- regulationThe White House restored the UK AI Safety Institute’s access to pre-release frontier models from Google, OpenAI and Anthropic. Ministers said the ban threatened the UK’s ambition to lead global AI regulation.
- amazonAmazon is negotiating to buy Decart for about $7B after Anthropic abandoned a $6B bid. Decart’s chip-switching software could reduce Nvidia dependence, boost Trainium adoption and cut cloud AI costs.
- anthropicAnalysts estimate Anthropic’s 245 MW data-center site could generate about $1.3B a month, while sources say it signed a cloud deal with IREN. The figures highlight the economics behind its $530B in leasing deals.
- anthropicAnthropic will ban cruelty toward Claude from Nov. 12 but has not said what the rule protects. Enforcement will use the same conversation-ending tool Claude received in August 2025.
- agentSakana AI and UC Berkeley’s MASS framework organizes copies of one model into virtual subagents for research tasks. On a 27B open weights model, it raised scores per output token 1.2–1.6x across four benchmarks, but may reinforce blind spots.
- openaiOpenAI’s release of more than 700 mathematical manuscripts, including 16 problems studied by Hugo Duminil-Copin, is pushing researchers to reconsider public openness. Northwestern students are questioning the field, while AGMAI says its advisory role was not an endorsement.
- safetyMI5 chief Ken McCallum told the UK Cabinet on Sept. 8 that AI is a major national-security threat, especially from Russia, Iran and terrorist groups. A test found 407 vulnerabilities across nine government bodies for £13,000, including one critical flaw.
- agentOdyssey released Odyssey-3 as a free public research preview. The 14B world model generates interactive environments from text in real time and reportedly controls robots, drones and GTA V, though its physics benchmark results have methodological limits.
188 storiesUpdated 36m ago