Trump stands up a 'Super Intelligence Force' led by Jay Clayton
President Trump on Sunday announced a federal "Super Intelligence Force", led by Director of National Intelligence Jay Clayton, days after meeting the heads of OpenAI, Anthropic, Google, Meta, Nvidia and SpaceXAI at the White House. The group will coordinate federal AI work, engage industry, infrastructure providers and public-interest groups, and deliver a report on risks and opportunities within 120 days. It sits on top of a one-page voluntary accord the labs signed, committing to internal evaluations, external audits and board reviews. The implication: Washington is choosing coordination and self-policing over binding rules for now, and the 120-day report becomes the next real marker for US AI policy. CBS News
What else matters
Google pauses OSS bug bounty product reports over AI slop
Google stopped accepting product-vulnerability submissions to its Open Source Software VRP from October 1, citing a significant rise in automated reports, "the vast majority of which are not valid". Supply-chain reports and anything already filed still count, and Google promises an update in Q1 2027 rather than a reopen date. Google Bug Hunters rules
Gemini 4 hits internal skepticism on coding
Ahead of Gemini 4's full launch, Google staff with access say the model scores well on benchmarks but stumbles on real coding work, especially front-end design. Google already abandoned the promised Gemini 3.5 Pro, so a weak coding story would hurt in the market it most needs to win. Bloomberg via NWA Democrat-Gazette
Shanghai AI Lab releases 744B Atria-Dawn-Preview under MIT
Shanghai AI Laboratory published Atria-Dawn-Preview on Hugging Face and ModelScope: a 744B-parameter mixture-of-experts agentic model built on GLM-5.2, with 256K context and FP8 weights. The MIT licence lets companies fine-tune and ship it commercially without permission, so independent evals will decide whether it matters. ChatPicture
Google's RRSI keeps self-improving agents from gaming their tests
Google researchers proposed Regularized Recursive Self-Improvement of agent harnesses, which rewrites prompts, tools and memory around a frozen model without overfitting to optimisation tasks. Reported gains reached 4.7 points on unseen benchmarks, never fell below baseline, and used about 30% fewer tokens than unregularised tuning. Creati.ai
Worth your time
GPT-6 Astra cleared a World of Warcraft starting zone from raw server data, building its own pathfinding helper instead of reading the screen.
