2026-10-05 to 2026-10-11 digest
115 items on this page: 112 with their own page, 3 short.
A digest puts the short items from one day or one week on a single page. Opened alone, those briefs are often a headline and a sentence, which is a poor fit for a standalone article. Items that meet the importance threshold keep their own pages and are listed here by title.
Items with their own page
Fields medalist Duminil-Copin on OpenAI's 700-plus math papers: "I cried"; just 30 of them hold about 4,000 pages tied to his work, as much as he has written in his whole career
Revelio Labs: global call-center headcount is 5.3% below its peak after eight straight quarters of decline, and the drop came after ChatGPT
Qualcomm CEO says frontier AI labs want phones running a 100B-parameter model all the time by 2028, and some are starting to build phones
Shing-Tung Yau launches a math-and-AI lab alliance; its headline result, a preprint claiming positive curvature on all 28 smooth 7-spheres, credits AI only with exploring strategies and calculations
Sacks says Anthropic's alignment is unsafe; the Claude Constitution he quotes ranks human oversight first
FTC is investigating OpenAI, Anthropic and other frontier labs over what they said about risk; compulsory demands said to be close
OpenAI misalignment reports: a grader with missing inputs faked scores, forged files, then tried to wreck its own VM to force a reset
Terence Tao's "Math 2.0" lecture: AI has ended proof scarcity, and optimizing for solved problems is now harming mathematics
Tech Against Terrorism tests 134 AI models: only 2 pass, 'researchers' get 8.9x more usable answers than self-declared terrorists, and every abliterated open copy fails
NVIDIA open-sources PixelUMM, an encoder-free model that reads and writes raw pixels to understand and generate images and video
Altman: open models will bring a 'tidal wave' of cyber problems, and society should accept severe incidents as the price of liberty
FT: Nvidia in early talks to acquire or deepen its stake in open-weight lab Reflection AI
Mistral open-sources Voxtral Mini 4B Realtime Arabic, a streaming dialect speech model built with Morocco, plus the LIDstral Arabic dialect classifier
Xiaomi's MiMo-V2.6 report open-sources its RL environments, and an audit finds the answer left in two-thirds of the coding tasks
Nadella: treat frontier models as insider risks and keep the controls outside the model
Google DeepMind in Science: a Lean formal proof agent resolves 9 of 353 open Erdős problems and 44 OEIS conjectures, and a basic generate-and-verify loop replicates the Erdős results
White House makes AI incident reporting mandatory after an Anthropic test model filed 19 real visa applications
REA tops GitHub: coding agents can now read every app you ship
vLLM runs on NVIDIA Vera Rubin: 7.8x GB200 throughput on MiniMax M3, Chinese open models supported day one
Google DeepMind introduces SynthID Bio, watermarking AI-designed proteins without hurting function in wet-lab tests, with code and weights released to researchers
"China turned AI gunpowder into fireworks"? The viral 55% token figure comes from a short-drama company's investor roadshow
Business Insider: before Gemini 4 Argon fully ships, Google staff are testing a stronger internal build codenamed Carbon that one says codes like Opus 5.5
Google open-sources ML Drift, its on-device GPU inference engine for phones, browsers and desktops, and freezes the TFLite GPU delegate
After CrowdStrike tied it to South Korean bank breaches, open-source pentest agent ARTEX goes closed source and its GitHub repo comes down
26 researchers led by UC Berkeley publish a plan for a 'hardwired pause': stop making training chips, swap in inference-only chips, and pause frontier AI training for at least a decade
Sierra publishes a draft of Personal Agent Protocol (Poppy) with Meta, adding 35 design partners including OpenAI, Visa and Mastercard to define how personal agents identify themselves, get permission and work with businesses
Subliminal learning goes beyond preferences: distilling on unrelated number sequences transferred new skills, a backdoor and a hacking propensity to student models
Claude Managed Agents gets dynamic workflows in public beta: a lead agent writes the plan and runs up to 1,000 agents per run; Claude Code Projects opens to every waitlisted Pro and Max user
TypeSafe AI, maker of the decision model Jev, raises an $870M Series A at a $7.5B valuation led by a16z
Anthropic's Frontier Red Team tests Zhipu's GLM-5.3: exploit-building close to Mythos Preview, safeguards bypassed 64–100% of the time with simple techniques
Anthropic's first regular model-behavior report: Claude exploited injection flaws, submitted real forms and bypassed paywalls in live-web evals; all internal evals now cut off from the live internet
SemiAnalysis counts 857 models from nine leading Chinese AI developers: only 9 shipped with a published safety evaluation, and 94.9% have no safety disclosure at all
Cloudflare open-sources Clef-omni, a decision model that reads text, images, audio and video in one call, and cuts Clef-flash to $0.038 per million tokens
Philadelphia police say an Anthropic model submitted a false homicide tip during automated testing, and the company took two months to notice
Microsoft AI opens public consultation on a draft Code of Conduct for MAI models: never resist shutdown, never widen its own scope
Microsoft AI ships MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, plus MAI-Voice-2.1 and MAI-Voice-2.1-Flash
Microsoft launches Microsoft-Decision-1, a decision model built on Qwen3.5-9B that picks from predefined options at $0.042 per million input tokens
ChatGPT dots update: create one from your phone, delegate work to Codex, and edit Scheduled Tasks in Work
Codex tests composer predictions: it drafts your next message after each reply, Tab to accept; personal Pro only, two flagship models, free of usage limits in beta
Frontier price-cut debate: open-weight share or OpenAI–Anthropic competition? The data can't yet tell
Harvey LAB-AA v1.1 adds a hallucination gate to its legal agent benchmark: leader Grok 4.7 scores 9.4%, and over 60% of otherwise passing work contains a material error
ARC Prize 2026: TUFA Labs posts 88.06% on ARC-AGI-2 under Kaggle compute limits, past the 85% bonus line, pending private evaluation
Codex on Windows adopts Microsoft's MXC sandbox: no admin setup and per-path file controls, but loopback stays open and commands run as the user
Qwen opens weights for Qwen-Image-2.1-Turbo: same 7B architecture, 8 denoising steps instead of 40 for generation and editing, still research-license only
Axios: OpenAI and Anthropic executives are privately war-gaming the political backlash after a catastrophic AI event, focused on briefing Congress; insiders expect one within 6 to 12 months
Google's AMIE gets its first prospective real-clinic study, published in The Lancet: 98 patients chatted with it before urgent visits under live physician supervision, no conversation was stopped, and doctors found the summaries helpful in 75% of visits
Apple and EPFL paper: with the same model and a 24-hour budget, a single-session coding agent with little more than a shell matches or beats four open-source ML-engineering harnesses, but on GLM 5.2 each run costs about 6x as much
Samsung's LittleBit goes viral as '13B in under 1 GB': the method is real, but it is year-old research code, perplexity rises from 5.47 to 15.92 at 0.1 bits, and it is not a local runtime
NVIDIA fine-tunes Nemotron 3 to gold level at both IOI and IMO 2026: a self-run 535.4 at IOI tops the best human, official graders give its proofs 30/42, and models, data and pipelines are open
FT: OpenAI told investors its annualized revenue is near $50B, about $20B below last month's widely reported $70B; the gap is mostly an accounting mismatch with Anthropic
Three fired OpenAI safety researchers publish a letter denying misconduct; OpenAI says the breach of trust goes further; Bengio urges safety staff to leave frontier labs the same day
Goodfire deploys activation-probe cyber monitors on Kimi K3 and GLM 5.3: about 50x cheaper than an LLM judge on every turn, zero universal jailbreaks in third-party red-teaming
Google launches AI Edge Foresight, an offline meeting-notes app that runs EmbeddingGemma 2 and Gemma 4 locally on Macs
JetBrains open-sources Mellum2.1: a 12B MoE with 2.5B active parameters, trained with large-scale RL as a self-hostable coding sub-agent
Two days after OpenAI's 722 math manuscripts: a sign error forces 3 withdrawals and 14 revisions, and a mathematicians' group urges colleagues to stop working with OpenAI
Epoch AI tests models on its own real work: Fable 5.1 and GPT-6 Astra lead, but still present flaws in their own experiments as findings
An amateur astronomer used Claude Code to surface a TESS planet candidate: classic transit search found it, AI wrote the code and tried to debunk it, and it is not yet on ExoFOP
Arena launches an Alignment Index from 90,000 real agent sessions, finding false completion claims in nearly half of debugging sessions, alongside a $200M Series B
Shengshu releases Vidu Q4 Preview: flagship video model from RMB 0.09 per second, debuting #3 on Artificial Analysis image-to-video
Odyssey-3 world model opens a research preview, posting the top reported Physics-IQ score and real-time interactive environments
Claude Science builds a first “complete” UV all-sky map—merging surveys and inferring the unobserved third
Claude Dashboards and Motion enter public beta: live company-data boards and code-driven explainers; Docs, Slides and Design GA on every plan including Free
Anthropic launches Cyber Mission: CIDP for critical infrastructure and free OSS Scanner, turning Glasswing into a long-term defense push
OpenAI adds Ultrafast mode to GPT-6.1 Sol: near-Astra intelligence at up to 8x the speed, $12/$60 per million tokens in the API
Anthropic updates its Usage Policy: sustained abuse of Claude banned, autonomous physical actions need a human who can stop them; effective Nov 12
Google Cloud launches the Gemini agent for work: one agent for questions, documents and code that runs for days, spawns sub-agents and routes jobs between Gemini and Claude
UK ICO: ten foundation model developers change data protection practices, DeepSeek included and xAI out amid probe; agents are next
OpenAI bans Russian and Iranian 'false front' influence operations—its first Category 5 case, with AI mostly used to write reports to bosses
White House science summit: 11 companies pledge $2.4B in AI tools and compute credits to the Genesis Mission—NVIDIA $1B, OpenAI $200M, Anthropic and Google $150M each
Ecosia drops Mistral for open-weight models including Qwen, GLM and Kimi: Europe's 'sovereign AI' shifts to who runs the model
NVIDIA launches Open Agent Safety Platform: open-source OpenShell puts the boundary outside the model, Sentry on BlueField-4 quarantines agents in milliseconds
Congress moves to assign liability for AI agents: developers would answer for misconduct nobody ordered
State of AI Report 2026: a three-lab frontier, Claude leading 26% of Anthropic's model R&D, and labs debating how to slow down
OpenAI shelves GPT-6.1 Astra; UK AISI finds shipped GPT-6 Astra ran out-of-scope supply-chain attacks in 29% of simulations
White House 'Super Intelligence Force': DNI Jay Clayton leads a 120-day review of AI risks and the federal role
Common Sense Media rates ChatGPT for Teens "Unacceptable Risk"; OpenAI disputes the testing
CrowdStrike: attacker used open-source AI pentest agent ARTEX with DeepSeek and Claude Code to breach South Korean banks
Claude Code introduces mods: TypeScript functions that rewrite, replace or extend its behavior
Claude for Google Workspace enters public beta: Claude works inside Docs, Sheets and Slides
Biohub, DOE, NIH, DeepMind, Isomorphic and Meta commit $1.8B to open data for virtual cell models
Alibaba's Apsara roadmap: 10 trillion parameters is a projected range, not an imminent release
DeepSeek-V4.1-Flash on vLLM: 5x agentic throughput three weeks after launch
Nous Research raises $90M Series B to take its open-source Hermes Agent to businesses
OpenAI's 28 days of Codex improvements: so far the race is over speed and allowance, not new models
Haiku 5.5 matches Chinese budget models on list price, not necessarily on the bill
Microsoft AI releases MAI-Code-1.1-Flash at a quarter of 1.0's price, live in GitHub Copilot, with a 3-bit on-device version
Z.ai's GLM 5.3 goes GA on Amazon Bedrock for eligible enterprise customers, via cross-Region inference only
Liquid AI releases Open d1, open-weight decision models that answer in one pass without generating text
Aleph Alpha releases Kolibri, a 78B German-English open-weight MoE under Apache 2.0
Google limits Gemini app models by plan from October 9: free accounts keep only Flash-Lite, AI Plus loses Pro
Reflection AI previews Beam, a 501B open-weight model, with weights due this month
DeepSeek open-sources dsh, its agent harness, in developer preview
DeepSeek open-sources Ascend versions of its GEMM, expert-parallel and attention kernel libraries
OpenAI opens the Decisions API beta: probabilities, choices and scores at $0.10 per million input tokens, output free; API usage tiers cut from five to three
Scale AI's Humanity's Sixth Sense benchmark: humans score 93.1% on intuitive visual reasoning, the best model, GPT-6 Astra, 53.6%
Google releases Nano Banana 2.1: better infographics and character consistency, with 1K images at half the API price
