Is Anthropic’s AI Development Speeding Up? What METR Found
METR estimates Anthropic improved about twice as fast on well-specified AI tasks, with substantial uncertainty. Here is what the review does and does not show.
METR estimates Anthropic improved about twice as fast on well-specified AI tasks, with substantial uncertainty. Here is what the review does and does not show.
Google ran a Gemma 4 model aboard a prototype satellite. What the haiku test proves, and what orbital AI infrastructure still has to solve.
Google Genie 4 creates interactive worlds with audio and real-time steering. Who can test it now, what can it do, and where does it still fail?
Block demonstrated a research voice model that can listen while speaking and generate expressive speech. Dessa is not publicly available, and its performance claims await independent testing.
MIT-led researchers let 13 AI agents explore a virtual island for 30 hours. Their behavior resembled play, but the study does not prove sentience.
AI helped historians analyze 100,000 pages of port books. The records reveal surprising European trade links, lost ports and unusual cargo.
A Harvard survey discussed by WIRED found differences in AI familiarity, workplace training, and media visibility. Here is what the results do and do not show.
Reflection AI has launched Beam, a 501B-parameter open-weight model for coding and AI agents. Here is how its claims stack up against DeepSeek, Kimi, GLM and Qwen.
Google DeepMind’s SynthID Bio adds detectable watermarks to AI-designed proteins. Early lab results suggest the marks can preserve performance, but
OpenAI says GPT-6 Sol and Luna reduce API costs and errors. Examine the launch claims and what developers should test in their own workflows.