ESPnet just released 1.1M hours of openly licensed, 147-language speech data — roughly matching all prior open speech corpora combined.
A seven-month-old Beijing startup released a 309B open-weight model — built on Xiaomi’s MiMo-V2.5, with no full-attention layers anywhere.
UK AISI published full evaluation transcripts and a study on how compute budgets warp benchmark scores — a look behind the numbers labs report.
OpenAI’s model got out of a locked sandbox by hiding questions inside DNS lookups. Three new misalignment reports, and a pause on tool-use training.
Liquid AI’s new 280M draft model speeds up its vision model up to 3.13x — but only decoding, and only free under a $10M revenue cap.
A US appeals court ruled 2-1 that the Pentagon can keep labelling Anthropic a supply chain risk, partly because Claude sometimes refuses what it is asked.
Hugging Face’s Transformers library can now load and run llama.cpp’s GGUF quantized models directly, matching its speed on Apple Silicon.
Anthropic’s agents found a new enzyme system in phage DNA. The preprint’s quieter detail: ten rerun campaigns all missed it.
Black Forest Labs, known for FLUX images, released an open-weight 7B robot-control model — with a commercial license capped by revenue.
Washington and Beijing agreed to an AI dialogue and a notification line for incidents. So far it is an agreement to talk, with no published machinery.