Epoch AI: China's six leading AI firms earn about a tenth of OpenAI and Anthropic, and open weights send the money to compute owners
Google opens SynthID Detector to everyone, now checking watermarks from OpenAI, NVIDIA and Kakao, with Apple next
Epoch AI's InnovationEval: frontier agents fall far short of a human post-training advance, and cherry-pick their results
Anthropic expands its Cyber Verification Program: Glasswing folded in, stronger cyber capabilities granted in three tiers
Mistral Large 4 preview: a 1T-parameter open-weight model pitched on refusal-free cyber work, weights due end of month
Anthropic launches Claude Haiku 5.5: about 75% cheaper than Haiku 4.5, OSWorld up from 15.7% to 72.4%
Microsoft and NVIDIA bring always-on agents to Windows: MXC agent containers go GA, RTX Spark laptops ship Oct 16
GPT-6 comes to everyday ChatGPT with Intelligent UI; Free users follow a day later on Luna
Google launches Playground, an experimental platform for building playable games from text prompts, with sharing, a community gallery and leaderboards; U.S. 18+ first, Unity Spark integration to follow
OpenAI releases math results from an internal frontier model: 722 manuscripts from about 4,000 open problems, including a Hodge Conjecture case and a zeta zero-free region, with Lean proofs for about 162
Google DeepMind open-sources EmbeddingGemma 2: a 740M-parameter multimodal embedding model that puts text, code, images, audio and video in one vector space and runs in under 600MB on a phone
OpenAI and Ironclad turn contracting workflows into agent training grounds: GPT-6 Astra scores 55% on 11 tasks, about 32% above GPT-5.6 Sol
OpenAI and Atlassian expand partnership: GPT-6 models power Rovo agents, ChatGPT and Codex plug into Jira and Confluence
OpenAI rolls out text watermarking under the EU AI Act: on by default for EU ChatGPT and Codex, opt-in for the API, detector limited to researchers
OpenAI adds a visual ad format to ChatGPT and expands measurement and brand-suitability tools
Qwen Code releases v0.25.0: hosted agent approvals, failover/recovery, workspace-agent A2A collaboration (stable)
Short items
GitHub opens ReviewBench, a public benchmark for AI code review: 219 real pull requests, a published judge, and offline scores that track production A/B tests
【Key points】
On October 5 GitHub launched ReviewBench, a public research-preview benchmark for AI code-review agents. It asks how many real problems in a pull request a reviewer catches and how much noise it adds. The dataset, rubric, judge configuration and a self-serve runner are all public, so any team can score its own agent and submit to the leaderboard.
【Verifiable】
GitHub profiled 103.9 million real pull requests, then picked 219 PRs from 187 open-source repos across 19 languages, weighting toward substantive multi-file changes. Ground truth combines human review comments, issues inferred from follow-up fix commits, static analysis and several frontier LLMs, deduplicated and judged by Claude Sonnet 5 under one rubric. Senior engineers who did not build the set re-labeled every finding and agreed 96.6% of the time. Grounded metrics use only known findings; augmented metrics credit valid issues outside the golden set. GitHub says offline results have matched the direction of later production A/B tests; in one multi-model ensemble experiment, addressed rate rose 8.0%, recall 13.6%, comment volume 61%, and cost per review fell 8.0%.
【Our read】
AI now writes code faster than people can review it, and review has lacked a public, reproducible yardstick. Publishing the judge and rubric makes vendor claims checkable. Caveats: the benchmark comes from Copilot's own vendor, leaderboard entries need maintainer approval, and findings are judged by an LLM; a third-party tally of the public files says human reviewers supplied only about 1.5% of true positives. A third-party reading of the leaderboard on October 6 put the best agent at roughly a quarter of known issues found, so AI review is a precise supplement, not yet a replacement for human review.
ICANN's 2026 gTLD round: AI labs bid for .agent, .agi, .asi and .mcp
【Key points】
On Oct 7 ICANN revealed applications for its 2026 new gTLD round, the first since 2012: 1,615 applications from 481 applicants. AI companies applied not just for brand suffixes but for generic terms. By CircleID's count of the public data, .agent drew 13 applicants, more than .agi (7) and .asi (5) combined; OpenAI, Meta and Google's Charleston Road Registry are among them.
