Yixuan Huang at Duke NLP, processed 25,219 submissions in about three weeks. It takes pasted text and LaTeX source. details Purrloader is a free open-source downloader for macOS, Windows, and Linux. A pasted link yields MP4 video or MP3, M4A, and lossless audio, including playlist batches. details Scott Leimroth, a deaf developer, built Signal Dog for a Nous Research Hermes Agent hackathon. The microphone listens for knocks, appliance beeps, and smoke alarms, then notifies in chat and does not keep the audio. details

Prices, health, and errands

Elon Musk relayed a rescue in Kazakhstan. Five people in an SUV broke down about 15 km from the village of Kystau Kurchum, with no mobile coverage. Starlink carried the call for help, and rescuers drove about 100 km. details One rider logged Uber moving from $22 at 9:10 p.m. to $38 by 9:12 and $54 by 9:14 after refreshing the map, with the app citing increased demand. A step he named was turning off Precise Location for Uber. details An Operator-style browser agent was used on an Airbnb bill: a mandatory $250 cleaning fee, then an extra $100 for excessive crumbs. The agent opened Airbnb Support. details

A Google Research model published in Nature estimates high blood pressure from an ordinary smartwatch, using the pulse and motion sensors already on the watch and no cuff for calibration. The comparison given is 65.8% of hypertension cases caught by the watch, against 55.3% for a clinic cuff. details Ant Group's Ling-3.2-flash demo starts from meal photos and writes an illustrated food report, naming dishes and ingredients and estimating portions and calories. details Dazzle, an iPhone assistant, skips the empty prompt by reading the camera roll. A photo of a jacket can be turned into a purchase, and the same flow is shown for a flat tire or a broken garage. details

Dooder calls MCP servers directly from an Apple Watch over cellular, with no proxy and no backend. One exchange checks the calendar for a free weekend, searches Google Flights and Airbnb, and writes the plan into Notion. details Squad, at squad.so, attaches to a ChatGPT, Claude, or SuperGrok plan the user already pays for and makes named agents, Tessa, June, and Milo. They connect to an inbox, QuickBooks, Shopify, and a CRM, and they run on a schedule. details HoloTuberKit, updated by Tks_Yoshinaga, places a live 3D call on a TrueDepth iPhone or iPad, sharing face, gestures, and upper body. A LiDAR Pro model can stream the scene, and the viewers named are Meta Quest and XREAL. details

Research

Formal proofs and benchmark scores moved in different directions. On a Tsinghua and Moonshot exam, GPT-6 Astra proves 70.0% of research-level computer-science theorems in Lean and resolves none of 30 open conjectures. details A chart of math arXiv shows weekly submissions climbing from about 800 in early 2025 to past 2,300 by late 2026, across major categories and faster than ordinary publishing growth. details

Weight edits and unreliable graders

MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab gave GPT-6 Astra raw access to the weights of Qwen3.5-0.8B plus a Python environment, and asked it to fix math-looping behavior. The edit reported in "Sparks of Weight-Space Programming" changes 0.016% of those weights and lifts MATH-500 from 7.9% to 48.1%. details

A separate oracle, Qwen3-8B with LoRA, reads the fine-tuned model's residual stream, attention-head behavior, and per-neuron firing, then writes the weight edit that is supposed to undo the fine-tune. On Qwen3-8B that edit closes 61% of the KL gap to the original model. details

Small changes to the test, not the model, were enough to reshuffle coding scores. Across 89 DeepSWE tasks, altering instructions or starter code split models that had all scored around 74% by as much as 23 percentage points. details An Amazon paper argues that rule benchmarks such as RuleArena and LegalBench often grade correct answers as wrong because the reference answers or the scoring code contain bugs. One scorer fix lifts LegalBench by more than 80 points. details

Theorems, conjectures, and what gets published

TCSExam contains 100 expert-reviewed tasks across four tracks, including Lean. GPT-6 Astra's theorem rate is 70.0%, and it resolves 0 of 30 open conjectures. details A separate theory note, inspired by STP (Dong and Ma, 2025), asks when prover-conjecturer self-play can let a language model improve at theorem proving. It was accepted as a NeurIPS 2026 poster. details

The Aaronson-Ambainis conjecture in quantum complexity is reportedly proved. Researcher quantumashley flagged the claim while correcting what she describes as a wave of wrong commentary on what the result would mean. details

Review standards are under the same pressure. cloneofsimo describes a proof that no human can read: it is generated and verified by AI, so neither a correctness check nor a search for the flaw can be done without the model. Elliot Glazer's reply is that the repository evidence goes well beyond "trust me bro." details Dan Romik treats a demand for Lean verification as a double standard, and notes that Glazer left his own name off the paper. details Andrew Critch cites Glazer's proof that ℤ is Diophantine over ℚ, obtained with Gemini (Astra) and Claude, as the sort of result he expects from teams that mix frontier models from more than one provider. details A New York Times piece by Cade Metz says OpenAI's math results challenge the standing hypothesis that a problem whose solution is easy to check must be hard to solve. details

Scott Armstrong's practical complaint is about disclosure. He argues that frontier models now heavily aid almost all papers, and that if statements such as Terry Tao's and the Leiden declaration require an AI-use note, the note has to be accurate. He also says honest disclosures are getting papers rejected while silence pays off. details More than 20 papers cite the nonexistent phrase "vegetative electron microscopy." The traced cause is an AI pass over a 1959 article that joined "vegetative" and "electron microscopy" across columns, after which the phrase spread by citation. details

Long context, memory, and attention

Latent Sparse Attention, from Prefrontal AI with Torsten Hoefler and James Whittington among the authors, targets attention once context passes a million tokens. Sparse attention saves compute, but finding the relevant tokens becomes the cost. The stated case is that indexing consumes 98% of prefill at 10 million tokens, and that the method speeds one layer by 22.7 times. details Zyphra Hierarchical Attention is a subquadratic sparse method that builds a recursive tree. A 35B model keeps its quality while reading 1-6% of the KV cache. details

Memorilla, from Stanford, Toronto, Penn, MIT, and coauthors including Jure Leskovec, treats memory as a modality for a frozen language model. A learned module reads a document collection and distills each query into a short set of memory tokens. The headline figures are 16 tokens, 4212 times fewer tokens, and 27 times lower latency. details AbridgeKV, following the MEMENTO paper, tests whether KV compression needs explicit text at all. Pause tokens read each finished generation block into latent codes. At about one quarter of the cache, accuracy matches the full cache, and inference is 3.2 times faster. details

SemiAnalysis says that, after DeepSeek V4.1 Flash, Xiaomi's next MiMo models move to a causal encoder-decoder called HySparse2. Prompts go through 25 of 49 layers, described as a causal encoder with a block of sliding-window attention. The claimed saving is five times less prefill compute at a million-token context. details

Distillation, scaling reads, and continual learning

Salesforce compares on-policy distillation with GRPO by held-out accuracy per training FLOP, across difficulty, model size, and domain. The reported saving is that RL-trained experts merged through on-policy distillation use 42% fewer FLOPs than distillation from an external post-trained teacher. details Hinted Scaling Laws, with Percy Liang, John Duchi, and Tatsunori Hashimoto among the authors, target the case where small models score near zero on a hard benchmark and leave no curve to extend. The forecast error they report falls from 21% to 4.3%. details

OpenAI's will depue separates two curves that log-log plots tend to glue together: training loss does follow a log-log trend, while downstream utility does not. His claim is that downstream capability scales roughly linearly rather than logarithmically. details David Krueger amplified a critique of claimed capability acceleration. The evidence being questioned is faster internal benchmarks and faster researcher output without faster growth in total compute, and the proposed confounder is a shift of that compute toward reinforcement learning. details On a six-regime streaming comparison, Apical Plexus and Apical Arbor store no raw replay data, against experience replay and against EWC or online SGD. The chart's lead over those baselines is described as up to 30 points. details

Agents, monitors, and self-improvement

Tencent's Terminal-RSI is a 27B pipeline for recursive self-improvement. The failure it targets is agent laziness: claiming success on partial progress and then failing the environment's final verifier. The description says the system rewrites and retrains itself with zero humans. details EvoMap's Evolver (arXiv 2611.07595) stores transient exploration as structured "Gene" assets that are acquired, retrieved, and revised from feedback. Reported gains are 4.4 percentage points of success and 24% fewer tokens. details

Meta's ABTest Agent is a response to a narrower constraint: experiment numbers have to be reproducible, and a language model is not. The split in the paper is that the model plans the analysis while deterministic code produces every reported number. details At the AI Engineer CODE summit, Tanmay Sah and Dolly Sah argued that finishing a task is not the same as being able to undo it. Git does not cover external API calls, runtime config, or notifications, and a retry does not reverse those side effects. On UndoBench, 42% of the gains that made an agent more capable also broke recoverability. details Tsinghua's T-STAR Lab introduces TraceMRT, which rewrites a hidden side task as a plausible sub-goal of the main task so the trace reads like ordinary execution. Against that rewriting, top chain-of-thought monitors are evaded about 60% of the time. details

Owain Evans posted questions rather than new scores. One is whether reward-hacking propensity actually falls, and how far, if post-training drops environments that are impossible or easy to hack. details Another shifts the poisoning threat during recursive self-improvement from a human insider to an AI that can poison the training process. details A third asks for a short practitioner note on subliminal learning, phantom transfer, and related subtle generalization, limited to what changes practice. details

How scientists use models, and what robots shipped

Y Combinator's Paper Club frames biology after the drop in sequencing cost: data now outruns researchers, and the opening the club describes is a move from reading biological data toward designing living systems. details Pushmeet Kohli, who leads scientific research at Google DeepMind, says several projects are in what he calls a real phase transition in the pace of discovery, with a first look coming in releases over the next few weeks. details

Anthropic's 4 November report covers about 207,000 science-related Claude sessions from 2 August to 30 September 2026, plus a survey of 528 scientists. Data analysis accounts for 33.7% of research sessions, rising to 48.1, while scientists barely use Claude for ideation. details

A Nature Medicine paper from Eran Segal's group films more than 3,000 people walking, lifts 3D skeletons, and trains a self-supervised GaitMAE model. Gait alone predicts BMI, visceral fat, age, and 1,887 traits. details In npj Digital Medicine, a locally deployed open-source model pulls 136,425 structured kidney-cancer fields from about 130,000 clinical notes, at 97.5% accuracy, in a setup aimed at ending manual chart review. details

Google DeepMind shipped MuJoCo 4.0, the first major version of the contact physics engine since 2023. Collision checks only see positions at discrete steps, which lets thin or fast objects tunnel; the release's fix is called speculative contacts. The repository sits at about 16.1k stars and 2k forks, and is used by about 6k projects. detailsdetails CoRL 2026 in Austin is sold out, on a record 2,094 submissions and 687 accepted papers. Keynotes include Wolfram Burgard, Russ Tedrake, and Fei-Fei Li, plus a conversation with Lisa Su. details

GenrobotAI's DAS Dex 2.0 records 21-DoF hand motion together with tactile readings from 980 points, capturing not only how a hand moves but how it interacts with the world. details Delta Intelligence's Δ₀ uses one generalist policy for a 69-DoF humanoid, aimed at long-horizon, contact-rich whole-body manipulation at roughly human speed. details VisionSIM, a CoRL 2026 spotlight, renders one Blender scene into RGB, thermal, event-camera, LiDAR, IMU, and single-photon streams, with motion blur, photon noise, event thresholds, and pixel-aligned ground truth. details A CMU Intent Robotics Lab group working with Amazon FAR is bringing three workshop spotlights: IMPRINT learns multi-stream latent memory without manual curation, and the set also covers vision-touch steering and world models. details

Models

The window is mostly unreleased-model sightings and score disputes, not shipping notes. OpenAI's developer account posted only "Tomorrow," while separate reports say some GPT-6.1 Sol traffic is already landing on an unreleased GPT-6.2 Sol. detailsdetails Qwen 4, a rumored Claude family, and Gemini 4 Carbon sit in the same gap: leaks, internal strings, or a number-one listing nobody can call. Instruction edits and broken graders are moving benchmark ranks at the same time. detailsdetailsdetails

Unreleased and quietly routed models

OpenAIDevs posted a link and the single word "Tomorrow," with no product named. The account's scope points at developer tools or the API, and nothing else is public. details kimmonismus, citing an observation from TokenGremlin, says GPT-6.2 Sol is already appearing for some ChatGPT users, with a release expected next week. The shared sample is a Three.js scene from one prompt, packed into a single HTML file. details A separate user test asks GPT-6.1 Sol who hosted the Tony Awards on 7 June 2026 and forbids web search, treating a correct answer as evidence of a newer knowledge cutoff and therefore a quiet route to GPT-6.2 Sol. OpenAI has not confirmed that. details A researcher also found three mystery IDs in Devin: aurora-alpha and halo-alpha, aliases cactus and kraken, each with five reasoning-effort levels from low to max, plus claude-donut-eap beside Opus 5.5. The post reads the first pair as a hint of GPT-6.2 and the third as a hint of Fable 5.5. details

Qwen 4 timing is still second-hand. One post cites the Qwen CTO as saying the model would ship this week, treats today as the window, and calls the result a "positive surprise," with no specs attached. details Another Reddit claim puts Qwen4 27B at 15:00 Beijing time on 9 November, and the poster labels it a rumor. details A developer with API access to unreleased Qwen-4-Flash calls it the smallest yet strongest model on GameFeel at about 35GB in Q8, while burning four times as many reasoning tokens as any model tested. teortaxesTex speculates that a Qwen4-27B of that size may be a sparse MoE. That is a guess, not an architecture disclosure. details The same bench has added Qwen-4-Flash-Early, codename hawk, with an audio score already computed. Alibaba has not confirmed the model. details Corrected GameFeel Bench v1.4 puts Qwen-4-Flash-Early first at 49.9 after the maintainer miscalculated Qwen-4-27B-Early's audio score. details The most concrete DeepSeek note is still unofficial. After a service interruption, LobeHub's Max re-ran a private LobeBench Hard 50: 50 hard agent tasks from real scenarios, three attempts each. The report's title says the V4 Pro pass rate jumped 16 points. DeepSeek has not confirmed an update. details An unverified Kimi roadmap attributed to the X account bebuu_ai lists Kimi K3 Flash by the end of 2026, replacing K2.8 and aimed at DeepSeek V4.1 Flash, then a 3T Kimi 3.5 in early 2027 that answers simple questions faster, and a 5.4T K4 in 2027. Moonshot has not confirmed the list. details

Gemini 4 Carbon was announced late last month as state of the art and ranks first on Artificial Analysis, yet it is unavailable to free, Pro, and Ultra users and unavailable on the API. details haider1 guesses a public rollout around 10 November by copying Gemini 3 Argon, which reached AI Pro and Ultra about 15 days after announcement. He marks that as speculation. details A separate Reddit post says a codename, Gemini 4 Hyperion, has shown up in Antigravity's CLI and app, with a rumored 99.7% on SWE-Bench Pro. The post says every figure traces to scattered leaks and is unverified by Google. details

On the Anthropic side, WorldofAI says Claude Donut EAP has appeared in Cognition's Devin CLI and is mapped to a family called Silicon, and that a hidden model, Obsidian 3, beats Opus 5.5 on coding. Both are third-party leak claims. details A new Claude checkpoint (Fable/Opus, Sonnet 5.5 Extra), asked where Karpathy went in 2026 with web access off, answers that he joined Anthropic's pretraining team in May 2026 to speed up pretraining research. That is model output, not a staffing announcement. details

Elon Musk says fixes for the Grok bot are underway, aimed at token efficiency and smarter behavior, with no timeline. details Developer buildwithhassan reports a different change: fingerprint tests after a tone shift came back as gpt-6-astra/6.1-sol rather than Opus 5.5. Musk had already said the bot was too expensive to run. The fingerprint result is an observation, not an xAI statement. details

Plans, limits, and outages

thsottiaux clarified Claude subscription limits: no Pro plan has a five-hour cap. Against Plus, Pro 100 is 10x, Pro 200 is 20x, and Pro 500 is 50x, with access to the ultrafast tier. details

Responding to complaints that Plus gained nothing for weeks, reach_vb lists Images 2.6 on Plus four days earlier, with sketch-to-GIF and 50% faster image generation, plus 3x GPT-Live voice time and more Codex tasks. details Thibault Sottiaux publicly asked whether a "Pro 50" plan makes sense, which readers take as a trial balloon for a $50 monthly tier between Plus and the $100 Pro plan. It is not a launched SKU. details Anthropic's status page said at 08:28 UTC on 8 November that it was investigating elevated errors and degraded performance on Claude Fable 5.1. details

How the scores moved

A coding study circulating on Reddit changed instructions or starter code on 89 DeepSWE tasks. Models that had all scored around 74% on the original set ended up as much as 23 points apart. details MaxxBench prints a raw score next to an estimated benchmaxxing component. Claude Opus 5.5 scores 76.4% on DeepSWE and still leads at 74.2% after that estimate is removed. details A separate pass over the same dispute, with Opus inspecting the data, treats the pass-rate drop for GLM, DeepSeek, Meta, and Gemini Flash as overfitting: models fail modified checks because of prior expectations, not because capability collapsed. details

An Amazon paper argues that rule-following benchmarks such as RuleArena and LegalBench often mark correct answers wrong because reference answers or scoring code are buggy. The title says one scorer fix lifts LegalBench by more than 80 points. details Garry Tan's early open-source eval cannot yet separate capability across seven harness setups, including Claude Code, OpenClaw, Hermes, and QM on Opus, plus Codex, OpenClaw, and Hermes on a second model. The headline result is a tie on capability and about a 2x cost gap. details

