2026-10-12 digest
19 items on this page: 17 with their own page, 2 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
NVIDIA opens Vera Rubin rack and compute tray designs to OCP (MGX 1.2, up to 250kW per rack) and adds a Vera CPU rack spec for AI agents; Meta adopts Spectrum-6 as four firms publish a grid resilience spec
Bill Gates and Microsoft president Brad Smith say AI self-regulation is no longer enough; Smith: 'your washing machine is subject to more safety rules than AI'
Qwen introduces QwenGyre, elastic RL for agents whose rollouts approach 1M tokens, with up to 1.85x speedup; the paper puts flagship Qwen 3.8 at 2.4T parameters
Microsoft bets on its own silicon: Azure Maia to scale to multi-gigawatt deployments from 2027 as its primary inference system; Maia 300 claims 40% more tokens per megawatt than leading GPUs
UK unveils AI risk package: wider frontier-AI cyber defence pilots, possible legal reporting of AI incidents and near misses; government confirms AISI is evaluating Gemini 4 Argon
Codex Day 7: GPT-6.1 Sol default context in Codex and Work rises from 272k to 500k, about 80% more
LeCun's AMI Labs open-sources H-JEPA, a hierarchical world model that plans in abstract latent space and lifts maze success from 18% to 73%
WSJ: Amazon in talks to buy Decart for about $7 billion, a month after Anthropic walked away at $6 billion; neither company has confirmed
The Information: Anthropic has an undisclosed multi-year cloud deal with former bitcoin miner IREN; neither company has confirmed
Coinbase's fraud agent test: newer Opus, Sonnet and GPT versions caught less fraud than their predecessors, and a post-trained Qwen3.5-9B beat Opus 4.5
Sakana AI and UC Berkeley release MASS: a 27B model acts as its own team, optimizer and judge for two rounds of recursive self-improvement
Telegraph: White House reverses course, UK AI Security Institute regains pre-release access to Anthropic, OpenAI and Google DeepMind models
Karpathy: the gap in understanding AI is now a funnel, with only ~5,000 people using internal frontier systems
Hugging Face Bio open-sources Carbon-A, a 1.2B gene-finding model, and a database of 566 million predicted coding loci across 22,617 species; Iso-Seq checks in cat, chicken and Arabidopsis back some genes missing from RefSeq
China's Party Central Committee and State Council issue opinions on new quality productive forces: full 'AI+' action, AI in cars, phones, PCs and humanoid robots, a national compute network, and AI risk monitoring and emergency response
White House Accord on Superintelligence: Google, Anthropic, Meta, OpenAI, xAI and Nvidia sign voluntary four-layer controls and audits; same-day executive order renames AI 'Super Intelligence' in federal documents
US senators propose an AI Safety Board with 45-day pre-release access to frontier model weights; Cruz blocks fast-track passage
Short items
Harness Learning paper: a 4B model trained to rewrite an agent's harness code beats its 35B teacher on average across 21 unseen reasoning tasks
【Key points】
On Sept 28, Alvin Zhang, Ruslan Salakhutdinov, Daniel Khashabi, Yuda Song, Andrea Zanette and co-authors posted "Harness Learning Enables Generalizable Test-Time Adaptation" (arXiv 2609.35738). It treats an agent as model plus harness, the code that organizes model calls, tools and information flow, and trains a proposer model with RL to rewrite a solver's harness from execution feedback. At test time no weights change; only the code does.
After DAIR.AI's weekly list on Oct 11, Pedro Domingos posted "the model swallows the harness" (~353K views) and Elon Musk replied that Tesla FSD's network likewise ate the C++ code (~246K views).
【Verifiable】
Paper: https://arxiv.org/abs/2609.35738
On Reasoning Gym (Qwen3.5-4B proposer and solver, SFT on a 35B teacher's revisions, then RL), mean single-step revision score on 21 unseen families rises from 0.32 to 0.62, above the teacher's 0.56; under oracle best-of-eight the teacher still wins. A proposer trained only on HotpotQA transfers to MuSiQue and 2WikiMultihopQA; on MuSiQue, exact match is 0.15 for independent revisions versus 0.27 after ten rounds.
X: Domingos https://x.com/pmddomingos/status/2109420932955804040 , Musk https://x.com/elonmusk/status/2109468545054880123 ; ~150 related posts in 24h, 24-33 per hour over the last five hours.
【Our read】
1. "The model swallows the harness" is half backwards. The harness is not folded into weights; another model rewrites it, and it stays readable, reviewable code. That is the opposite of FSD replacing code with a network, and better for auditing.
2. The novelty is making harness revision a trainable skill that transfers to unseen tasks, on the same line as Sakana's MASS (e/820).
3. Evidence is small: 4B/8B models on puzzles and QA, not frontier coding agents. Whether it holds on long real tool tasks decides if this is a method shift or just a good paper.
Google pauses product vulnerability submissions to its open-source bug bounty after a surge of mostly invalid automated reports
【Key points】
On October 1 Google's official bug bounty account, @GoogleVRP, said it is temporarily no longer accepting product vulnerability submissions to its Open Source Software Vulnerability Reward Program (OSS VRP), which covers Google-maintained projects such as Go, Angular and Protocol Buffers. The reason given: "a significant rise in automated submissions, the vast majority of which are not valid." Supply chain reports and reports filed before October 1 are unaffected. Google committed to an update in Q1 2027 but gave no reopening date.
【Verifiable】
The announcement is in the @GoogleVRP post of October 1; the program rules page later carried the same notice, and SecurityWeek, BleepingComputer and The Hacker News covered it from October 5. Google gave no figures and did not explicitly say the submissions were AI-generated, only "automated." Per Google's program page, product vulnerabilities in flagship projects paid $500 to $7,500, and supply chain compromises up to $31,337. According to The Hacker News, Google began requiring stronger proof in some tiers in March 2026, and the Go project added a section on LLM-generated reports to its security policy in early September. According to SecurityWeek, Google changed Chrome and Android rewards in May, and HackerOne's Internet Bug Bounty paused new submissions in March.
【Our read】
The point is not that one bounty program paused, but that AI has made vulnerability reports cheap enough to break bounty economics: filing costs almost nothing, while verifying still costs maintainer time, and rewards encourage volume. Keeping supply chain reports open suggests the flood is in the category tools can guess at in bulk. Watch whether programs move to proof machines cannot easily fake, such as working reproductions or merged patches, and whether vendors' own scanners cut false positives. Until then, the gap between finding bugs at scale and fixing them falls on maintainers.
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy