Anthropic is improving Claude about twice as fast, mostly thanks to Claude, METR finds
METR got inside Anthropic’s data. Its verdict: progress has roughly doubled, mostly thanks to AI, and most of the evidence stays redacted.
METR got inside Anthropic’s data. Its verdict: progress has roughly doubled, mostly thanks to AI, and most of the evidence stays redacted.
The self-hosted interface lets models drive a desktop and a browser, lifts default limits and ships security fixes that are not yet described.
In an OpenAI test, GPT-6 Astra completed about 40% fewer rubric items on tasks framed as AI safety research than on matched ones.
OpenAI describes an agent pipeline that found and exploited 948 Linux kernel flaws and drafted a candidate patch for each one.
Bans, moratoriums and failed rezonings: of 41 local ballot measures on November 3, data centers won only one, in Harwood, North Dakota.
On Linux, an RTX Spark laptop’s GPU stays throttled on wall power. Nvidia just released a driver that adds the interface at issue.
OpenAI’s Images 2.6 brings animated GIFs to ChatGPT for everyone, two API models at unchanged prices, and a safety card with numbers.
Hy4 leaves preview with vision and computer use. Tencent promises the weights with no date and admits it trails the top closed models.
OpenAI and Tencent report AI models designing a device and tuning an open-source processor, and Anthropic reportedly uses Claude on its chips.
Across 60 long stays at Stony Brook University Hospital, doctors preferred the AI-written discharge summary in 95% of cases.