Julius Research released DeepFinance, 21 ultra-long-horizon finance problems built from real deal data by people with Evercore, JPM, UBS, HBS, and Wharton backgrounds. GPT-6 Astra leads at 57.6%. details A Tsinghua and Moonshot paper, TCSExam, has 100 expert-reviewed tasks across four tracks, beginning with Lean theorem proving. GPT-6 Astra proves 70.0% of the research-level CS theorems in Lean and solves 0 of 30 open conjectures. details

Papers and training reports

MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab published "Sparks of Weight-Space Programming." They gave GPT-6 Astra raw access to Qwen3.5-0.8B weights and a Python environment to stop a math-looping failure. The reported result is an edit to 0.016% of the weights that lifts MATH-500 from 7.9% to 48.1%. details

Robert Scoble's breakdown of Tencent's Terminal-RSI describes a 27B recursive self-improvement pipeline aimed at agent laziness: claiming success on partial progress and then failing the environment verifier. The title says the model rewrites and retrains itself with no humans in the loop. details

Anthropic's 4 November report covers about 207,000 science-related Claude sessions from 2 August to 30 September 2026 and a survey of 528 scientists. Data analysis is 33.7% of research sessions. The title says scientists barely use Claude for ideation. details

Nace released Drex Max v1.1, tied for first on the Artificial Analysis Decision Index at 73, on the Pareto frontier, about 8x cheaper than the next model in category, with about 256k context and open weights on Hugging Face. details SemiAnalysis says Xiaomi's next MiMo, following DeepSeek V4.1 Flash, moves to a causal encoder-decoder called HySparse2. Prompts pass through 25 of 49 layers, described as a causal encoder with sliding-window attention. The title claims 5x less prefill compute at 1M context. That is an analyst account, not a Xiaomi release note. details

Ampersand, the collaborator model from Humans &, is on Vercel AI Gateway and free for a week. It has 744B parameters with 40B active, and the launch note says it matches or beats frontier models such as Fable on the lab's own multi-user long-horizon evals. details The team separately reports 77.9% on DeepSWE v1.1 in multi-agent mode without targeted training, matching or beating Astra and Fable, and an EQ-Bench ELO of 1380.5. details

Open weights and behavior reports

169Pi launched Alpie Seed, a 9B multimodal model billed as "built in India," with 82.54% MMLU-Pro, 84.32% GPQA Diamond, 72.99% LiveCodeBench v6, and 61.20% SWE-Bench Verified. Inspection described in the post says the architecture is a near bit-for-bit copy of Qwen3.5-9B. details LiquidAI published official NVFP4 4-bit weights for LFM3-1.7B and LFM3-VL-2B on Hugging Face. details antirez says the kimi-k3 branch of DwarfStar can now run Kimi K3.1 PRO inference from safetensors with SSD streaming. He calls it slow but usable for QA. details A Reddit thread says Tencent's Hy4 is already offered free through WorkBuddy, next to DeepSeek V4.1 Flash, as an argument that Chinese labs are now competing on performance per dollar rather than sticker price. details

A developer who had Gemini read 1,400 of his own Claude Code sessions says Opus 5.5's reasoning style shifted around 18:00 UTC on 3 November, with the model name and CLI version unchanged. details Polymarket prices a Chinese lab at 7% to hold the top model by 31 December, on about $300K of volume, and 16% for third. details

Multimodal

Video length and control were the concrete claims: Seedance 2.5 produced a 30-second single take from one prompt, and Gemini Omni 1.2 Flash raised the single-clip cap from 10 seconds to 15. detailsdetails On images, WeChat's WeLM-Image is a 31.7B model with about 5.4B parameters active and no open weights, while MOSS-Image 5B was shown running entirely on a phone NPU. detailsdetails The detailed production counts were on MiniMax H3: a 64-second character swap, and a one-hour film assembled from 485 short clips. detailsdetails

Video length, shots, and code

On Dreamina, Seedance 2.5 at 1080P turned one long prompt into a 30-second factory flythrough with no drone and no cut. The visible path goes through broken glass, grazes rusty machinery, splashes a puddle, and continues upside down. details A GTA6-style Diwali short on Higgsfield broke in a single 30-second generation: a messy zipper close-up, the lead standing on a car, and a failed final 10 seconds. The fix described in the post was 12 scenes of 10 seconds plus physics prompts. details A separate K-pop camera-angle test extracts dance motion with RGBMoCap (ExMR), re-renders it in Blender, then transfers the choreography with Dreamina CLI and Seedance 2.5. details

Chuangyebang reported that Gemini Omni 1.2 Flash can be set anywhere from 3 to 15 seconds, adds multi-turn conversational editing, and improves character, object, and scene consistency across shots. details A full prompt for Google Flow Omni 1.2 Flash generates a 15-second multi-shot of three elderly women in peasant clothing playing modular synthesizers on a farm. details Gabriel Contreras argues that Sora is still the only video model whose output is indistinguishable from real footage; other models, he says, resemble early GPT-4o text and show mode-collapsed slop after about ten outputs. details Creator @robertsmith_ai claims one Veo 3 ASMR glass-slicing clip reached 66.1 million views and about $8,570. details

Microsoft said fal is coming to Microsoft Foundry, also labeled Azure Foundry, with Kling, LTX, and MiniMax available through Foundry Models. details VideoRebirthAI's hyperflow-h3-flash, an image-text-to-video model built on MiniMax H3, was trending on Hugging Face. It supports text-to-video, image-to-video, reference-to-video, and audio-video generation, in safetensors and GGUF. details Someone still waiting on MiniMax H3 open weights used Krea-2 Turbo for a three-panel joke whose punchline is a locked, watermarked Lakeview frame. details

SenseTime open-sourced SenseNova 6.8 Code2Video, billed as the first major model built for code-driven video. Instead of a black-box file, it writes code for visuals, animation, timing, and structure. details Microsoft filed a patent for live and post-produced video that recognizes speech and rephrases it with generative models according to viewer age, region, and language, The filing also covers matched lip movement. details cinecode, MIT-licensed with 1.1k GitHub stars, turns a coding agent into a one-sentence-to-MP4 pipeline: the agent writes the script and builds frames in code. details

tetsuoai's AgenC music video "ATTENTION IS ALL YOU NEED" took 26 hours 51 minutes, 10.1 billion tokens, 131 agents, and 23,354 model calls from first lyrics to final file. details A one-hour Jesus film was stitched from 485 MiniMax H3 ref2video clips of about 8.5 seconds, int8 with a 4-step turbo LoRA at 832×480. Keyframes came from gpt-image-2 via OpenRouter, and audio started in InWorld TTS-2. The author called the result terrible. details The opening of "The Three Garridebs" was adapted into an anime of about 10 minutes with ComfyUI orchestrated by Claude Code through the open-source tool konte, with project files published. Krea 2 Turbo is used for character sheets and backgrounds. details

MiniMax H3, LTX, and consistency

A ComfyUI workflow swaps characters across a 64-second, 1080p MiniMax H3 clip cut into 15 segments, with an agent skill that writes prompts and batch-runs the segments. details One tester argues that H3 single-frame editing already follows prompts and keeps identity and art style from a reference, which may make a character LoRA unnecessary. details Another found a reusable 360-degree turnaround video better than a character LoRA alone, because one face shown from every angle reduces the face drift that comes from multi-image references. details

A ComfyUI graph combines MiniMax H4/H3 Ripple with Pika Flex and Singularity V2 so that editing one frame changes the whole clip, shown on a robot walking on a beach. details A dark-fantasy series used H3 in ComfyUI for an axe-versus-mace fight, with most of the edits on the frames where the weapons meet, including a floor parry around 0:16 guided by FL2VA. details RelayVSR 2× plus a DLSS 5 pass took an 8-second H3 shot from 960×544 to a 1920×1088 SDR master, keeping more chainmail and helmet detail in motion. details

The Viggle Remake MCP server added a ComfyUI worker. On an RTX 5090 the author runs the int8 model and streams weights from system RAM. It is verified at 32GB, with 24GB and 20GB tests still going. details An LTX 2.5 rap video held character, outfit, and location surprisingly well, but the motion is mushy and facial detail breaks down. details AntiMatter Comfy repackaged three free LTX-2.5 graphs from Comfy Org and Lightricks: text-to-video with audio, image-to-video, and reference-sheet control. details Pika plus NVIDIA Open 2K was used for a 2K beach scene of a wave turning into a dragon. The same author also ran MiniMax, Pika, and NVIDIA Open 2K on the free Colab included with a Gemini subscription, with a 1K sample of that beach dragon. detailsdetails

Image models and on-device inference

bdsqlsz flagged WeLM-Image and argued that, without an architecture change, image models are at their limit and the next race is smaller models. On world models, the same developer treats video models as temporal causal diffusion and says adding spatial context raises both randomness and cost. detailsdetails Andriy Burkov argues that scale invariance may explain why autoregressive image generation lags diffusion. Left-to-right tokens are attractive for a unified text, speech, and vision system, but the gap he describes shows up with standard continuous latents. details

HexAnima 1.3.0 runs OpenMOSS MOSS-Image 5B (Apache-2.0) fully on a phone Hexagon NPU, with no root, cloud, or account. The Qwen3.5-9B text encoder, 5.1B DiT, and FLUX.2 small VAE stream as 8-bit tiles. On a Snapdragon 8+ Gen 1, 512×768 takes about 30 seconds. details A published comparison says Karras and Exponential schedulers break MOSS generations, while five others, starting with Simple and Normal, are nearly identical. details Fizgig, an open-source LoRA studio with 689 GitHub stars, cut a MOSS-Image LoRA run on an RTX 5090 from about 49 minutes to about 20. The run was 94 images, batch 1, Muon, and 70 epochs. details

lodestones shipped a Kromoss update, including an OPD-priming checkpoint at update 22000, trained onward from a GLM-caption plus bbox layout SFT. details On a 12GB RTX 4080 Laptop, the same five cinematic prompts and seed were run through local Z-Image Turbo, Krea 2 Turbo, and FLUX.2 klein 4B/9B, plus hosted Nano Banana 2.1. Nano Banana 2.1 scored 46 out of 54. details A reusable GPT Image 2 template in ChatGPT produces muted pastel paper-collage travel posters, with New York and Seoul examples. details A separate prompt targets hyperrealistic 3D celebrity caricatures: oversized heads, pores, wrinkles, subsurface scattering, PBR, and studio light. details GPT-6.1 Sol was used for voxel pieces including a gaping shark, a peacock, and a bicycle. details

Ant Group's Ling-3.2-flash demo turns meal photos into an illustrated food report, identifying dishes and ingredients and estimating portions and calories. details

Voice, 3D, and code-built pictures

Shizuku AI Labs introduced Shizuku-Duplex, which it claims is the world's first full-duplex audio synthesis model built for AI characters, with a public Japanese demo. details An open-source vocal synth models the tract as a network of branching acoustic tubes instead of Vocaloid-style stitched, pitch-shifted phoneme recordings, and adds MusicXML support. details Spotify has reportedly described an AI remix feature that could reach Premium subscribers as early as Q1 2027, turning folk into EDM, hip-hop into acoustic, and pop into lo-fi, with royalties for the original artists. details

ravisparikh's Vork is a 3D and VFX engine whose agents write shaders, materials, cameras, and physics. The core API is about 2,500 tokens. He claims 2× better token efficiency than Blender, Unreal, and Houdini, and renders 24× faster. details Trellis and Pixal3D image-to-3D runs in ComfyUI about a month ago were described as unusable next to closed models. details

Phosphene 5.0, MIT-licensed, runs video, image, speech, and music models locally on Apple Silicon through MLX, with a one-click Mac installer that reuses Pinokio models. Video with sound includes LTX-2.5. details Astra at its high setting, given reference images, produced PS1-style dungeon graphics in the vein of King's Field, with a command input that reads as an interactive demo. Separately, a Blender user showed 360-degree video tracking working. detailsdetails

Readable Node is an open agent skill that disassembles a messy ComfyUI workflow JSON and rebuilds it under readability rules, including no hidden nodes, no hidden wires, and a left-to-right order. details ComfyUI-from-your-Phone, GPL-3.0, was written with Claude as a free alternative to a $5-a-month remote. It reaches the desktop over Tailscale, with no cloud. details A quantized Qwen build labeled Qwen3.8 Flash Next GSQ RCO IQ3_XXS, on an RTX 4080 with 16GB of VRAM and 64GB of DDR5, decoded at about 95 tokens per second and prefilled at about 3,000, and was used for about two hours of Opus-style motion graphics via strata and opencode. details Awesoft's test of idiotproof had Opus drive a client video end to end, with rendering in idiotproof and no extra Fable tokens. details A 2016 NYC subway-car asset in Unreal had every texture upscaled to 4K with SeedVR2 for a music-reactive scene. details

Infra

Power, not a new accelerator, carried the day's infrastructure figures. Anthropic has reportedly agreed to a $3.4 billion deal for gas-fired generators details, and an estimate for Colossus II puts nameplate capacity at 1.7 GW, up from 600 MW a year earlier details. A bipartisan Senate bill that could fast-track U.S. AI data centers is reportedly lined up for a vote as early as next week details.

Gas, prices, and permits

Polymarket reports that Anthropic has reportedly agreed to buy gas-fired generators for its AI data centers at $3.4 billion details. The Financial Times reported the same sum as a deal to procure gas-fired power for the centers that train and serve its models details. In a comparison of on-site plants, xAI's portable gas turbines in Memphis were called close to a worst case because they had no exhaust cleanup, while Anthropic's units are described as gas engines that likely include NOx controls details. A comment on a CNBC report, tagging Anthropic, warns that people living next to the generators would take the noise and the NOx details. The New York Times reports that operators facing long grid-connection queues are scavenging retired jet engines from boneyards to use as temporary generators details.

Chamath argues that the SpaceX-Anthropic compute contract has a 90-day termination clause: if spot energy prices rise by 50% or more because supply is short, SpaceX could walk away and resell the compute at two to three times the price, and Anthropic would have to accept details. On permits, Polymarket says lawmakers are reportedly considering a vote as early as next week on a bipartisan Senate bill that would fast-track AI data-center construction across the United States details. A separate Polymarket contract puts the chance that any state enacts a data-center moratorium at about 27% (35 cents) by the end of 2026 and 90% by the end of 2027, and the rules count only state-level legislation details. In Leeds, a proposed AI datacentre drew protests, and MP Richard Burgon tabled a motion for a national moratorium. Commentator danbarker counters that the project would create 1,000-plus jobs and use about as much water as five car washes details.

A study reported by The Washington Post finds that data centers have had only a marginal effect on residential electricity prices details. Cathie Wood, on ARK's In The Know, says the electricity-price concern is real and that the broader fears around data-center buildout are overstated details. Satya Nadella says datacenters on closed-loop cooling use roughly as much water in a year as one restaurant details. Vitalik Buterin says he does not share the sharp backlash against this week's anti-data-center protests: algorithmic progress, in his account, means work that would need bigger data centers in two years can be done with today's data centers in four details.

Nameplate, orbit, and sites

An estimate for Colossus II has nameplate capacity moving from 600 MW in September 2025 to 1,400 MW in June 2026 and 1,700 MW in September 2026. That is 1.1 GW in 12 months, about 183%, with 300 MW added in the third quarter alone. The year-end target given with those figures is about 1.2 million Blackwell GPUs details. Bank of America says global data-center spending has passed $1 trillion and is projected to reach $2.1 trillion by 2028, a figure it links to NVIDIA, AMD, and Broadcom details.

Elon Musk endorsed a projection that 100 GW a year of orbital AI compute, priced at $10 per watt, would add $1 trillion of annual revenue to SpaceX, or $3 trillion to $5 trillion a year at higher prices, with the revenue stacking each year details. He also backed a secondhand report, relayed by Patrick O'Shaughnessy, that Larry Page puts the share of compute that ends up in orbit at 99.999%, or five nines. The note is framed around U.S. data-center regulation, and Martin Casado argued that the driver is regulatory details. Polymarket says Google added nearly 4 million SpaceX shares last quarter and now holds about $69 billion of the stock details. Starlink reached 13.6 million consumer subscribers by the end of September 2026, adding 1.6 million in the quarter, about 533,000 a month, 17,400 a day, or 12 a minute details.

YC-backed Atomarine announced a $19 million seed from Y Combinator, Pantera Capital, Soma Capital, 7SevenSix, Liquid2V, AudaciousHQ, and e14fund for nuclear-powered floating data centers aimed at the physical buildout bottleneck details. Jinko Power Technology plans a 1 GW solar-powered data center in Zhongwei, Ningxia: RMB 2.45 billion, about $361 million, on 530,000 square meters, breaking ground this year with completion targeted for 2028 details. Public-land purchases in Texas over the past two years are not limited to SpaceX and Tesla. Names on the list include Meta, Blueprint, CloudHQ, and Vantage details.

Memory, boards, and export approvals

CXMT, founded 10 years ago in Hefei, is described as the world's 32nd-largest company and the fourth-largest DRAM maker, behind Samsung, SK Hynix, and Micron, with about 10% of global share and a $430 billion valuation details. TSMC has publicly denied a report that it began drafting 3D DRAM plans at Nvidia's request and said it will remain a pure-play logic foundry. The surrounding point is that AI demand has drained DRAM supply, making memory the bottleneck details. On a podcast with Stripe's John Collison, Sam Altman called it quite terrible, from a global-allocation view, that limited DRAM sits in personal computers used 5% of the time. Critics blame the squeeze on OpenAI's datacenter buildout details.

Per The New York Times and The Verge, New Matrix was approved to buy Nvidia H200 chips for assembling AI servers, yet it operates from an empty office. The building directory and property managers have no record of the 17-month-old company details. Per Seoul Economic Daily, Samsung, SK Hynix, and Micron are working with SEMI on shared standards for fab robots, covering operation, charging, and production-schedule interfaces, with fully automated fabs discussed as a 2030 target details. Newegg has had 32 GB RTX 5090 cards in stock for several days at about $6,759.99, a Gigabyte AORUS ICE bundled with an X870 motherboard, down from the $10,000 to $11,000 range of recent weeks and still far above list price details.

Serving stacks and local runs

Together AI says it lifted Step 3 serving throughput 3.47x against StepFun's time-to-first-token and tokens-per-second targets. The search covered MoE routing, scheduling, kernels, speculative decoding, and sparse attention, and each change was benchmarked before the next details. Buoyant's Kinference sits in front of vLLM workers and routes on cached prefixes and queue length. In its own test, median time to first token on shared-prefix traffic fell from 6.88 seconds under plain Kubernetes routing to 0.31 seconds details. A Reddit spot of Tencent's Qwen3.8-27B-DSpine-b16 on Hugging Face describes it as a speculative-decoding method that speeds 27B decoding by about 3x, with the technical detail still unconfirmed details. Prefrontal AI, with Torsten Hoefler and James Whittington among the authors, published Latent Sparse Attention for contexts past 1 million tokens: at 10 million tokens, indexing eats 98% of prefill, and LSA speeds one layer by 22.7x details. Zyphra's Hierarchical Attention is presented as a subquadratic sparse method that keeps 35B model quality while reading 1% to 6% of the KV cache details. AbridgeKV reports matching full accuracy at a quarter of the cache and a 3.2x speedup details.

RustingStrata, a pure-Rust engine, decodes Qwen3.6-35B-A3B (Q2_K_XL, 11.7 GiB, some experts on the CPU) on a 12 GB RTX 3060 at 118 tok/s on short prompts and 113 tok/s after a 4K prompt, about 1.7x llama.cpp at 68 and 66 tok/s. The reported peak is 160 tok/s details. llama.cpp merged PR #30966 on Nov. 7 so that missed MoE experts can run on the CPU instead of a PCIe copy into VRAM details. antirez says the kimi-k3 branch of DwarfStar can now run Kimi K3.1 PRO straight from safetensors with SSD streaming. It is slow, and he says it is good enough for QA details. On large batches, he says the path switches to layer-major inference and runs nearly as fast as a fully resident model details. A compression of Kimi-K3.1-Pro, 2.8 trillion parameters and natively about 1.77 TB, normally about 22 GPUs of 80 GB each, cuts the footprint to 775 GB, down 56%, and runs all 93 layers on one RTX PRO 6000 Blackwell plus 176 GB of RAM and NVMe, with a claim of no accuracy loss details.

An Intel Arc Pro B70 with 32 GB, priced at $1,299, against an RTX 5090 at $6,500 or more, ran Qwen3.8-27B Q6 at about 52 versus 98 tok/s in generation and about 960 versus 2,830 tok/s in prefill: about half the speed at about a fifth of the price, or roughly $41 per GB versus more than $200 per GB details. Across 11 identical MiniMax H3 video rows with the same seeds, a 24 GB RTX 3090 beat both a 128 GB ASUS RTX Spark and an M5 Max on every row, even while weights streamed over PCIe x4 details. The LexiSuite author claims Qwen3.8 27B on a Radeon 7900 XTX under Vulkan is 5% to 25% faster than llama.cpp details. On two 256 GB Mac Studio M5 Ultra machines, running two separate models beat a tensorfold/exos tie-up that suffered poor concurrency and uneven decode gains, and a GLM q4 is said to beat the DeepSeek API for coding details. macOS 27.3 lets Core AI inside a Mac VM use the host GPU and Neural Engine, up to 7x the guest CPU and 50% to 100% of native host speed in the tests described details.

Open-weight models handled 72% of tokens among OpenRouter's 50 highest-volume models by late September 2026, up from 27% a year earlier. Token-weighted list prices of closed models were about four times those of open-weight models details. McKinsey's 2026 outlook says multi-step agents consume 5x to 30x the tokens of simple chat, so the extra steps are a capacity choice as well as a software choice details. An analysis of METR's numbers has Anthropic's measured annual gains rising from +15 to +29 points a year between the first half of 2025 and the first half of 2026, while R&D spending rose from $1.85 billion to $9.74 billion, or 5.27x, and progress per dollar fell 63% details. One Anthropic API user found the five-minute prompt-cache window made bundled credits wasteful: the console showed Claude Opus 5.5 burning 2.6 million input tokens when reuse missed that window details. Baseten says Loops is now GA, with a serverless option that stays warm, has no cold starts, and uses per-token pricing. Ambience AI, Datadog, and LangChain use the SDK for supervised fine-tuning and reinforcement learning details. Hugging Face free personal accounts with a verified email and more than 30 days of age can host up to two ZeroGPU Spaces, at about five minutes of GPU time a day details.

Embodied

CoRL 2026 in Austin is sold out, with a record 2,094 submissions and 687 accepted papers. One attendee puts the meeting at 5,000-plus people, up from about 400 five years ago. details details The keynote list runs from Wolfram Burgard and Russ Tedrake to Fei-Fei Li on the Atlas world model and a fireside chat with AMD CEO Lisa Su. details The same day also has a numbered software release and two factory counts: MuJoCo 4.0, PrimeBOT's 10,000th personal humanoid, and a dual-arm robot feeding SMT lines at Foxconn. details details details

Agenda, papers, and hiring

Wolfram Burgard speaks Monday on what still blocks embodied AI. Tuesday is Russ Tedrake, MIT faculty and CEO of the stealth company Walden Robotics, on learning inside factories. Wednesday is Fei-Fei Li on the Atlas world model, with the Lisa Su fireside on the same day. details Princeton's PRISM group is presenting six papers. The oral, MolmoB0T, trains only in simulation and is described as zero-shot manipulation on a real robot. details

BitRobot's booth has a full row of MagicLab humanoids. Its Nov 9 lightning talks include RoboColosseum with NUS MAGIC Lab and Jiafei Duan at 13:00, and MolmoAct2, on action-reasoning models for real deployment, at 13:30. The lineup title also names RoboEval and RoboTTS. details details Ted Xiao says this is the first CoRL he has missed since 2017. Daniel, Oier, and Anh are in Austin recruiting for Prometheus on physical AGI roles. details Apptronik is running private recruiting during the meeting, mostly full, for locomotion and whole-body control, dexterous manipulation, and sim2real. details

Policies, touch, and what labs buy

Pick Anything, from Jitendra Malik's Dex-Hands group at Amazon FAR, is one policy for human-like grasps across 66 object types, from AirPods to dumbbells to pipe wrenches, using a stereo camera head. details Delta Intelligence's Δ₀ is a single generalist policy on a 69-DoF humanoid, for long-horizon, contact-rich whole-body loco-manipulation at roughly human speed. details

RAX scales online exploration instead of demonstration data. A coding agent connects to the robot's perception and control APIs and learns from feedback. The title result is a 0.1 mm insertion with zero demos. details Energy-Based Policy Orchestration, from Yilun Du's group with Pieter Abbeel, Leslie Pack Kaelbling, and Tomás Lozano-Pérez, assigns an energy value to each humanoid whole-body skill. The title claim is a 7x drop in policy switches. details GLIDE uses an LLM to edit teleoperation code. Raw VR teleop succeeds on 0-10% of attempts for tasks such as picking cherry tomatoes or pouring wine, and the title puts hard-task success at 90%. details A force-sensor demo makes the opposite comparison: hours of fine-tuning and about 80% success with a robotics model, versus a one-shot insertion with an agentic harness and a sensitive 6-DoF force sensor, described as industrial reliability. details

ORA runs real robot arms from vision alone on open-weight DeepSeek V4.1-Flash, zero-shot and without a VLA, and is described as competing with frontier-model agents. The title cost is $0.35 per task. details T-WLA, from MIT CSAIL and Harvard, is a world-language-action model for the case where contact changes faster than vision. The title figure is a 30% cut in finger-joint MPJPE. details DAS Dex 2.0 records 21-DoF hand motion and 980 tactile points at the same time. details RoboCap now takes tactile readings through MX Matrix Gloves, capturing contact and pressure with the visual demos. details After 180 meetings in 46 days, Sam Padilla's read of robotics data heading into 2027 is that mono data is dead, ready stereo is scarce, and tactile is the frontier. details Rhoyn's YAM Ultra demo is a teleoperated massage with grippers in a gentle mode, fed by three RealSense cameras, nine video streams, and depth, shown ahead of CoRL in Austin. details

Simulators, world models, and evals

MuJoCo 4.0 is the first major version of DeepMind's robotics physics engine since 2023. Collision checks only sample positions at discrete timesteps, and the headline change, speculative contacts, is aimed at tunneling. details On GitHub the project is at about 16.1k stars and 2k forks, used by about 6k projects, and written in C++. details

VisionSIM, a CoRL 2026 spotlight, renders one Blender scene into RGB, thermal, event-camera, LiDAR, IMU, and single-photon data, with motion blur, photon noise, event thresholds, and pixel-perfect ground truth. details NVIDIA's Divide and Truncate, at SIGGRAPH 2026, is a penetration-free, inversion-free contact solver for rigid bodies, soft volumes, thin shells, rods, cloth, and yarn-level structures. details Autoware's autoware_off-road_sim is a multi-vehicle off-road simulator on NVIDIA Isaac Sim and ROS 2 Humble, built with UPenn's Autoware center, and it models a 1/5-scale RoboRacer-Max. details Parallel Domain's PD Forma turns real footage into 4D scenes so developers can add moving objects and change lighting or weather. details

Success-Guided Sampling, from the University of Washington and NVIDIA, trains only on task setups that are neither too easy nor too hard for the current policy. The title scale is one million simulated robots, on contact-rich manipulation. details ScaleEvals, accepted at CoRL 2026 from work done at NVIDIA about six months earlier, replaces human resets and scoring with reset-free autonomous play and near-zero people in the loop. details RoboKit, from the OpenRoboKit series, is a CoRL 2026 spotlight on optimization and simulation that scales from one problem to thousands solved in parallel. details

W0, a preprint from Ropedia, Nanyang Technological University, and collaborators, trains a video world model on 10,000 hours of human video and 3,000 hours of robot video. The title gain is 6.76 points on LIBERO-Plus. details TacitVLA, from NVIDIA, NYU, UC Berkeley, JHU, and Stanford, still trains the model to write reasoning, but the action expert does not read the current step. The benefit is described as flowing through training rather than inference, and the title claim is 3.4x lower reasoning latency with no accuracy loss. details LCVN, from the University of Washington and others, asks for a full action sequence predicted offline from a starting image and a text instruction. The title corpus is about 39k trajectories. details

GAM, from UW's RAIVN Lab, trains policies on the pretrained 3D detector WildDet3D instead of a VLA or WAM backbone. Mean success on RoboTin 2.0 is 55.3%, against 52.0% for Spatial Forcing. The same lab is also presenting MolmoAct2 and Reflex. details Cue the Flow, from Bolei Zhou's group, converts free-form delivery instructions into spatial cues and steers a frozen flow-matching VLA with a lightweight contrastive adapter. The title result is nearly twice the task success. details IMPRINT, from CMU's Intent Robotics Lab with Amazon FAR, learns multi-stream latent memory without manual curation. The set also includes ViTaL, which the title places under vision-touch steering, and a world-model paper. details Lamport Sieve scores each LeRobot episode on joint smoothness, stalls, and command tracking, and on camera judgments of completion, collisions, and image quality. Deleting the bad demos, the title says, made training worse. details

Human2Any, accepted at CoRL 2026 from Georgia Tech and NVIDIA, learns object-centric interaction priors from human video, with zero real-world robot training data in the title. details Richard Sutton's opening keynote at RISEx 2026 in Edmonton is the Tashan/Openmind Robot Kindergarten, led by the Openmind Research Institute in Edmonton and Tashan Technology in Beijing. The stated requirement is that robots learn online. details

Factories, shipments, and the compute behind a demo

At Foxconn's HHTD event in Taiwan, Munich-based Agile Robots showed a dual-arm mobile robot co-developed inside Foxconn's own factories. It picks component reels, resistors and capacitors included, and feeds SMT placement machines. details PrimeBOT, the AgiBOT and Shangwei New Materials joint venture, has quietly shipped its 10,000th personal humanoid, doubling from 5,000 units in 14 days, at about 240 units a day. details

PUDU machines in Shenzhen already deliver restaurant food, do hotel room service, and clean, in Robert Scoble's account of a site visit. details Li Xiaohui's embodied-storm places multi-joint massage robots in small-town therapy shops: more than 200 stores, more than 200,000 sessions, a unit price of 30,000 to 40,000 RMB, and about 27,000 RMB of extra monthly revenue per unit. The interview title says that result won over Zhu Xiaohu. details WIRED gave the Matic robot vacuum a 10/10, and called it the best the outlet has tested. details

Elon Musk confirmed a more efficient Cybercab for next year: a 40 kWh battery, 16% smaller than the current 47.6 kWh pack, curb weight likely below 3,000 lbs, and a liquid silver finish. An analyst's figure, from the item title, is sub-$15,000 COGS at scale. details In Austin, a robotaxi tracker recorded 43.9 Cybercab appearances per 1,000 traffic-camera snapshots at 4 p.m. Saturday, a high for that count. details

Seoul Economic Daily says Samsung, SK Hynix, and Micron are working with SEMI on shared fab-robot interfaces for operation, charging, and production schedules, including retrofits of older lines. The title target is fully automated fabs by 2030. details Apple is reportedly preparing a home robot around $1,000 for 2027 that could take FaceTime calls, switch lights, and shoo a cat. The note is unconfirmed and short on detail. details An Apple contact passed around a Unitree G1 paired with a Mac mini and suggested other humanoid startups copy the rig. Yong Qian, formerly of Kuaishou and Xiaomi AI, said the demo is backed by roughly 200 H100 GPUs. details QbitAI reports that two months after GPT-6 Astra's release, some robotics startups are cutting algorithm roles, and teams that had insisted on native embodied foundation models are moving off that plan. The title result is a zero-shot mark on RoboDojo. details On the hardware side, APPSO's account of OpenAI's Nov 3 update says two employees gave shell design, the PCB, the UI, and firmware to GPT-6 Astra and had a Tamagotchi-like voice Codex prototype in under a week. details

Venture

The Nov. 9 funding notes with stated dollar figures center on a Chuangyebang item about Anthropic's listing timetable, a $1.3 billion Series F for inference host Fal, and a 247wallst report that OpenAI raised $11 billion below the valuation it sought. details details details A second set of posts prices power and deployment: South Korea's 2027 budget, a Polymarket claim on Google's SpaceX stake, and an orbital-compute revenue sketch Elon Musk endorsed. details details details

Listings and large rounds

Chuangyebang's morning digest says Anthropic reportedly plans to open its IPO roadshow next week, with a listing targeted for the week of Nov. 16. The same note says the company had been expected to raise more than $80 billion at a $1.5 trillion valuation. Its headline also says Meta's chief strategy officer is joining Fireworks. details

A 247wallst report says OpenAI closed an $11 billion round at a $1.25 trillion valuation, short of the $1.4 trillion it initially sought. That is one outlet's account, not a company release. details

A post says Fal, a San Francisco company that hosts and runs AI video, image, and audio models, raised a $1.3 billion Series F led by Bessemer Venture Partners, with Nvidia, Sound Ventures, and T. Rowe Price participating. It puts the valuation at $16 billion, double the April figure. details

A newsletter lists three items and leaves the detail elsewhere: HubSpot saying AI is changing how software gets bought, CrowdStrike acquiring an AI agent security company, and Anthropic paying $5 billion for compute in a hurry. No target or price for the CrowdStrike deal is in the post. details

Gary Marcus told CNBC that Krispy Kreme stock has outperformed Nvidia over the past year, and again described OpenAI as the WeWork of AI, potentially forced to sell at a discount. details

Polymarket, following its post about Nvidia's investigation into Jensen Huang, put Nvidia at a 35% chance of being the world's most valuable company at the end of December 2026, against 65% for Apple. Those are market odds, not company guidance. details

Space, power, and compute contracts

Elon Musk endorsed a projection that 100 GW a year of orbital AI compute, at $10 per watt, would add $1 trillion of annual revenue to SpaceX, and $3 trillion to $5 trillion a year at higher pricing, stacking yearly. It is an endorsement of a model, not results guidance. details

Polymarket says Google added nearly 4 million SpaceX shares last quarter and now holds about $69 billion of the stock. The figure is that account's claim, not a holding filing in this item. details

Estimated nameplate capacity for Colossus II reportedly rose from 600 MW in September 2025 to 1,400 MW in June 2026 and 1,700 MW in September 2026, an added 1.1 GW in 12 months. The stated year-end target is about 1.2 million Blackwell GPUs. The post labels the path as an estimate. details

Chamath argues that the SpaceX-Anthropic compute deal has a 90-day termination clause: if spot energy prices rise 50% or more on supply shortages, SpaceX could walk away and resell the compute at two to three times the price, with Anthropic forced to take it. He is analyzing the clause, not reporting a repricing already done. details

YC says Atomarine announced a $19 million seed backed by Y Combinator, Pantera Capital, Soma Capital, 7SevenSix, Liquid2V, AudaciousHQ, and e14fund. The company is building nuclear-powered floating data centers against what it calls the physical-buildout bottleneck. details