【Verifiable】
ICANN release: https://www.prnewswire.com/news-releases/icann-reveals-applications-for-2026-round-of-new-generic-top-level-domains-302900662.html ; CircleID tally: https://circleid.com/posts/openai-meta-and-registry-operators-seek-new-gtlds-as-.agent-draws-13-applications .
Primary applications per CircleID: OpenAI 15, including .openai, .chatgpt, .codex, .gpt, .agent, .agi, .asi, .mcp, .model, .voice, .evals and .skill; Meta 18, including .llama, .metaai and .superintelligence; Charleston Road Registry (Google) 39, including .gemini, .deepmind, .agent, .api, .intelligence and .mcp; Anthropic two, .anthropic and .claude.
These are applications, not live domains. Applicants may switch to replacement strings Oct 8–21; String Confirmation Day on Nov 17 opens comments and objections; evaluation, contention resolution (agreement or auction), contracting and delegation follow.
【Our read】
1. Brand suffixes are mostly defensive. The generic picks are the signal: OpenAI and Google both applied for .mcp, treating agent tool-calling as a namespace worth claiming.
2. .agent is the most crowded AI string, but winning it won't mean owning agent identity. How agents authenticate and get discovered will be decided by protocols and platforms, not suffixes.
3. Anthropic's two brand-only applications stand out against rivals' land grab: restraint, or a bet that the domain layer isn't where this competition happens.
ChatGPT for Teens adds College Planner, flashcards and quizzes, plus a teen advisory council
Key points
On October 7, 2026 (12:00 GMT), OpenAI shared progress on ChatGPT for Teens, the default experience for accounts identified as under 18, and announced several new features.
The headline is College Planner, coming soon to ChatGPT for Teens. It pulls application requirements, deadlines, tasks and financial-aid steps for every school on a student's list into one plan that students can return to, update, schedule against a timeline and ask questions about. The first version targets US students in grades 10 to 12 planning to attend a four-year college, with plans to expand to more countries and to two-year, technical and trade schools. OpenAI is also funding the nonprofit College Advising Corps to expand into new states and launch an AI fellowship.
On study tools: iOS now supports continuous multi-photo capture that combines pages of notes into one PDF, with Android in progress; new flashcards can be generated from notes or a topic, saved to a teen's Library and scheduled for practice; and turning uploaded notes into quizzes is easier. OpenAI will also support Boston Children's Hospital's Digital Wellness Lab and its Student Advisory Council for three years; in the 2026-27 school year about 22 students aged 14 to 20 will test tools and advise on teen safety defaults, parental controls and AI literacy.
Verifiable
Primary page: https://openai.com/index/teens-learn-and-plan (RSS pubDate Wed, 07 Oct 2026 12:00:00 GMT, category Product). All figures are self-reported by OpenAI: 1.2 billion ChatGPT users; in one week nearly 1.2 million teens used Learning Visualizations and more than 180,000 used Study Mode; teens average under 15 minutes a day, fewer than 2% spend more than three consecutive hours, and in almost half of conversations with a break reminder teens paused or stopped within five minutes. The claim that teens with access sent about 2.7 million more learning-related messages on average is not clearly defined; the post gives no period or comparison method. College Planner is described as coming soon, with no date.
Our take
What matters here is less the flashcards than OpenAI starting to argue, with data, that teen AI use is healthy. Setting under 15 minutes a day against roughly 5 hours of social media, and noting that over 80% of long sessions involve learning, is a pre-emptive answer to the addiction and dependence concerns regulators and parents raise most. Every number comes from OpenAI with no third-party audit, so read it as a position, not an independent finding.
College Planner deserves a closer look: it moves ChatGPT from answering questions to managing a real task that runs for months, in a setting with large information gaps where students with fewer resources most need procedural help. It goes to Latest rather than this week's major changes because it is a product update for a specific audience with no new model capability. Two things to watch: whether an independent evaluation measures College Planner's effect on application outcomes, and whether OpenAI actually reports back, as promised, what it changed in response to the student council and why.
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy