Thore Graepel, who helped build Google DeepMind's AlphaGo, has founded Metis Reasoning in London. The lab's starting premise, set out in a white paper, is that large language models cannot be trusted to reason, and it wants to apply AlphaGo's lessons to general reasoning.
NVIDIA has presented Cosmos3-H-Dreams on the Hugging Face blog, describing it as a way to recreate clinical surgery through generative simulation.
A Hugging Face post introduces FairStress, an approach for evaluating fairness in decisions made by large language models in more depth.
Startup Electric Plant has built an AI-powered translator that interprets the faint electrical signals running through plant stems, aiming to let humans understand what plants are communicating.
A computer program has for the first time outperformed humans at solving cryptic crosswords. The Guardian compares the milestone to IBM's Deep Blue beating Garry Kasparov at chess in 1997.
A Hugging Face post describes ALTK-Evolve, which uses learned memory to reach the top of the τ²-bench leaderboard on customer service tasks without relying on a frontier model.