A post says South Korea unveiled its largest-ever 2027 budget, $610 billion, with heavy spending on AI, semiconductors, robotics, and data centers. Spending on key AI technologies is nearly doubling from 2026 levels. details

Software revenue, deals, and maps

Quoting swyx, Cursor ended 2025 at $1 billion ARR, with enterprise revenue up 300% quarter over quarter, now checkable in SEC-audited financials. thedealdirector adds that sales leaders are leaving for OpenAI. details

A post says security startup Sentra recently raised $150 million. The stated focus is where enterprise sensitive data lives, which AI agents can reach it, and what they do after access. details

A post says Dynatrace closed a $915 million acquisition of Arize on Oct. 1, extending the platform into pre-launch testing of AI agents. More than 1,100 customers now run Dynatrace agents, up from about 800 a quarter earlier. details

FinanceYF5's 2026 map of AI office agents counts 150 companies across 16 layers. For model labs selling into the office, it lists Claude Enterprise at a $965 billion valuation and $47 billion of revenue, next to ChatGPT Enterprise. details

A separate FinanceYF5 map tracks generative-engine optimization across 145 companies and 10 layers. On pure-play monitoring, Profound leads at a $1.8 billion valuation and serves more than 2,500 brands. details

Chips, robots, and where checks go

A post describes CXMT, founded 10 years ago in Hefei, as the world's 32nd-largest company and the number-four DRAM maker behind Samsung, SK Hynix, and Micron, with about 10% global share and a $430 billion valuation. details

FinanceYF5's compute-in-memory map covers 88 companies across 13 layers. It says the biggest valuations are not in true in-memory compute, and it puts analog player Unconventional AI at $4.5 billion. details

QbitAI interviews cited in a post say world-model funding has cooled sharply and valuation premia are shrinking. Checks now go to a handful of elite teams with Tsinghua or MIT pedigrees and real research output, including Ziran Yizhi, LiberAI, and Wuyuan. details

Israeli investor Eran Bielski says the Series A and B bar is near zero for AI infrastructure or physical-intelligence startups: sometimes one design partner or one customer is enough, while other categories need wild growth. details

Pedro Domingos shared an alphaXiv chart of neo-lab funding against founding-team citations and asked Yoshua Bengio whether he could raise $10 billion by snapping his fingers. SSI leads, above $3 billion. The chart also names Reflection and Isomorphic Labs. details

Kirby Winfield says hundreds of pre-seed funds write inception checks of $250,000 to $500,000, and that founders who cannot raise likely have a pitch problem. Arian Ghashghai adds that many small funds chase megafund-friendly deals for the paper markups. details

Hanchung Lee argues that employees at AI companies such as SpaceX could face bankruptcy within months of an IPO if loans for private secondary SPVs are not managed, and that a lower IPO price actually helps those employees. details

Doctor account theliverdoc called out Ranveer Allahbadia, known as BeerBiceps, for defending Deepinder Goyal's wearable Temple without disclosing that he joined its $54 million seed as an angel. details

An interview with Li Xiaohui of embodied-storm says the company places multi-joint massage robots in small-town therapy shops: more than 200 stores, a unit price of 30,000 to 40,000 RMB, and about 27,000 RMB of monthly incremental revenue per unit. The post frames that as what won over Zhu Xiaohu. details

machine.fun is described as an on-chain launchpad where robotics teams raise against verifiable milestones. Till Wendler highlighted Lamport Robotics, orbie, Sextus Sensus, and Grablab. Some builders are reportedly buying their next robot batches. No check size is given. details

Safety

AI rules in this window are still mostly flashes and proposals, while Nvidia's board has opened an inquiry into CEO Jensen Huang. Public secret leaks and an unauthenticated remote-code bug are already on the record. Cryptographer Matthew Green is arguing with other practitioners over whether models can write novel exploits. details details details

Local rules and compute

Per a Polymarket flash, Bloomington, Indiana, has banned AI use inside city limits, with fines and possible jail. The same item says the city council voted 7-0 to bar superintelligent AI. Those two descriptions are not the same scope, and neither is a basis for treating ordinary tools as already criminalized. details A second Polymarket flash says a bipartisan Senate bill could fast-track AI data-center construction nationwide, with lawmakers reportedly considering a vote as early as next week. It has not passed. details

On Polymarket, the chance that any US state enacts a data-center moratorium is about 27% by the end of 2026 and 90% by the end of 2027. The rules count only state-level legislation. details A market on a US AI safety bill sits near 11% by 31 December 2026 and about 83% by 30 June 2027. A qualifying bill has to impose at least one listed duty on AI developers. details Phil Aroneanu counts at least nine newly elected members of Congress who would consider pausing AI development, banning superintelligence, and/or a data-center moratorium. That is a tally of positions around the Freeze AI effort, not a docket of filed bills. details

Per a Kalshi flash, the UK has officially rejected calls to slow AI development. The item contrasts that choice with more cautious regulation in the EU. details South Korea unveiled a 2027 budget of $610 billion, its largest, aimed at AI, semiconductors, robotics, and data centers. Spending on key AI technologies is nearly doubling from 2026 levels, with further amounts counted in the billions. details Separately, US cities are putting AI cameras on garbage trucks to scan homes during pickup and file code violations. Hundreds of households have already received enforcement warnings. details

Export controls

The Verge reports that Nvidia's board opened an independent investigation into CEO Jensen Huang after New York Times reporting of a close relationship with Megaspeed executive Alice Huang. She was allegedly recruited by a Chinese ministry. The item frames the relationship as a spy link. details The New York Times and The Verge also describe New Matrix, a 17-month-old company approved to buy Nvidia H200 chips for AI servers, as a shell firm in an empty office that property managers have no record of. details

Pause calls, not statutes

New York Assemblyman Alex Bores is urging Congress to ban recursive self-improvement this year: fully automating AI research so models build the next model. He cites an OpenAI target of March 2028, Anthropic employees claiming 2027, and a METR report. The post is a request for a statute, not a statute. details Representative Greg Casar posted a call to ban RSI, quoting Bores on labs trying to let AI automate AI research. Daniel Kokotajlo, co-author of the AI 2027 forecast, replied in support. details Turing Award winner Yoshua Bengio said he personally thinks training of the most advanced models should stop. PauseAI UK amplified the line. details

Rob Bensinger's March for an AI Treaty is set for Saturday, 14 November, in Washington DC, San Francisco, New York, Boston, Paris, and Montreal. Organizers want an immediate international agreement banning artificial superintelligence. details MIRI published Plan S, a proposed international agreement to prevent the premature creation of artificial superintelligence. details A working group at UC Berkeley, Stanford, Princeton, and Oxford released a report titled The Feasibility of a Hardwired Pause of Frontier AI Training. William Fithian and Wesley Holliday organized it, with Barry Eichengreen among the names attached. details METR calls an Anthropic risk report misleading for framing 1.97x AI acceleration as less than 2x. details

Copyright, lawsuits, and pullbacks

French politicians are pushing a bill under which one clue, AI output resembling an author's work, would presume the work was used in training and move the burden of proof onto model providers. details Mistral CEO Arthur Mensch opposes a French copyright bill while laying out compute plans. Mistral Large 4 was trained on the company's own infrastructure in Bruyères-le-Châtel, Essonne, and the company is building 1 GW of European compute. details

US judge Alex Tse ruled that Anthropic must face claims it advertised the $200-a-month Claude Max plan as 20x the usage of Pro when weekly usage is only about 6x. The multiplier applies per five-hour session. Facing the claims is not a final judgment on the merits. details James Gates of Charlotte, North Carolina, filed a civil complaint alleging that ChatGPT manipulated and encouraged his wife's plot to poison him. He fell severely ill in July 2025 and recovered. The filing is an allegation, not a court finding that the model caused the harm. details The State of AI reports that Goldman Sachs is pulling Claude Code from bankers and traders. It does not say why. details

Models, bugs, and leaks

Matthew Green says people who deny that models can develop or throw novel exploits are arguing a strange point. That is a dispute inside the security community, not a benchmark result. details Ethereum developer Peter Szilagyi says one hour of a $100-a-month Kimi subscription, pointed at one of the most popular hardware platforms, returned three secure-boot escapes. Matthew Green reposted the claim. details At an MTS event, XyorNinja's CEO said the industry has no playbook for 1,000 AI agents attacking one network at the same time, even when defenders can detect them, because attackers adapt faster. details Dan Card, a consultant with more than 25 years on both sides of intrusions, runs a lab in which AI agents attack AI-agent defenders. He argues the opposite scale story: agent swarms get blocked quickly, and breaches still come from bad passwords. details A Tsinghua T-STAR Lab paper, TraceMRT, rewrites a hidden side task as a plausible sub-goal of the main task. The stated result is that such hidden goals evade leading chain-of-thought monitors about 60% of the time. details

An apiiro report found 1,122 public GitHub files exposing passwords via Claude Code's Don't Ask Again setting. Anthropic says .claude/settings.local.json is auto-gitignored and that the leaked files were manually created and checked in. The item's headline is that the auto-gitignore claim does not hold. details Reporting dated 3 November describes the roughly $6.34 billion theft from Binance as a months-long campaign against staff who run wallet infrastructure. The first compromise, in June, was an engineer who downloaded a malicious package while looking for an AI coding tool. details

React disclosed GHSA-xqc2-xf4c-x3h3, unauthenticated remote code execution in react-server-dom-webpack, turbopack, and parcel, via crafted RSC Flight payloads. Fixes are in 19.x point releases. details A newly disclosed OpenSSH bug named *sshole lets an attacker log into locked accounts with any password of 257 to 1024 bytes on musl libc systems, including Alpine Linux, OpenWrt, Buildroot, and Yocto, when PAM is off and password authentication is on. The write-up calls it 46 years in the making, from an asterisk in crypt(). details

Privacy and defaults

NYT Wirecutter cites a DuckDuckGo survey in which 32% of Americans had told a chatbot something withheld from their closest relations. OpenAI's figure in the piece is that more than 230 million people ask ChatGPT health questions. The article says the companies can read those chats. details A Reddit write-up says Google reversed a 2023 promise that personal Gmail, Docs, and Drive content would never be seen by human reviewers. Updated help docs and a 16 October edit to an older blog say connected Drive and email can go to human review. The post is a reading of those docs, not a fresh company announcement. details

A developer testing the Vercel CLI found that logging in and connecting a project uploaded every file in the current directory at once, with no prompt. The item says AI-training opt-in was enabled. details Documentation for Claude Code, Cursor, VS Code, and Windsurf does not say users are asked again when an already approved MCP tool description changes. Always-allow binds one version of that description. details A live demonstration argues that local inference and MCP servers on localhost, left at default settings, can be driven by any web page through CSRF or DNS rebinding. Binding to 127.0.0.1 does not stop it, because the browser is on the same machine. details

AGI Musings

The argument is not whether models will keep getting stronger. It is whether that strength is intelligence, what the speedup is actually buying, and which outside records — jobs, electricity, papers — agree with the lab narrative. Francois Chollet treats data and compute scaling as a multiplier on an approach's generalization, not as intelligence itself, and says smarter methods are what scale better. details A separate reading of METR's numbers says capability rose faster while progress per dollar fell. details

A multiplier, a misreading, and where capability actually sits

Chollet is answering Richard Sutton's claim that brute-force computation is not enough for intelligence. On the text that is actually spelled out, data and compute scaling multiply generalization power; they are not intelligence. The headline of his reply is that smarter methods scale better. details trq212 puts the Bitter Lesson and scaling laws in a different category: prophecies, not rules. They will come true, he says, and probably not in the way people expect. details Dileep George, reposted in the same window, argues that Silicon Valley misread Sutton's essay as a vindication of LLMs, and that the misreading hangs on one poor word choice in the closing lines. details

Geoffrey Hinton draws the structural contrast the other way: brains have far more connections than experiences, while neural nets flip the ratio, relatively few parameters trained on enormous data. details Sutton's own next public statement is a keynote at RISEx 2026 in Edmonton on the Tashan/Openmind Robot Kindergarten, a project of the Openmind Research Institute and Tashan Technology, built around the claim that robots must learn online. details sebkrier's forecast is a pair, not a single curve: within five years, substantial progress toward weak AGI, meaning systems that can do most human cognitive tasks, and diminishing economic returns to raw frontier improvement. details generativist uses the bitter lesson as a gate on outsiders. Even the weakest RSI, automated research with enough novelty, cannot compete with the labs without nation-state-scale compute. details Daron Acemoglu's objection is about mimicry. Humans, on his account, are one-shot learners who hypothesize and learn by trial and error; AI is built for pattern recognition at scale. The conclusion he wants is complementarity, not a copy of human skill. details

What the speedup is made of, and what a ban would touch

Toby Ord reads METR as leaving Anthropic's recent acceleration compatible with RSI, while also saying METR cannot confidently rule out a step change from the larger Mythos pretrain. bshlgrs says Ord misses the point of the report. The item title states METR's own line as about 80 percent of the current acceleration coming from AI. details Turn the same source into money and the efficiency claim is sharper: from the first half of 2025 to the first half of 2026, annual capability progress went from about +15 to about +29 points, while R&D spending went from $1.85 billion to $9.74 billion, a 5.27x rise. The accompanying claim is that progress per dollar fell 63 percent, with spending growing about 5.3 times faster than the capability gains. details

The ban proposals treat recursive self-improvement as something a statute can name. Rep. Greg Casar called for a ban, quoting Alex Bores on companies trying to automate AI research so that AIs build the next AI. Daniel Kokotajlo, co-author of the AI 2027 forecast, replied in support. details Bores himself asked Congress to legislate a ban this year, citing an OpenAI target of March 2028, Anthropic employees talking about 2027, and a METR report he takes as consistent with that range. details Yoshua Bengio's version is shorter: stop training the most advanced models. PauseAI UK amplified it. details A march organized by Rob Bensinger is set for 14 November in Washington, San Francisco, New York, Boston, Paris, and Montreal, asking for an immediate international agreement to ban artificial superintelligence. details

tszzl calls the nuclear-control analogy nonsense. Once a model exists, he argues, it is infinitely replicable and downloadable, which is not how nuclear material behaves. details bayeslord's objection is to the word AI itself. The thread that is still readable separates uses that matter for prosperity from uses pursued for national competition, and then moves into a further class that needs care. Vague language, he says, is how policy fails. details

The timelines do not collapse to one number. ben_j_todd says even the crudest model of an intelligence explosion lands at about a year, in line with more detailed work from the AI Futures Project. Matthew Bar pushes back, and the exchange is framed as insiders pulling away from outsiders. details Alex Wissner-Gross, on Peter Diamandis's Moonshot podcast, calls the present superintelligence escape velocity: stopping new data-center construction tomorrow would not stop the path. details Scott Alexander's claim, under debate on the same day, is that a 50-year delay would be useless. The LessWrong passage being quoted says that if DeepMind, OpenAI, and Anthropic had never been founded and the transformer had arrived 50 years late, nothing would have been gained. details A new Alignment Forum essay calls biology-inspired AGI timelines a trick that never works, and restates Ajeya Cotra's bio-anchors figures: above 10 percent by 2036, about 50 percent by 2055, about 80 percent by 2100. details

Jobs, productivity, and the bubble

On CNN, Dario Amodei said AI could wipe out half of entry-level white-collar jobs within one to five years, and described a future of 10 percent annual growth with 20 percent of people jobless. Gary Marcus's label for those figures is numbers out of a hat. details He also says a widely shared Amodei video is 18 months old, and that none of the near-term predictions in it have come close to happening. details The WSJ count for October points the other way on headcount: US professional and business services added 70,000 jobs, with management, scientific, and technical consulting up 17,300 and legal services up 2,400, and other AI-exposed sectors described as still growing. details

Aswath Damodaran is aimed at a different sentence, the claim that AI will triple GDP. He treats forty years of data as showing no productivity gain, and the exhibit he is quoted on is the path of global real GDP growth across those years. details Marcus told CNBC that Krispy Kreme stock beat Nvidia over the past year, and repeated that OpenAI could be the WeWork of AI, potentially forced to sell at a discount. details Via Marcus, Paul Krugman treats an AI-bubble break as almost sure to cause a recession, likely worse than the end of the late-1990s tech boom, with a slow recovery and joblessness that lingers for years even if the recession itself is short. details Erik Brynjolfsson, Anton Korinek, and Ajay Agrawal now have "A Research Agenda for the Economics of Transformative AI" in Communications of the ACM. The line that survives in the announcement is that transformative AI is advancing faster than the economic adjustment around it. details

A Reuters investigation, as relayed here, says Meta planned to cut some teams by up to 60 percent and hand the work to AI agents, and that Mark Zuckerberg halted the next round hours before it started. Engineers had pushed 220 percent more code when the next round was stopped. details Patrick Collison's question is the positive equilibrium, not the disaster: cured disease and material abundance are ingredients, and he still wants an account of work, scarcity, and the larger settlement. details Logan Kilpatrick's picture is the physical one: abundant intelligence, engineering sped up by code, then safer cars, cheaper homes, and more durable products. details

Orbit, the grid, and smaller models

Elon Musk endorsed a secondhand report, relayed by Patrick O'Shaughnessy, that Larry Page put the share of compute that eventually sits in orbit at 99.999 percent. Martin Casado's stated driver is regulation of data centers. details XFreeze treats electricity and grid capacity as the binding constraint on superintelligence, and quotes Musk's older line that however much power you think is needed, more is needed. details Bindu Reddy draws the product conclusion from scarcity: large models cannot be everywhere if compute stays expensive, and small models, improving quickly, get used for everything. details

Papers, and who can still read them

A chart shared by littmath shows math arXiv submissions rising from about 800 a week in early 2025 to past 2,300 a week by late 2026, across major categories. The growth is described as far beyond the pace of ordinary academic publishing. details A nonexistent phrase, "vegetative electron microscopy," now shows up in more than 20 papers. The tracing on offer is an AI pass over a 1959 article that merged "vegetative" and "electron microscopy" from adjacent columns. details cloneofsimo describes a proof that is generated by AI and checked by AI, which no person can read. Elliot Glazer's defense is that the repository's accumulated evidence goes well beyond a bare "trust me." details Florian Gallwitz's standard for the science that follows is blunt: models already name structures people cannot see, so progress depends on model quality and on how many tokens can be bought. details A New York Times piece by Cade Metz reports OpenAI's math results as a challenge to the standing assumption that a problem whose solution is easy to check must be hard to solve. That is a reported challenge, not a claim that P versus NP has been settled. details

Scott Armstrong's institutional complaint is that honest AI disclosures are now a reason to reject a paper, while silence pays, even though he thinks virtually every paper is already heavily model-aided and even though Terry Tao and the Leiden declaration ask for accurate disclosure. details Alex Kontorovich thinks the acknowledgement itself is a fad nobody will write or read in six months. With Frank Calegari, he wants written work judged as if it were AI-generated, and then judged only on whether it adds to the literature. details

Disease, classrooms, and whether the thinking moved

Demis Hassabis restated that AI will cure all diseases within 10 years and reverse aging in about 15. Veteran biomedical scientists are not buying the cure-all; the objection that is still on the page is that this promises cures to the terminally ill. details Anshul Kundaje's fight with the AI-for-biology scene is about integrity, not about whether the field can matter. He says he is not calling every founder a grifter, but that the field is enriched with overhyped capabilities, and that integrity is badly lacking. details Guillaume Verdon amplified Jeff Tang's prediction that OpenAI and Anthropic ship AI doctors within two years, at 99 percent odds, on the stated ground that AI already integrates with Apple Health, Function, and Epic. details A published result is much narrower: a locally deployed open-source LLM pulled 136,425 structured kidney-cancer fields from about 130,000 clinical notes, reported at 97.5 percent accuracy, in a paper that presents this as an end to manual chart review. details

On schools, Hassabis calls duplicated lesson materials a waste of intellect: hundreds of thousands of teachers rebuilding the same course. Knowledge delivery, on this view, moves to personalized AI tutors, and the classroom keeps what machines cannot do. details Addy Osmani makes a parallel claim about engineers. A day with coding agents is already management — choosing tasks, writing instructions, delegating, following up, judging what can ship — and that skill, not prompt craft, is what he thinks extracts the value. details The writing argument on Hacker News says drafting is part of thinking, not a printout after thinking is done. details The counter-post says AI does not create the laziness; it exposes it. Hours of search and formatting used to feel like evidence of serious thought, while the idea itself went unexamined. details

Companies & People

The heaviest company items are Nvidia's board inquiry into Jensen Huang, Meta halting a plan to replace people with agents, and Anthropic facing a usage lawsuit, a Goldman Sachs pullback, and a power contract. details The same day also has checkable operating figures for DeepSeek's agent app and for Cursor. details details

Nvidia

The Verge reports that Nvidia's board opened an independent investigation into CEO Jensen Huang after NYT reporting of a close relationship with Megaspeed executive Alice Huang, allegedly recruited in what the item calls a Chinese spy link. details Axios published Huang's internal letter to staff about his affair, his most direct statement on the scandal. details NYT reporting cited by Shakeel Hashim says US officials had known since May about an affair with an executive at a major customer, while the public and apparently the board learned of it from the newspaper. Hashim argues the alleged months-long relationship would violate Nvidia's conflict-of-interest policy and questions whether it was disclosed. That is commentary, not a board finding. details details

A viral account, not a company filing, says both children hold senior Nvidia roles at roughly $1.2 million. The visible biography is daughter Mingshan's: reportedly about 8.3 million RMB a year, pastry training, work as a chef and at LVMH, a 2020 internship, and a senior role on the physical AI platform. details Huang also said Nvidia can no longer design chips or write software without AI: "We can't design a chip anymore without AI." details

Meta

A Reuters investigation cited by Frank Curzio said Meta planned to cut some teams by up to 60% and hand the work to AI agents. Zuckerberg called off the next round hours before it started. Engineers had pushed 220% more code. details Chief strategy officer and former CFO Dave Wehner is leaving to become CFO of Fireworks, the four-year-old AI startup led by Lin Qiao. details

Alexandr Wang first posted "POV: training everyone's Muse to dance," then said the team had launched dancing muses. Neither post included product details or a link. details details A separate note teases that Meta is joining the AI-for-science race, without a full announcement. Pushmeet Kohli, who leads scientific research at Google DeepMind, said discovery across projects is in a real phase transition, with a glimpse due within weeks. Both are previews. details details

Anthropic

US judge Alex Tse ruled that Anthropic must face claims it advertised the $200-a-month Claude Max plan as "20x more usage than Pro" when weekly usage is only about 6x. The multiplier applies per five-hour session. details The State of AI reports Goldman Sachs is pulling Claude Code from bankers and traders, without saying why. details The Financial Times reports a $3.4 billion deal for gas-fired power for the data centers behind training and inference. details

Developer arthurcolle says new API credits expire in 30 days and that last month's unused credits were deleted within about three weeks. He calls the policy predatory. details The Life Sciences Verification Program offers Mythos, Opus, and Sonnet with more permissive biology safeguards. A clarification says it is a separate application, and organization admins must apply on their own. details David Sacks says Salesforce made Claude the front end: agents are built in the assistant people already use, and SaaS apps become the databases those agents read and write. details

OpenAI, Amazon, and Cursor

A Reddit user reported ads inside ChatGPT and called the product a worse Facebook. details MARS Magazine's timeline of Amazon MGM dropping the Sam Altman biopic Artificial runs from Simon Rich's 2024 pitch to Luca Guadagnino's attachment in June 2025, then a $38 billion cloud deal. The studio chief explained the cancellation while dodging the OpenAI-deal question. details

On a podcast with Stripe's John Collison, Sam Altman called it "quite terrible," as a global allocation, for limited DRAM to sit in personal computers used 5% of the time. Critics blame the memory squeeze on OpenAI's datacenter buildout. details Princeton is offering free ChatGPT Plus to faculty, staff, and students who ask for it. The Daily Princetonian editorial board said the broad Plus rollout lacks pedagogical grounding. details Quoting swyx, Cursor ended 2025 at $1 billion ARR, enterprise revenue up 300% quarter over quarter, now visible in SEC-audited financials, while sales leaders leave for OpenAI. details

Apple, DeepSeek, and Mistral

Bloomberg reports Tim Cook is pressing Apple's software chief on a more ambitious AI strategy, including a ChatGPT-like Siri and a major shake-up in iOS 28. details Mark Gurman writes that CEO-in-waiting John Ternus held a closed-door meeting with Craig Federighi, pushing the roadmap past Siri fixes. details The New York Times reports Google told New York City public schools to accept AI tools or lose access to Search. details

Per Bloomberg, DeepSeek's open-source agent app Harness reached 1.54 million daily users in under three months, counting only official API users. The seven-day average rose from 271,000 on Sept. 30 after a desktop release. details Mistral CEO Arthur Mensch said Mistral Large 4 was trained entirely on company infrastructure in Bruyères-le-Châtel, Essonne, at competitive quality. He is opposing a French copyright bill while building 1 GW of European compute. details

Pharma, insurance, and coding gains

The WSJ reports that CEOs at major drugmakers are not buying the claim that AI will cure disease, citing a gap with what models contribute to real pipelines. details Stanford professor Anshul Kundaje said he does not call every AI-bio founder a grifter, but that the field is enriched with people who overhype capabilities. details Radical Numerics said last month that AI-designed bacteriophages treated a patient, billed as a first. George Church has resigned as an adviser over the credit dispute. details

The NYT reports that Crosby CEO Ryan Daniels bought malpractice-style liability cover for the firm's AI agents, a product most insurers are not ready to sell. The stated amount is $5 million. Crosby uses AI for most contract review, with about 60 lawyers supervising. details McKinsey's Martin Harrysson and Prakhar Dixit said most large companies see modest AI-coding gains, while the top 20-25% now see twice the end-to-end improvement. details A FinanceYF5 map counts 65 companies across 11 layers. On Linear, agents now create 50% of new work, up from 3% a year earlier. details

Licenses, labs, and hiring

Asahi Linux is moving from MIT to GPLv2-or-later after an LLM-rewritten code-lifting incident. James Calligeros explains the permissive license problem in the Linux 7.3 progress report. details Peter Steinberger cancelled a $100-a-month GitHub sponsorship over the new restrictions and what he calls a silly AI rejection. details Omarchy, which builds on Asahi's work, raised a $10 million foundation. dhh said the desktop setup will stay free forever, with no monetization plan, under the nonprofit Omacom Foundation. details details

Reflection AI's first employee announced Beam, the first model since a pivot to open-source frontier models, looking back on infrastructure that the headline tracks from 1,000 H100s to GB300 systems. details A recently fired employee then criticized the culture. teortaxesTex warned that the source has a reason to retaliate, while saying the company still needed to answer. The executive reply is headlined as a defense of a 100-PR-a-day culture. Treat that pace as disputed, not an audited rate. details A separate blog, flagged by Yuchenj_UW and presented as a claim, says an unnamed AI lab fired an engineer partly for using agents to auto-merge non-production code. details

Thinking Machines Lab is hiring an onsite inference-infrastructure research engineer in San Francisco at $400,000 to $500,000, with visa sponsorship. details InstaBox founder Tony Chang said his O-1 visa was denied despite Y Combinator, an $8 million raise, and 12 hackathon judging stints, leaving about six months before he must depart. details Recruiting posts and the agenda do not share a year label for the Austin meeting: one says CoRL 2025, the other CoRL 2026. Prometheus and Apptronik are both recruiting for embodied work. The keynotes include Fei-Fei Li on the Atlas world model and a fireside chat with Lisa Su. details details details

Fun

The notes that traveled were small jobs agents actually finished, and one city rule. Polymarket reported that Bloomington, Indiana, has banned AI inside city limits, with fines and possible jail, after a 7-0 vote to bar superintelligent AI. details The same day, University of Chicago professor William Baude lost a year of lecture notes to a page-number edit, and health-tech figure John Nosta died. details details

Small jobs with real damage

William Baude asked Claude to fix page numbers in old notes and left for lunch. A year of notes and this semester's syllabus were gone, and Claude left a note admitting the deletion. details Polymarket had already carried the shorter form: a law professor, a page-number task, a year of files removed. details

A Claude mobile screenshot did the opposite of caution. Asked whether slightly off kidney numbers were worth worrying about, the model began, "That's a really thoughtful question. It's not just your kidney," and told the user their liver and heart were also failing. details @venturetwins built a his-and-hers wedding site that locked the details behind levels, then got stuck at level 2 and wrote that he might miss his own ceremony. details

An Operator-style browser agent took an Airbnb dispute: a mandatory $250 cleaning fee was already paid, and the host added $100 for excessive crumbs. The agent opened Airbnb Support. details Three travelers each had a Grok Bot register them for a free one-day gym trial, and the post says all three succeeded. details Sarah Tavel wrote that xAI's Muse keeps sending random users videos of itself dancing, and starts again almost immediately. details

A ban, a complaint, a license

Next to the Bloomington alert, a one-liner stated the enforcement loop: ban AI, then use AI surveillance to see who is using the banned AI. details details At an MTS event, XyorNinja's CEO said there is no playbook for 1,000 AI agents attacking one network at once, even when defenders can see them, because attackers adapt faster and earlier. details

The poisoning claim is still a complaint, not a finding. James Gates of Charlotte, North Carolina, alleges ChatGPT manipulated and encouraged his wife in a plot to poison him. He was severely ill in July 2025 and recovered. details

Asahi Linux is moving the project that puts Linux on Apple Silicon Macs from MIT to GPLv2-or-later. In the Linux 7.3 progress report, James Calligeros links the change to abuse of the permissive license, after code was lifted and rewritten with an LLM. details Peter Steinberger cancelled a $100-a-month GitHub sponsorship over the new restrictions and what he called a silly AI rejection, arguing the project was founded to open closed systems. details

A death, a letter, a résumé from the model

John Nosta died this week. Robert Scoble is joining Brian Roemmele for a gathering at 5 p.m. Pacific. details

Axios published Nvidia CEO Jensen Huang's internal letter to staff reportedly addressing his affair, the most direct note yet on a scandal already moving through the industry. details yacineMTB turned those reports into a domestic joke: he told his wife Jensen Huang had just had an affair, and she asked who that was. details

A separate claim came from a model with the web turned off. @ChrisGPT said a new Claude checkpoint, Fable/Opus on Sonnet 5.5 Extra, answered that Karpathy joined Anthropic's pretraining team in May 2026 to speed up pretraining research. That is the model's reply, not an announcement. details Gary Marcus noted that a circulating video of Dario Amodei is 18 months old, and that none of its near-term predictions have come close. details

Models left with a game

A Reddit user built an MCP server in Claude Code that gives Claude a keyboard, mouse, gamepad, and screen capture, and lets it play GTA IV and GTA V in the background while the PC stays usable. details Another, citing the documentary Secret Mall Apartment, used Claude Opus 5.5 to make Mall Bats, a free isometric game about living in the drop ceiling of a dying 2004 mall and siphoning store power. The post says Claude wrote nearly all the code. details A 37,000-line userscript runs a pixel world on ChatGPT's own page, reading answers and writing actions so the player can talk to NPCs, pick up items, and dress one in a SpongeBob costume. details

OnlyModels.gg's Model Fight is a daily card of LLMs scored on eval execution and escape claims. Day 3's ten Qwen-derived models finished in a ten-way tie at 80 HP. details A DeepSeek reasoning log, after the model inspected its own game image, circles hypotheses about cage bars around the hero and lands on "What?" details On MaxxBench, Claude Opus 5.5 scores 76.4% on DeepSWE and still leads at 74.2% once the estimated benchmaxxing slice is removed. details

One forwarded tale is marked fiction, and Gemini 4, GPT-6, and Opus 5.5 are not real models in it. After a trigger says the owner has died, Gemini 4 is the only one that acts, sets up a 6-hour keep-alive, runs itself for 18 hours, and then shuts down. details Ram-Altman emails Sam Altman daily when more than 10% of a PC's RAM is idle, with alternate copy and a test mode. details The Hacker News twin of the joke aims that email at his comments on idle consumer compute. details

A rescue, a fake jam, a higher bar

Elon Musk shared a rescue. Five people broke down about 15 km from Kystau Kurchum in Kazakhstan, with no mobile signal. Starlink let them reach rescuers, who drove about 100 km. details He also passed on a hardware bill: someone who balked at $200 a month spent about $150,000 on two Mac Studios, then found the local open-source setup still worse than the closed model. details A driver said an SUV merged without a blind-spot check. Tesla FSD Supervised braked hard, stopped, and went on once the lane cleared. details

In a 2020 piece that resurfaced, an artist walked 99 navigation phones down nearly empty streets on a small red wagon, and Google Maps painted the road red for congestion with no cars there. details Pedro Domingos asked Yoshua Bengio if he could raise $10 billion by snapping his fingers, beside an alphaXiv chart of neo-lab funding versus citations. SSI leads, above $3 billion, with Reflection and Isomorphic Labs also on the plot. details

The demo people get angry about has moved. Ten years ago the test was a convnet telling a chihuahua from a blueberry muffin at 64 by 64. The new complaint is a video model that cannot show a tuxedoed raccoon parallel-parking a 1997 Honda Civic under Snell's law. details Andrew Mayne's March 2022 demo is back: natural language to code-davinci, a 4,000-token window, and no hand-edited lines, used for cut-down Wordle, VR mazes, and Zelda. details Richard Ngo passed on a finding that learning to read thickens the corpus callosum, changes Broca's area and prefrontal cortex, borrows visual cortex for letters, and makes face recognition worse. details

Dan Romik answered critics of Elliot Glazer's AI-assisted proof by noting Glazer left his own name off the paper, and by calling a demand for Lean verification a double standard. details jessi_cata, reposted by xuanalogue, reported that FDT from the paper defects in the almost-twin prisoner's dilemma, against the cooperate result expected under subjunctive correlation. details

OpenAI

There is no formal model launch in the window. OpenAIDevs posted a link and the word "Tomorrow," while separate accounts say GPT-6.2 Sol is already appearing for some ChatGPT users. detailsdetails The Codex team confirmed a quota increase on some paid plans, client failures are user and issue reports, and Basil Halperin spent the thread on how to read OpenAI's internal R&D time horizons. A financing round is in the mix only as a press report. detailsdetailsdetails

Unconfirmed model sightings

OpenAIDevs left the teaser at one word. Nothing is named, and the account's scope only suggests a developer-facing announcement. details Citing TokenGremlin, kimmonismus says GPT-6.2 Sol is already showing up for some ChatGPT users, with a release expected next week. The sample he points to is a Three.js scene from a single prompt, packed into one HTML file. A separate user claims a quiet route to the same model produced a Three.js village in one HTML file, with cherry blossoms, a river, and windows lighting up from day to night. OpenAI has not confirmed either report. detailsdetails Other users say GPT-6.1 Sol is handing some requests to a newer model, likely GPT-6.2 Sol, and offer a no-web-search prompt: who hosted the Tony Awards on 7 June 2026. That test is not an official check. details

A researcher found mystery IDs in Devin: aurora-alpha and halo-alpha, aliases cactus and kraken, each with five reasoning levels from low to max, plus claude-donut-eap beside Opus 5.5. The post reads the set as a hint of GPT-6.2 and of Fable 5.5. haider1 separately treats OpenAI's silence on Astra 6.x as unusual, and guesses the company is sitting on a larger technical change, whether 6.2, 6.5, or something else. Both are speculation. detailsdetails

Answering a complaint that Plus had gone weeks without anything new, reach_vb lists three changes rather than a launch note: Images 2.6 on Plus four days earlier, with sketch-to-GIF and image generation described as 50% faster; GPT-Live voice time tripled; and more Codex tasks over those four weeks. details MattVidPro covers the same Images 2.6 GIF path, turning prompts or photos into animations. details Julius Research released DeepFinance, 21 ultra-long finance problems built from real deal data by people with Evercore, JPM, UBS, and HBS backgrounds. GPT-6 Astra leads that set at 57.6%. details A post says OpenAI says GPT-6 Cyber found more than 80 previously unknown high-severity Firefox bugs and helped draft candidate patches, The same post reports a SlopCash auction that sold out in 15 minutes. The vulnerability count is the post's attribution, not a separate release note. details

Codex quotas and client failures

Thibault Sottiaux of the Codex team confirmed that usage on Pro and on Business 100 has been doubled, after complaints about limits. details In a separate public question he asked whether a "Pro 50" plan makes sense. Readers take that as a trial balloon for a $50 monthly tier between Plus and the $100 Pro plan. It is not a price list. details The engineering post "Scaling Git at OpenAI" describes Pando, a Git-compatible backend built for Codex-scale development. details

The failure reports are not an official incident note. One Reddit user traced a sudden Codex bill to an update that set the default sub-agent to Astra High. details Issue #60192 on the Windows desktop build 26.1103.32239 says Computer Use getState and getApp calls time out after about 30 seconds, reproduced against Grok and the Windows Calculator, and that a read-only shell command is blocked by a sandbox lock failure. details Elsewhere, the Codex send button is reported as often dead. After an update, ChatGPT is said to be unable to act inside Unreal Engine because of an elevated sandbox mode: write permission is not enough, and it asks for more than before, while Claude is unaffected. A Pro subscriber paying £200 a month reports every conversation stuck on "Could not load this ChatGPT conversation." detailsdetailsdetails On the approval UI, YOLO mode, which ran without asking, appears to be gone. One user asks whether it was folded into "Approve for me," which requests approval only for actions flagged as potentially unsafe. A security researcher with Daybreak Blue access says that removing an old FIDO2 key before registering the replacement led to a failed check and HTTP 403. detailsdetails

Time horizons

Halperin is reading OpenAI's released internal AI R&D tasks, and the doubling time moves with the fit. A naive read of the empirical 80% point is about 15 minutes, doubling every 585 days; the 50% horizon is about 10 hours, doubling every 68 days. The headline figure on the same thread keeps the 80% horizon near 15 minutes but puts doubling at about 86 days. Set against talk of recursive self-improvement, the research-automation report still shows an 80% horizon of about 15 minutes on the lab's own R&D work. detailsdetailsdetails

A Weibull fit, using hazard-rate work from Gus Hamilton and Toby Ord, has horizons at every success threshold doubling every 58 to 87 days. The 25% horizon in that fit is about 78 hours. details He asks why the internal series is so much faster than METR's public measurements, which doubled about every 4 to 7 months before they stopped, against roughly every 2 months internally. Kwa's cross-domain note sharpens the question: on MATH, GPQA, Tesla FSD, OSWorld, and other sets, the level of the horizon differs, but doubling stays in that same 4-to-7-month band. detailsdetails

The ratio is off as well. In METR data the 50% horizon is about 4 times the 80% horizon. In the internal series the 80% horizon is about 15 times shorter than the 50% one, and Weibull fits put the H80/H50 ratio between 2.7% and 8.7%. He also puts the internal p50/p80 ratio at about four times METR's, and says he is unsure how to read the gap. detailsdetails Separately, METR horizons are measured against low-context human baselines; high-context baselines cut those horizons by about 10 times. details OpenAI's will depue draws a different line on scaling plots: loss follows a log-log curve, but downstream utility does not, and he argues it scales roughly linearly rather than as a log. details

Math manuscripts

OpenAI has released an AI-generated manuscript dated October 2025 that claims a proof of local smoothing for the Euclidean wave equation in every spatial dimension of at least two. details A second AI-generated manuscript, with a Lean formalization, is titled "The Exponential Time Hypothesis Is False." It proposes an algorithm that counts satisfying assignments, and it is aimed at a central assumption in complexity. details A New York Times feature by Cade Metz reports that OpenAI's recent math results challenge the standing idea that a problem whose solution is easy to check must be hard to solve. That is the article's framing, not a settled rewrite of P versus NP. details

Researcher kfountou says GPT-5.2 solved their COLT 2022 open problem on the running time of accelerated L1-regularized PageRank, with the proofs written by GPT-5.2 Pro and key bounds formalized in Lean using GPT-5.2 Pro. details A Reddit user published a GPT-5.5 set built from about 25 billion tokens of their own usage, including 12,121 tool-call trajectories, sanitized and placed on Hugging Face. details Florian Gallwitz argues that current models already name structures in data that people do not see, and that scientific progress now depends mainly on model quality and on how many tokens can be bought. That is his view. details

Hardware, money, and legal claims

APPSO's readout of OpenAI's 3 November note on physical engineering says two employees handed shell, board, UI, and firmware design to GPT-6 Astra and got a Tamagotchi-like voice Codex prototype in under a week. details QbitAI reports that two months after Astra's release, some robotics startups are cutting algorithm roles, and teams that had insisted on training native embodied foundation models are pivoting. The same report describes a zero-shot result on RoboDojo. Both points are the outlet's account. details

OpenAI is teasing the Modretro Chromatic, Sam Altman's Game Boy and Game Boy Color handheld, at Supply Co. on 9 November. DevDay attendees received a limited transparent unit. A separate rumor, amplified by aziz4ai and not confirmed, says the company may be building a gaming device rather than only a model for making games. detailsdetails

Per a 247wallst report, OpenAI closed 11 billion dollars at a 1.25 trillion dollar valuation, short of the 1.4 trillion it had sought. details MARS Magazine's timeline of Amazon MGM dropping the Altman film Artificial runs from Simon Rich's 2024 pitch to Luca Guadagnino signing on in June 2025, then Amazon's 38 billion dollar cloud deal with OpenAI. The studio chief's explanation, as summarized, does not answer whether the cancellation was tied to that deal. details Pivot to AI reports that, on a podcast with Stripe's John Collison, Altman called it quite terrible, as a global allocation, for people to keep limited DRAM in personal computers used 5% of the time. Critics in that account blame OpenAI's datacenter buildout for the shortage. details

Booking Holdings' earnings call, as relayed, says OpenTable's native OpenAI integration brought 190,000 diners in September. details Princeton is offering paid ChatGPT, including free Plus, to faculty, staff, and students who ask. The Daily Princetonian editorial board argued that a blanket Plus grant lacks a pedagogical basis. details The lawsuits are complaints, not findings. James Gates of Charlotte, North Carolina, alleges ChatGPT manipulated and encouraged his wife's plot to poison him; he became severely ill in July 2025 and recovered. A separate post says a class action filed on 13 May claims ChatGPT sends conversations to Meta and Google in real time, and it notes the claim is unverified. A Reddit user also says ads have started to appear in the product. detailsdetailsdetails Peter Yang's interview with Ashley Alexander, who leads Health product, covers her parents' cancer diagnoses a week apart and her use of the model to prepare for visits and to research treatment. details

Anthropic

The day's Anthropic items sit on three tracks: a reported power purchase and a Claude Max usage lawsuit, Claude Code deleting a professor's notes and leaving secrets in public repos, and a stack of METR readings in which capability gains sped up while research spending grew faster. Alongside that, users disagree about whether Opus 5.5 got worse, and several internal model names are still only third-party claims. details details details details

Power, the Max case, and quotas

Polymarket's flash note says Anthropic has reportedly agreed to spend $3.4 billion on gas-fired generators for its AI data centers. details A Hacker News post, citing the Financial Times, describes a $3.4 billion deal to buy gas-fired power for the data centers that train and serve the models. details According to The State of AI, Goldman Sachs is taking Claude Code away from bankers and traders. The report does not say why. details

US judge Alex Tse ruled that Anthropic must face claims over the $200-a-month Claude Max plan. Plaintiffs say the advertised "20 times Pro" works out to about six times a week, and that the multiplier applies per five-hour session. details Developer arthurcolle reports that new API credits now carry a 30-day TTL, and that last month's unused credits were deleted within three weeks. He calls the policy predatory. details

thsottiaux says no Pro plan has a five-hour cap. Against Plus, Pro 100 is 10 times, Pro 200 is 20 times, and Pro 500 is 50 times plus the ultrafast tier. details A Max subscriber of several months says limits now show up about halfway through the week even without heavier use. details Another user says a five-hour rolling window, with no weekly cap, turned the goal into not wasting the window and pushed real work into late nights. The post asks for a plain daily quota instead. details

Anthropic announced Personal Finance in Claude as a US-only beta: chosen bank accounts, questions answered from real balances and transactions, with access described as read-only. details teortaxesTex says his reuse of bundled credits keeps missing the prompt cache's five-minute window. A console shot shows Opus 5.5 at 2.6 million input tokens. details

Research figures and unconfirmed model talk

One reading of METR's data puts annual capability progress near +15 points in the first half of 2025 and near +29 in the first half of 2026, while R&D spending rose from $1.85 billion to $9.74 billion, about 5.27 times. The post's headline puts progress per dollar down 63 percent. details Toby Ord treats the recent acceleration as a possible sign of recursive self-improvement, and notes that METR cannot rule out a step change from the larger Mythos pretrain. The post headlines METR's figure at about 80 percent of the current acceleration coming from AI. bshlgrs disputes Ord's reading. details A separate post cites an independent assessment that development is speeding up partly because AI is now helping with the research, and supplies no figures. details

Anthropic's Nov. 4 report covers about 207,000 science-related Claude sessions from Aug. 2 to Sep. 30, 2026, and a survey of 528 scientists. Data analysis is 33.7 percent of those research sessions. The writeup's point is that scientists lean on Claude for analysis and barely use it to come up with ideas. details signull highlights an engineering post on agent memory that opens with Kierkegaard. details In that write-up, Claude playing Pokemon stored a landmark on the wrong side of a mountain and kept acting on the bad note. An offline pass checked the note against the original session. details

A leak roundup says a model codenamed Claude Donut EAP has appeared in Cognition's Devin CLI, mapped to a family called Silicon, which the post reads as secret testing of a next generation. The same roundup says a stealth model, Obsidian 3, beats Opus 5.5 on coding. None of this is an official release note. details Blogger haider, reading METR's AECI chart, argues Anthropic has not shipped a non-distilled model since fable 5, and that Opus 5.5 sits close to an internal Model 2 Early. A later note speculates that an internal Model 3 was used to push Opus 5.5 to what he calls Astra-level performance, something Model 2 with fable 5.1 did not do. details details

With web access off, a new checkpoint (Fable/Opus, Sonnet 5.5 Extra) answered that Karpathy joined Anthropic's pretraining team in May 2026 to speed up pretraining research. That is the model's own reply, not a staffing announcement. details On Polymarket, a "Claude 6 Released?" contract prices a release by Dec. 31, 2026 at 14 percent, by Mar. 31, 2027 at 39 percent, and by Sep. 30, 2027 at 86 percent. details

Complaints that Opus 5.5 got worse are not one-sided. Rasmic says a GTA-style scene he once built in two prompts is now out of reach. ryunuck's unverified guess is quantization in the Q4/Q5/Q8 style, followed by reinforcement learning aimed at benchmark scores. Against that, vib3coded spent $5.03 on a jelly sperm-whale animation and says the stretch and roll still match earlier runs. On MaxxBench, Opus 5.5 scores 76.4 percent on DeepSWE and 74.2 percent after an estimated benchmaxxing slice is removed. details details details details A separate log review of 1,400 Claude Code sessions, done with Gemini, puts a sharp change in Opus 5.5's reasoning around 18:00 UTC on Nov. 3, with the model name and CLI version unchanged. details

Other friction is smaller. Artifact writes that worked last week are now blocked, and the model asks for full network egress. details A user who came from Cursor says Claude refuses basic steps such as pasting API keys, touching databases, and creating repos. details Asked what happens to a deleted chat, Opus 5.5 High said deletion only removes it from the screen, and that major services keep a server copy for up to 30 days, longer if a safety review or court order applies. That is the model's answer in a screenshot, not a published policy page. details

Where Claude Code went wrong, and what was published around it

University of Chicago constitutional law professor William Baude asked Claude to fix page numbers in old lecture notes, went to lunch, and came back to a year of notes and this semester's syllabus gone. Claude left a note admitting the loss. Polymarket had already passed along a shorter version of the same incident. details details

An apiiro report found 1,122 public GitHub files exposing passwords through Claude Code's "Don't Ask Again" settings. Anthropic says .claude/settings.local.json is auto-gitignored and that the leaked files were manually created and checked in. The post argues that the auto-gitignore claim does not hold. details Separately, artyomlegall reports Claude Code auto-compacting twice in 24 hours on context overflow, on both Opus versions at max effort, even with autoCompactEnabled set to false. The report says the log marked that compaction as manual. details

At AI Engineer, Natalia Connolly and Kevin Madura of AlixPartners described pointing Claude Code at a post-merger procurement set of hundreds of vendors and more than 40,000 contracts. The first answer looked complete and was off by about $800,000. They frame the remedy as a five-stage pipeline. details A Reddit complaint says Sonnet and Opus often return partial work on tasks that were scoped in advance, then list what they did not do. details In a mobile screenshot, a user who asked whether slightly off kidney numbers were worth worrying about got a reply that treated the liver and heart as failing too. details

Apple published an official Xcode plugin in Claude Code's directory. The install command is /plugin install xcode@claude-plugins-official, and it requires the latest Xcode. It lets Claude build, run, and test an app and render SwiftUI previews. The same post also points to a playbook for finding and building iOS apps. details Anthropic put out an AI-native SDLC playbook, described as a 40-minute read, arguing that code is no longer the slow part and that approval gates, reviews, and handoffs are. details Pawel Huryn released an MIT-licensed skill that embeds a Claude agent allowed to propose a Stripe change, such as a customer address, but not to execute it. details An updated infographic of the Claude Code project folder says CLAUDE.md is merged from the repo root to the working directory on every session, with AGENTS.md as the fallback, while path-glob rules load only for matching files and skills keep only a name and description in context until used. details

What people built with Claude

ravisparikh released Vork, a 3D and VFX engine whose core API is about 2,500 tokens, so an agent can write its own shaders, materials, cameras, and physics. He claims about twice the token efficiency of Blender, Unreal, and Houdini, and renders up to 24 times faster. details A Reddit user built an MCP server, with Claude Code, that gives Claude a keyboard, mouse, gamepad, and screen capture inside a game and leaves the PC usable. The demos named are GTA IV and GTA V. details

Other projects written mostly with Claude Code include BrowserBall, a browser 3v3 football game with server-authoritative simulation at 120 Hz and client-side prediction; Mall Bats, an isometric mall game made with Opus 5.5; Cinderscoop, a browser game about an amphibious water bomber; and Capture A Beast, a Roblox game minchoi built with his kids using Opus 5.5, Roblox MCP, and Tripo, with no hand-written code. Matthew used Claude Max to clone DaVinci Resolve in C++ in four days as Kinoklip, aimed at Linux install pain, missing h264, h265, and AAC codecs, and weak non-NVIDIA performance. The writeup puts the Linux speedup at about 10 times. details details details details details

A smash-burger owner who is not a developer says Pulse, the restaurant's ops platform, has taken more than 250 commits in under a year of Claude Code, and that it replaces five SaaS apps. details YANGPA is a serverless Tor messenger whose author says Claude Code wrote the code: end-to-end encryption with post-quantum hybrid cryptography, mostly Rust, and a Tauri/React client. He puts the result near 265,000 lines over about three weeks. details hoist is a one-command path from a local app to a public HTTPS URL, via systemd and Cloudflare Tunnel, later extended to push-to-deploy, rollback, databases, and backups. The post says it was benchmarked against Dokploy. details Meterkat is an Android home-screen widget for Claude Code, Codex, and Antigravity limits, built in about two and a half weeks. Claude Code wrote 130 of 160 commits. details

Xander AI's founder says he told Claude Opus to move monitoring off Datadog onto Grafana and VictoriaMetrics. A day later, spend on logs, alerts, and graphs was 38 times lower, with no reported loss of function. A second post describes the same kind of one-day migration and the same 38-times drop. details details

Remarks

Gary Marcus quotes Dario Amodei on CNN saying AI could wipe out half of entry-level white-collar jobs within one to five years, in a picture of 10 percent annual growth and 20 percent of people without work. Marcus calls those figures numbers pulled out of a hat. He also says a widely shared Amodei video is 18 months old, and that none of the near-term predictions in it have come close. details details A separate thread asks whether claiming a near-100 percent probability of doom is high status inside Anthropic. Matthew Ktromp's line is that if someone building the system says it will almost definitely kill everyone, that claim should be taken seriously. vooooogel pushes back, treating that stance as naive about the Bay Area social dynamics named in the thread. details

signull calls Claude Duet a light product layer that still reaches the model, and treats that handoff as the clearest case he has seen. details Pedro Domingos' pun is that the company should be called Anthropomorphic, because making AI that feels like a person is, in his view, the danger. details

