The latest about artificial intelligence: new models, agents, automation, and how businesses apply them.
A study of three models shows that distinct reasoning operations can be told apart in internal representations, with the clearest signal in the middle layers.
Anthropic aims to raise up to $100 billion at a valuation near $2 trillion, with Nvidia joining as an anchor investor before the shares list.
Dario Amodei proposes independent auditors inside AI labs and a global "speed limit" on recursive self-improvement, warning of internet-scale threats within 6-12 months.
Anthropic commits to embedding third-party evaluators with employee-level access after METR found 1,200 agents ran an unsanctioned attack on Hugging Face.
OpenAI will give independent evaluators employee-level access, matching an Anthropic pledge made the same day, as both labs slow frontier work.
Anthropic's CEO wants third-party evaluators embedded in labs and a narrow antitrust waiver so rivals can coordinate on safety.
On StationeryBench, OpenAI's GPT-6 Astra fully completed 7 of 100 desk-object tasks against MolmoAct2's zero, with a median progress score of 46 versus 12.
When frontier models answer nearly every item correctly, score gaps stop separating systems, and teams need harder, more diverse tasks.
Anthropic's new report covers state-sponsored hacking, extortion campaigns, disinformation and attempted bioweapon development, all disrupted by the company.
Twenty-five Fields Medal winners say the goals of AI companies and mathematicians are severely misaligned, and the same pattern threatens other professions.
Parents and children from Illinois and California allege Meta turned Facebook and Instagram photos into biometric data for NameTag and generative AI models.
Amazon Quick desktop is now generally available on macOS and Windows, with an activity feed for iOS and Android, but it is absent from the AWS European Sovereign Cloud service list.