Latest AI Breakthroughs

Open medical AI models built on Gemma 3 beat their general-purpose base across medical images and text

Asked to rate over 4,400 AI-generated ideas, 100-plus group leaders called a quarter interesting

Children, adults and AI agents in a Minecraft-like game seem to seek novelty first, then control

A language-model system simulating a panel of physicians correctly diagnosed 83.2% of NEJM case-conference cases in a benchmark, against 36.1% for generalist physicians.

One training recipe gives a humanoid cartwheels, spin kicks and sprinting, and one distilled model handles new tasks

Simulated networks of 1,000 LLM agents talking politics drifted into polarized, like-minded camps

A multi-agent AI system on Gemini that generates and validates scientific hypotheses

A new algorithm could make 2020 Census tables more accurate without weakening their privacy protection

Pretrained on 500 million simulated collisions, one model handles new searches and rediscovers a known particle

A framework using Kolmogorov-Arnold networks to discover symbolic physical laws from data

An RL agent writes machine-checkable Lean proofs for competition mathematics.

A multimodal whole-slide foundation model for pathology that works without fine-tuning

A GPT-style model forecasts more than 1,000 diseases decades ahead and generates synthetic health records.

An open model develops self-reflection and verification with no human reasoning examples.

An AI model generates spatial proteomics from H&E slides to improve lung cancer prognostic accuracy

A two-level, language-guided policy autonomously completes the clip-and-cut phase of gallbladder removal and corrects its own mistakes.

An evolutionary coding agent finds new, provably correct algorithms and optimizes Google's own infrastructure.

A foundation model for spreadsheets outperforms tuned gradient-boosted trees on small data.

A blinded trial pits a medical language model against 20 primary care doctors on text consultations.

A vision-language-action model drives real robots through dexterous, open-vocabulary tasks.