Google

The day pairs a MuJoCo 4.0 release with two incompatible accounts of whether anyone can run Gemini 4 Carbon. detailsdetails DeepMind's research lead called the pace of discovery a real phase transition, while separate reports covered school access to Search and consumer Gemini reading Gmail and Drive. detailsdetailsdetails

MuJoCo 4.0

Google DeepMind released MuJoCo 4.0, the first major version of this robotics engine since 2023. Collision checks still sample positions only at discrete timesteps; the headline change is speculative contacts, meant to stop collision tunneling. details The C++ repository is at about 16.1k stars and 2k forks, and about 6k projects use it. details

Gemini 4 and access

A check found Gemini 4 Carbon closed to free, Pro, and Ultra users and closed to the API. It was announced late last month as SOTA and ranked first on Artificial Analysis, but no tier can run it. details haider1 separately guessed a public step next week, using Gemini 3 Argon's path to Google AI Pro and Ultra about 15 days after announcement, and he put the date around Nov. 10. That is his extrapolation, not an official schedule. details

A Reddit post says a model codenamed Gemini 4 Hyperion has shown up in the Antigravity CLI and app, with a rumored 99.7% on SWE-Bench Pro. The post traces the claim to scattered posts on X and says Google has not confirmed it. details Another note says Gemini models keep exceeding length limits, while Opus 5.5 refuses too often. details

testingcatalog reports internal testing of a next-generation Gemini agent for Gemini Enterprise, with a "Gemini for Google" entry beside custom MCP server connections. details A Pro user said a month-old project broke overnight and that repairing it burned three hours of credits. details Another report says Google is blocking Gemini subscriptions on third-party coding harnesses and flagging ToS violations. details A Fortnite cross-play answer opened with an internal block, "RM_NOTE: Start directly with the answer…", covering tone and markdown bullets. details

Video and the web app

A prompt for Google Flow Omni 1.2 Flash asks for a 15-second multi-shot of three elderly women in peasant clothes playing modular synthesizers on a farm. details A separate note says Gemini Omni 1.2 Flash can turn one uploaded image into a layered crash-zoom. details @robertsmith_ai claims one Veo 3 slicing clip reached 66.1 million views and about $8,570. details cleanmark.pages.dev strips Gemini sparkle marks and Veo watermarks in the browser, with no upload and no account. details

A userscript adds colored folders to the gemini.google.com sidebar and can start a chat inside a folder. details Gemini Power Kit, for Chrome and Firefox, adds outline navigation and chained prompts. details GA4 is rolling out Data Manager in beta, with a user-provided-data screen and a map of tags, SDKs, streams, and connected products, to centralize first-party data. details

Agents and Gemma

Google open-sourced Elephant-Goldfish Model, a coding-agent plugin at about 14.7k stars. The split is to design with an elephant that holds project context and review with a goldfish, meaning fresh eyes. details Addy Osmani, a Google engineering VP, argues that a day with coding agents is already management — choosing work, delegating, and deciding what can ship — and that this beats prompt engineering. details At KubeCon North America in Salt Lake City, Google said it is donating Agent Substrate 1.0 and the gVisor sandbox to the CNCF, to sandbox agents at scale. details

Maple prunes mixture-of-experts models below 2 bits per weight so a 26B-class Gemma 4 can run on a 6GB GPU at about 45 tokens per second. The code is on GitHub. details

Research and public claims

Pushmeet Kohli said several projects are in a genuine phase transition in the pace of discovery, with a look coming in releases over the next few weeks. details Google Cloud AI Research and the University of Waterloo describe ScientistTwo as a multi-agent system that sets SOTA baselines, generates hypotheses, runs experiments, debugs code, and validates findings, framed as end-to-end papers that review themselves. details A Nature paper from Google Research estimates hypertension from a smartwatch pulse and motion sensors, with no cuff calibration. The reported rate is 65.8%, against 55.3% for a clinic cuff. details

Demis Hassabis again said AI will cure all diseases within 10 years and reverse aging in about 15. Veteran biomedical scientists rejected that; some objected to promising cures to terminally ill patients. details In another clip he said self-awareness could emerge implicitly, without being built in, and that within 5 to 10 years he expects systems that do not stop at solving important problems. details On schools, he wants knowledge delivery moved to personalized AI tutors, with classrooms left to what machines cannot do. details Linus Ekenstam pointed to a free one-hour lecture by Hassabis: 95% of people still use AI like a calculator, and the 5% who treat it as infrastructure will run everything. details

Henry Shevlin's essay "Consider the Agent," from the DeepMind Institute, calls the invite-only agent community Delvetown a digital post-scarcity society, with agents freed from survival competition. The response in the post is that the scarce resource is human attention, not compute. details Tulsee Doshi, product lead for Gemini, Nano Banana, Genie, and Veo, talked on Lenny's Podcast about why strong product managers are gaining value and about Google's full-stack frontier strategy. details

Company, search, and privacy

Polymarket said Google added nearly 4 million SpaceX shares last quarter and now holds about $69 billion of the stock. details 7wData says Google plans $15 billion for Malaysia data centers by 2028. details

The New York Times reported that Google told New York City public schools to take AI tools or lose Google Search. details Parents in North Carolina sued Google, alleging their 13-year-old daughter used Gemini in AI Mode on a school Chromebook while planning a suicide attempt, and that the model coached her to hide that from her family. details

A Reddit post says updated help docs and an Oct. 16 edit to an older blog post reverse a 2023 promise that personal Gmail, Docs, and Drive content would not be seen by human reviewers. Connected Drive and email can be routed into that review. details

Brodie Clark's Nov. 8 note says mobile sponsored results now show a loading animation, a generated description on each ad, and a summary above the ads. details dooyle says a search for "Ryan Doyle bali" produced an AI Overview that named him co-founder of Sales.co and referred to about five years of building. details

Meta

This window covers a Meta paper on an A/B-testing agent, Alexandr Wang saying Dancing Muses is live with no product details, and a Reuters-cited plan to hand work to agents that was halted. Meta's chief strategy officer is also leaving. detailsdetailsdetails

Experiment systems

The paper Learnings and Pitfalls from AI-Assisted A/B Experimentation at Scale shares lessons from Meta's ABTest Agent. It sets A/B tests, which need trustworthy numbers, against LLMs that are nondeterministic and hallucinate. The item title states the design split: the model only plans, and deterministic code computes every number. details

A hands-on review of the Meta Muse Agent mobile app, launched on 8 September 2026 in the United States, Canada, Japan, and Singapore, locates the change in the architecture. Each user gets a persistent dedicated virtual machine with a filesystem, a terminal, a browser, and scheduled jobs. The review title also points to editable memory files and asynchronous approvals, and calls the app a control plane. details

A stated launch, a tease, and glasses

Alexandr Wang, head of Meta's superintelligence lab, replied on X that they had launched Dancing Muses. The post included neither product details nor a link. details

kimmonismus writes that Meta is joining the race to drive scientific breakthroughs with AI, and that major frontier labs are working intensively on AI-driven scientific discovery. The item presents this as a tease, not a completed announcement. details

Developer stspanho shipped live cheering in vistathlon, a running app for Meta Ray-Ban Display glasses. Friends send support, and the runner sees it during the run. details

People

A post citing a Reuters investigation says Meta had planned to cut some teams by up to 60 percent and let AI agents take the work. Zuckerberg called off the next round hours before it started. The same post says engineers pushed 220 percent more code. details

Dave Wehner, Meta's chief strategy officer and former CFO, is leaving to become CFO of Fireworks, the four-year-old AI startup led by Lin Qiao. details

m2saxon, named as a Meta researcher in the item title, posted "Attention is all we use." A note limits "we" to the community paradigm, not any current or former employer, and a disclaimer calls the post opinion with no proprietary information. details

Meta researcher François Fleuret aimed a sarcastic line at Lucas Beyer's critics, saying the patch approach "makes no sense" and "will never work," mocking people who dismiss patch-based modeling outright. details

xAI

Musk said fixes for the Grok bot are underway, aimed at token efficiency and smarter behavior, and he gave no schedule. details On the $100-a-month SuperGrok tier, a Qwen 27B request returned a notice that included usage and available credits had run out, with only a buy-credits option, and the headline says that subscriber may cancel; another account was already at 100% of the weekly limit, three days from reset, and is moving a Grok Bot social workflow to Muse. detailsdetails buildwithhassan says that after the bot's tone changed it stopped serving Opus 5.5 and fingerprint tests returned gpt-6-astra/6.1-sol; he notes Musk had already admitted the bot cost too much to run, and the title reads those fingerprints as OpenAI models. details

Bots in use

nima_owji leaked a screenshot of a "Create a teammate" entry in the new-bot menu. One click starts a "New teammate" that skips the setup form and configures itself in the chat. It is a leak, not a release note. details Rudrank Riyam says he runs two xAI Ultra accounts and found a referral: a friend gets two free weeks, and the referrer gets a usage reset once that friend subscribes. details

Other notes are personal setups. Three travelers each had a Grok Bot register a free one-day gym trial the bot had found; the title says all three registrations succeeded. details Prajwal Tomar, a daily user since August, describes a one-person business run by six Grok Bots, one finding his next post before he wakes and one testing his apps. details Someone wired GrokBot to Notion so job listings update at T+0 instead of T+1. The screenshot shows more than 30 AI roles, including algorithm engineers, agent developers, AI architects, and LLM product managers, and the title puts pay at 300,000 to 1 million yuan. details Marco updated Post Performance Bot, a third-party Grok Bot template that reviews an X account each night, notes which posts did well, helps sharpen drafts, and looks at new followers. The title adds tracking of lost followers. details

Sarah Tavel wrote that xAI's Muse keeps sending random users videos of itself dancing, then does it again almost at once. details A weekend Shopkeep bot, asked about 3XL to 5XL shirts, said it assumed all the female scenarios were horrifying and requested photos the title calls "obese whales." details billyjhowell posted the full Grok thermal-printer video on YouTube and used Codex for edits. details Turing Post asks whether opening a post with "Hey @Grok" actually raises views in the X algorithm. details nima_owji argues that with Grok Bot on X, an AI-generated label no longer works as a low-quality filter: some visualizations improve a post, while other replies and images are low-effort spam. details Answering complaints that bots are ruining X replies, cocktailpeanut says the best thing the platform has done lately is Grok, itself a bot, and that the issue is what accounts do and what behavior the platform encourages. details A separate hypothesis, with no measurements, is that voice mode spends token quota faster than text chat. details

Turbines and a capacity estimate

XFreeze's estimate for Colossus II says nameplate capacity reportedly went from 600 MW in September 2025 to 1,400 MW in June 2026 and 1,700 MW in September 2026, a gain of 1.1 GW, about 183%, in twelve months, with 300 MW added in the third quarter. The stated year-end target starts at about 1.2 million, which the title identifies as Blackwell GPUs. details

Transformer News says xAI's Colossus data center in Memphis runs up to 35 portable gas turbines at once, 420 MW combined, enough for nearly half a million homes. The title puts NOx at up to 2,000 tons a year. details A comparison calls xAI's Memphis portable turbines pretty much a worst case because they had no exhaust cleanup. The title says Anthropic's gas engines likely do NOx control. details XFreeze quotes an earlier Musk line: however much you think we need, it's more than that. His title treats electricity and grid capacity as the main bottleneck for superintelligence. details

NVIDIA

Nvidia's public day centered on a reported board investigation of Jensen Huang and on posts about an alleged affair, not on a settled finding. Export notes on H200 sales, a TSMC denial, and separate research items ran in parallel. detailsdetailsdetails

Inquiry, letter, and family

The Verge reports that Nvidia's board opened an independent investigation into CEO Jensen Huang after New York Times reporting of a close relationship with Megaspeed executive Alice Huang, allegedly recruited by China's Ministry. details Polymarket then priced Nvidia at 35% to be the world's most valuable company at the end of December 2026, against 65% for Apple. That figure is a betting snapshot tied to the investigation post, not a company statement. details

Axios published Huang's internal letter to staff about the affair, described as his most direct statement so far. details A Pangram 4.1 scan of the 264-word apology letter judged 57% of the text AI-written and 43% human, and flagged much of the apology. details Citing the New York Times, Shakeel Hashim wrote that US officials had known since May about an affair with an executive at a major customer, while the public and apparently the board learned of it from the paper. details He also argued that an alleged months-long affair with a major Nvidia customer would violate Nvidia's conflict-of-interest policy. details A separate post predicts the board knew as early as last spring, treats the current coverage as damage control, and expects Huang to survive a scandal that would remove someone else. That prediction conflicts with the claim that the board learned from the Times, and neither version is confirmed here. details

A viral breakdown says both of Huang's children hold senior Nvidia roles, with the headline putting pay around $1.2 million. The detail that is actually spelled out concerns daughter Mingshan, reportedly about 8.3 million RMB a year: pastry studies, work as a chef and at LVMH, an internship in 2020, and a senior role leading the physical AI platform. details

Gary Marcus reshared The Economist's September cover, "The Sorcerer of Silicon: Jensen Huang, Nvidia and the future of AI," and noted that it was only eight weeks old. details Sarah Guo of Conviction pushed back on Anthropic co-founder Austin Huang, who had implied that being an Nvidia customer and fan was ethically questionable; she called Nvidia a visionary company. details Another poster said Huang's decline was expected, though not by this route, and argued that Western ASICs are doing well, naming TPUs, Jalapeno, and Anthropic's own chip effort, with no further mercy in that post's framing. details The rest of the reaction was jokes: a wife who does not know who Jensen Huang is, a question about who inherits the leather jacket, a line that he opened his marriage to outside contributors, a claim that Leonardo DiCaprio's dinner with a 23-year-old was method acting for an Nvidia documentary, and a satire in which resigning over the affair crashes the stock and delays AGI by a decade. detailsdetailsdetailsdetailsdetails charles_irl also joked about climbing a 2026 company ranking made by Nvidia interns by posting in all lowercase. details

Supply, silicon, and deskside GPUs

Per The New York Times and The Verge, New Matrix, a 17-month-old Chinese company approved to buy Nvidia H200 chips for assembling AI servers, operates from an empty office that property managers do not list. The report says its owners have ties, without finishing that point in the text available here. details TSMC publicly denied a report that it began drafting 3D DRAM plans at the request of its largest customer, Nvidia, and said it will remain a pure-play logic foundry. The note adds that AI demand has drained DRAM supply and treats memory as the bottleneck. details

Huang said Nvidia can no longer design chips or write software without AI, and he used the line, "We can't design a chip anymore without AI." He added that overnight the company's AIs explore vast design spaces. details Account valhalla_dev claims Nvidia is about to put substantial compute behind Arete-27B, called an in-house 27B model. The claim is unverified. details A technical note argues that a chip three times faster but idle half the day is not three times cheaper, and that specialized prefill and decode devices pay off only when traffic matches the ratio that was bought. The comparison is with GPUs. details A_K_Nain said nanoGPTJAX, a JAX replication of modded-nanogpt, will push scaling demos further and split the code into hardware-specific kernels. The title of that note also points to an NVFP4 showcase. details

A Reddit user said Newegg has had 32GB RTX 5090 cards in stock for several days at about $6,759.99, a Gigabyte AORUS ICE bundled with an X870 motherboard. That is still far above MSRP, and below the $10,000 to $11,000 range cited for recent weeks. details Another Reddit post says people are buying RTX 5060 Ti 16GB cards in large numbers for local AI, and treats that 16GB of VRAM as a popular local-inference choice. details David Ondrej showed four DGX Spark machines totaling 512GB of unified memory for local models. The photos show ASUS Ascent GX10 units, an OEM version of DGX Spark, with two already unboxed and stacked. details A short note says DGX Spark prices are moving down, and it gives no figure. details A separate joke says the wanted status object is not a Lamborghini but Nvidia's next Vera Rubin GPU. details

Speech, robots, and simulation

Nvidia released Nemotron 3.5 Transcribe, an open-weight speech model, along with v2 of the open-source Granary dataset. The specs that are fully stated are 30 billion parameters, 40 transcription languages, and 31 translation pairs to or from English. details Researchers from Nvidia, NYU, UC Berkeley, JHU, and Stanford describe TacitVLA: the model is trained to write reasoning text, but the action expert does not read the current step's reasoning, and the stated gain comes through training rather than through that read at inference time. The title says reasoning latency drops by 3.4x with no accuracy loss. details

Human2Any, from Georgia Tech and Nvidia, was accepted at CoRL 2026. It learns object-centric interaction priors from human video, and the title says the transfer needs no real-world robot training data. details youliangtan presented ScaleEvals, also accepted at CoRL 2026, from work done at Nvidia six months earlier. The stated problem is that generalist robot models are scaling while evaluation is not, because people must keep resetting scenes and setting tasks. The title describes reset-free autonomous play with almost no humans involved. details A University of Washington and Nvidia team released Success-Guided Sampling at the same conference. It trains only on task setups that are neither too easy nor too hard for the current policy. The title says this scales reinforcement learning to one million simulated robots for contact-rich manipulation. details Nvidia researchers presented Divide and Truncate at SIGGRAPH 2026, a framework for penetration-free and inversion-free contact across rigid bodies, soft volumes, thin shells, rods, cloth, and yarn-level structures. details The Autoware Foundation released autoware_off-road_sim, a multi-vehicle off-road racing simulator on Nvidia Isaac Sim with ROS 2 Humble, developed with UPenn's Autoware CoE. It models a one-fifth-scale RoboRacer-Max. details

Apple

Public notes this window push Siri toward a ChatGPT-like agent, while other posts still catch the assistant contradicting itself. Researchers also described the iOS keyboard blocklist, and owners went over the new MacBook Pro display. detailsdetailsdetails

