Topic: Research

A frontier model’s training method clears formal scientific review.

General reasoning models solve olympiad problems without task-specific tooling.

A Chinese lab ships frontier reasoning trained by reinforcement learning on verifiable rewards.

Baker, Hassabis and Jumper are recognised for protein structure work.

Neural network foundations are recognised by science’s highest honour.

Structure prediction covers DNA, RNA and drug-like molecules.

DeepMind shows leading models are badly undertrained for their size.

DeepMind predicts protein structures at roughly experimental accuracy.

DeepMind’s system wins the Wuzhen match against the top-ranked human Go player.

DeepMind trains one network to play many games from raw screen images and score alone.