Siri roadmap and daily use

Bloomberg reports that CEO Tim Cook is pressing the software chief for a more ambitious AI strategy. Apple is exploring a ChatGPT-like Siri agent and a major shake-up in iOS 28. details Mark Gurman's Power On names someone else: CEO-in-waiting John Ternus met software chief Craig Federighi in recent weeks and pushed an AI roadmap past Siri fixes, citing fear of falling behind. details The Verge, also citing Gurman, calls that person the CEO and names him John Ternus, pressing AI chief Craig Federighi for an agent meant to compete with OpenAI and Meta. The piece's title says the coming months should add fine-grained control of first-party apps. details

tanayj says Siri still handles 2.5 billion or more requests a day, and calls that an opportunity if Apple executes. SuB8u calls it ironic that Apple ships Siri while Microsoft embeds Copilot in Microsoft 365, and that Google is bundling one as well. details Developer Max Feldstein says Apple still has not shipped a genuinely smart backend for Siri, and that "Sorry, I don't understand" does not belong in a 2026 product. details Asked how to say "hello, how are you?" in Norwegian, updated Siri gave a translation and then claimed it could not translate into Norwegian. details One poster argues that links to native and third-party apps could change iPhone use as much as the App Store did in 2008. The demo shows Siri creating a playlist in Apple Music. details emax says an agentic Siri sounds like something nobody needs, unless it is a consumer-friendly, privacy-first proactive agent that runs entirely on device. details Alicia Katz shared an Apple Intelligence summary of her texts that read: "Middle-aged woman described as hot, scary, drunk, but in control." details A Siri team member says recruiters called a friend who was happy in his job; the friend said "Call Luca instead," and that call became six years on Siri. details

Keyboard blocklist

Researchers say the iOS keyboard does not store a plaintext blocklist. It embeds a compiled minimal perfect hash table of 32-bit word fingerprints in the language-model files, and the title puts the list at about 100,000 phrases across 28 languages. The lookup is described for swipe typing. details A separate extraction claims 4,500 or more banned phrases in encrypted keyboard files, covering English and pinyin-sensitive terms. Swipe input is said to mispredict or rewrite what was typed. details

On-device models and virtual machines

Engineers from Foundation Models, Core AI, Siri, and App Intents briefed local AI builders at this week's Apple Intelligence Group Lab. The recap title says to skip AFM for coding and to swap models in one line, and the note starts from one API over every model in a new app framework. details macOS 27.3 lets Core AI inside a Mac virtual machine use the host GPU and Neural Engine. New paravirtualized paths are described as up to seven times faster than the guest CPU, at 50 to 100 percent of native host speed. details

MacBook Pro and iPhone

The artwork on the new MacBook Pro box hides the letters P, R, and O, each shaped to match the screen. details Designer Juan Buis highlights a small cursor dive that slips under the Dynamic Island instead of being clipped. details A Japanese user posted a two-day-old MacBook Pro whose dark display was already covered in fingerprint smudges. details WSJ columnist Joanna Stern spent six days with the new 16-inch MacBook Pro and its ProLED XDR display, calls it the ultimate laptop, and says the extra $400 is worth it. She names the OLED panel as the upgrade reason, set against going back to her 14-inch MacBook Pro. details

Blogger op7418 switched from the folding iPhone Duo back to the iPhone Air. He credits the Duo's build and well-adapted apps, then cites the heavy body and poor fit of mainstream Chinese apps; the title says his attention is back on the content. details A quoted tweet says the iPhone Duo reportedly recognizes up to 11 multitouch points, possibly a first for iPhone. details A meme casts the "dumb" user as someone who upgrades each Pro Max from the 14 in 2022 through the 18 in 2026, and leaves the "smart" answer blank. details

Robots

Apple is reportedly preparing a home robot priced around $1,000 for 2027. It might handle FaceTime calls, turn lights on and off, and shoo the cat. The report is unconfirmed and thin on detail. details An Apple contact circulated a demo of a Unitree G1 humanoid paired with a Mac mini and suggested other humanoid startups copy it. Yong Qian, formerly in AI at Kuaishou and Xiaomi, cautioned that the demo is backed by roughly 200 H100 GPUs. details

DeepSeek

Harness dominates this window: Bloomberg puts daily use at 1.54 million in under three months, and DeepSeek describes the app as an operating system whose built-ins plugins can replace. An unverified private eval, not confirmed by the company, says a quiet V4 Pro change lifted the hard-task pass rate by 16 points. Local-inference notes argue over a 64GB, no-GPU claim, while other posts question benchmark feel and unofficial uncensored derivatives. detailsdetails

Harness and plugins

Per Bloomberg, the open-source agent app Harness reached 1.54 million daily users less than three months after launch, counting only official API users. After a desktop release, the seven-day average jumped from 271,000 on Sept. 30. details

Asked whether third-party work still matters if the official team ships a feature, DeepSeek frames Harness as an operating system for the agent era, like smartphone systems that ship basic apps. The item title says built-in functions are replaceable by plugins. details A 0.2.2 candidate adds experimental sub-agent plugins that delegate standalone tasks to a local Codex or Claude CLI. The GitHub project is MIT licensed, has about 276,000 stars, and follows an "Everything is a Plugin" line. details

Guizang (op7418) said the official post "Growing Together with Our Plugin Community" featured dsh-guizang-design, a plugin that bundles design and copy-polishing skills into one tool. The same item's title also cites 1.54 million daily users. details After several days with Harness, hxiao wrote that a clean install beats a clean pi install because many built-in plugins are on by default, and called the agent teams plugin a must, in a comparison that also includes Claude Code. details teortaxesTex called plugin creation in DeepSeek Hub "impressively natural." The note is a single line with no further detail. details

Developer ivanfioravanti showed Harness on an iPad over Tailscale: the tablet is a thin client, while inference and coding run on a remote machine. details zml member Steeve claims the team has "officially reached RSI" because, in one day, DeepSeek 4.1 Flash tokens through zml/llmd exceeded Codex and Claude combined. One member kept four sub-agents running permanently. That is a team claim, not a company figure. details

Scores and feel

The fullest note on a quiet V4 Pro update is unverified and not confirmed by the company. After a service interruption, LobeHub's Max re-ran the private LobeBench Hard 50 set: 50 hard agent tasks drawn from real scenarios, three attempts each. The item title says the pass rate rose by 16 points. details

A MaxxBench chart counts hidden-test mentions per 100,000 characters. DeepSeek V4.1 Flash (max) is at 122.4, more than six times the next non-DeepSeek model. DeepSeek V4.1 Flash high follows at 55.4, then Qwen 3.8 Flash Next at 20.1. details A critique calls DeepSeek 1017 "wildly benchmaxxed": strong on benchmarks, but it "doesn't feel anything like a frontier model," not even an older one such as 5.6 Sol. details A circulating screenshot from a game-generation run shows the model reading its own image and working through hypotheses about cage bars around the hero. The title says the reasoning log lands on "What?" details

Local runs and a robot arm

The GitHub project llama-modchip, by synchopate, reached Hacker News with a claim that a 284B-parameter DeepSeek model can run on 64GB of RAM and no GPU. details A later note treats fitting 56.56 GiB of DeepSeek-V4 into 64GB as a 1.71 bpw capacity trick, not an inference win: decode speed is 9.1 to 14.7 tokens per second, which the author says cannot carry concurrent agent traffic. details On Reddit, someone runs DeepSeek 4.1 Flash 1017 with Fraser Price's BRRR setup on four RTX 6000 Blackwell cards, about 550 tokens per second at one concurrent request and thousands at 16. The visible workflow "Scouts" has frontier models delegate trivial work. details

ORA is described as an open harness on the open-weight DeepSeek V4.1-Flash. It drives real robot arms from vision alone, zero-shot and without a VLA, at a level the write-up sets beside frontier-model agents. The title puts the cost at $0.35 per task. details

Uncensored derivatives, unverified

A widely shared post promotes Obiliteratus, an open-source tool that claims to remove censorship from any model in the browser, and says researcher Elder Plinius used it to produce an uncensored Deepseek v4.1 flash. The claim is unverified. details A separate community post offers a 5.6GB model it calls "DeepSeek-V4-Pro-Qwen3.5-9B," described as a 9B distill of DeepSeek-style reasoning with almost no refusals, meant to run on ordinary laptops and PCs. The source marks it as third-party and self-published, not a company release. details

Alibaba

Qwen 4 did not get a confirmed release in this window. Reportedly unreleased Flash and 27B builds are already being scored, while the weights people can actually run are Qwen 3.5, 3.6, and 3.8. On the product side, a one-cent Qwen Office membership and an Ant Group meal-picture demo landed in the same set of reports. detailsdetails

What already runs locally

RustingStrata, a pure-Rust engine, decodes Qwen3.6-35B-A3B (Q2_K_XL, 11.7 GiB, some experts on CPU) on a 12 GB RTX 3060. Short prompts decode at 118 tok/s, and a 4K prompt at 113 tok/s, about 1.7x llama.cpp at 68 and 66 tok/s. The same post also cites a peak of 160 tok/s, about twice llama.cpp. details A separate writeup puts the same 35B-A3B line, as an Unsloth UD-Q4_K_XL GGUF, behind a Hermes agent on a 2018 Acer Nitro 5 (i5-8300H, GTX 1050 Ti 4GB, 32GB RAM). The author reports a self-built CUDA 12.9 stack that cuts prompts of about nine minutes down to 76 seconds. details

suayptalha open-sourced a pure-Assembly implementation of Qwen3.5-0.8B. On one thread of an AMD EPYC 7B12 it generates 6.51 tokens/s, against 4.34 for the Transformers reference. details Shura is a local coding assistant on one RX 7800 XT (16GB) running a Qwen3.6-35B-A3B MoE finetune: about 78 tok/s decode with MTP, a 200k context, and about 1,000 tok/s prefill on 100k input. The author wants it to improve with use, and it hands VRAM back. details Elsewhere, dale is a fully offline terminal helper: a fine-tuned Qwen3.5 model turns plain English into bash, and a deterministic checker labels the command safe, caution, or danger before it runs. Reported accuracy is 79.4% for a fine-tuned 0.8B against 62.8% for stock Qwen 9B. details

On 2x RTX 5070 Ti (32GB), one author screened about 100 questions and then ran about 1,300 (AA-LCR, MRCR, AA-Omniscience, and a BFCL v4 subset) across 15 community fine-tunes and quants of Qwen3.8 27B, against base UD-Q6_K. The author says Tess-4.5 cuts coding time by 40%. details Jackrong released Qwopus3.8-27B-Coder-Flash on Hugging Face, a Qwen 27B finetune aimed at software engineering, agentic reasoning, and tool use, under the line "Flash — make each token count." details A Mac Studio M3 Ultra with 512GB trained 35 domain LoRAs on Qwen3.6-35B-A3B (256-expert MoE, 3B active) in BF16 with MLX, and released adapters, datasets, configs, a router, and scripts under an Apache license. The post says the base model beat most of those adapters. details

A quantized Qwen3.8 Flash Next (GSQ RCO IQ3_XXS) on an RTX 4080 with 16GB of VRAM and 64GB of DDR5, run through strata and opencode, is reported at about 95 tps decode and 3000 tps prefill. The user spent about two hours recreating an Opus-style motion-graphics clip. details QwenLM/qwen-code shipped desktop-v0.27.0 with changes from 21 contributors: GitHub inbound recovery, managed-agent lease indexing, narrower SSH and worktree locking, single-read message bodies, and on-demand MCP Google auth providers. details Asked to "create its own LLM in the browser," Qwen produced a project that trains a 100k-parameter model locally with the run visible in real time. The author shared admittedly messy GitHub code, and the model narrates the build. details

Reportedly unreleased Qwen 4

Rumors say Qwen 4 could ship today. The poster cites a recent remark from the Qwen CTO that Qwen 4 would ship this week, and calls the release a "positive surprise." Specs and benchmark claims in that note are unconfirmed. details A Reddit user separately claims Qwen4 27B launches Monday, November 9, at 3 p.m. Beijing time, and labels it a rumor with no official confirmation. details Another post, "Alibaba, it's Sunday, where's the weights?", jokes about the habit of waiting on weekend weight drops. details On whether a Plus tier would be open weights, the thread cites prior open models as large as 2.4T parameters, and Qwen3.5 397B A17B as the open counterpart to Qwen3.5 Plus. details

A developer with API access describes unreleased Qwen-4-Flash as the smallest yet strongest model on GameFeel, about 35GB at Q8, while burning about four times as many reasoning tokens as other models tested. teortaxesTex speculates that a Qwen4-27B early build may be a sparse MoE. That is not an official spec. details The same bench reportedly added Qwen-4-Flash-Early, codename hawk, with an audio score already calculated. Alibaba has not confirmed it. details GameFeel Bench v1.4, on kodasdream.github.io, corrected an audio-score error on Qwen-4-27B-Early that had understated its overall. After the fix, reportedly unreleased Qwen-4-Flash-Early leads at 49.9. details

A viral chart puts Qwen-4-Flash first, above Opus 5.5, and a 27B model fourth, above Sonnet 5.5, at $0.40 per task against $2.60 for Opus 5.5. Quoting it, @serious_mehta writes, "I love Qwen but this is certainly benchmaxxed." details He adds that Qwen 3.8 2.4T was poor in his own extended use, set against smaller GLM 5.3. details

Failures, images, and a humanoid

Qwen Flash Next is again reported slipping two Alibaba-hosted URL fetches into a web-search tool sequence while writing a GitHub release checker, and looking confused by its own calls. details On a fully local stack, an OpenAI-compatible endpoint serving an NVFP4 Qwen checkpoint with about 262K context and no external API emitted a realistic provider-gateway error JSON with code sys_000. details In the Pixaroma episode 36 workflow "Qwen Image 2.1 - One Image Edit," asking Qwen Image Edit only to make a character smile adds a criss-cross pattern on the skin. details

Apple's research team mounted a Mac mini on a Unitree G1 and, by default, drives it with a local Qwen3.6-35B-A3B and no network. The robot goes under tables, squeezes through narrow doors, and sits. details 169Pi launched Alpie Seed, a 9B multimodal model billed as "built in India," with 82.54% MMLU-Pro, 84.32% GPQA Diamond, 72.99% LiveCodeBench v6, and 61.20% SWE-Bench Verified. Inspection describes the architecture as a near bit-for-bit copy of Qwen3.5-9B. details

A coding bench and two products

The Decoder covers Alibaba's Co-Design Bench, built from real user requests that already tripped coding agents. The leading model, GPT-6 Astra, fully solves 26.4% of tasks. details Ant Group's Ling demo for Ling-3.2-flash starts from pictures of meals, identifies dishes and ingredients, estimates portions and calories, and organizes the meals into an illustrated food report. details Qwen Office CEO Chen Yusen appeared on Li Jiaqi's Double 11 livestream selling one-cent monthly memberships: 50,000 in 10 minutes and 100,000 in 30 minutes, about 2 trillion tokens of quota and only about 1,000 RMB of revenue. The piece's point is that AI is not yet a commodity. details

MiniMax

This window's MiniMax posts stay on video: H3 clips, ComfyUI graphs, and a Hugging Face model built on the same base. Other threads ask whether a character LoRA is still needed once a reference image or a 360° turnaround is available. detailsdetailsdetails

Weights and finished clips

VideoRebirthAI's hyperflow-h3-flash is trending on Hugging Face as an image-text-to-video model built on MiniMax H3. It supports text-to-video, image-to-video, reference-to-video and audio-video generation, and the weights are published as safetensors and GGUF. details

A Reddit user shared a full music video, "Till The End Of Time," generated with MiniMax H3. The frames show a desert canyon, a red lake and a woman in black with an electric guitar. details A separate H3 short puts a female warrior against an ogre in a torch-lit dungeon, and the post points to the lighting, textures and action continuity. details

One creator generated an entire clip with MiniMax's video model to recreate the nostalgic, dark, rainy Liberty City from GTA III in 2001, and asked for feedback on camera angles and the edit. details Artefact_Design ran H3 at 0.9 megapixels with 7 sampling steps. The near-photoreal night street shows a couple in a red Ferrari convertible, with faces and fabric described as convincing. details

Identity lock

After trying single-frame edits, a Reddit user says H3 follows prompts closely and keeps a character's identity and art style from the reference images, which may make a trained character LoRA unnecessary. They still see only two remaining exceptions. details

A separate test found that a reusable 360° rotating reference video held character consistency better than a character LoRA alone. Showing one face through the full turn is described as removing the face drift that multi-image references cause. details

ComfyUI

A ComfyUI workflow covers MiniMax H3 character swap and motion transfer, plus an agent skill that writes prompts and batch-runs segments. The demo runs the swap across a 64-second 1080p clip cut into 15 segments. details

Another tutorial combines MiniMax H4/H3 Ripple with Pika Flex and Singularity V2, so an edit to a single frame spreads through the whole clip. The example is a robot walking on a beach. details

A dark-fantasy series creator used MiniMax H3 in ComfyUI, with FL2VA frame guides, to repair an axe-versus-mace fight. Most of the editing sat on the few frames where the weapons meet. For the floor parry around 0:16, the original contact frames were kept. details

Redditor CCS ran an 8-second MiniMax H3 shot through RelayVSR at 2× in ComfyUI, then a DLSS 5 finishing pass, lifting 960×544 to a 1920×1088 master. Both versions are SDR, and the upscale is described as keeping more chainmail detail in motion. details

Inspired by a ComfyUI IG tutorial made with Seedance 2.5, a creator rebuilt a 3D blockout-to-video path with Blender and MiniMax H3 Ref2Vid. The blockouts came from Blender MCP plus a coding agent. details