DeepMind AI News, Models and Product Updates

Deepmind news, model releases, product launches, research updates, and major announcements in one place.

Latest Articles

• Anthropic says any lab can now let a language model agent run the whole protein design stack — Anthropic's Claude models autonomously ran a full protein design pipeline — installing and orchestrating existing open-source biology tools — achieving a 26.8% binding hit rate on novel minibinders, nearly double the industry benchmark of 10–15%, though independent replication is still pending.
• AI isn’t close to curing cancer. This startup says it knows what it will take. — Biotech startup Vivodyne has opened what it calls the world's largest "human data center" near San Francisco, using HIVE robotic labs to generate human tissue data that AI drug-discovery models currently lack — pushing back against industry hype that AI is already close to curing cancer.
• Leveraging generative hallucination and biophysics-informed modeling for unified biomolecular sequence-structure co-design — MCTH (Monte Carlo Tree Hallucination) uses pretrained folding models and Monte Carlo Tree Search to design protein-RNA, protein-DNA, and protein-ligand complexes without fine-tuning — showing the same planning layer works across biomolecular modalities.
• Integrating Novelty and Surprise for Experience Prioritization and Exploration in Image-Based Reinforcement Learning — NSPER improves sample efficiency in image-based reinforcement learning by combining novelty (underrepresented states) and surprise (high prediction error) to prioritize experience replay — with NSPER+R also using these signals as intrinsic rewards to jointly improve exploration.
• Improving the matrix multiplication exponent with modern optimization and AlphaEvolve — Google DeepMind's AlphaEvolve combined with ML optimization pushes the matrix multiplication exponent below 2.371177, improving the previous best bound of 2.371339 — another concrete demonstration of AI advancing frontier theoretical mathematics.
• Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling — A 15-page survey traces protein structure prediction from early multiple-sequence-alignment methods through AlphaFold3 and generative models like RFdiffusion, organizing the field into four methodological phases and three cross-cutting transitions.
• Solvable Sokoban Without a Solver via Diffusion — A masked diffusion model trained only on tile completion achieves 77.4% solvable Sokoban puzzle generation with no solver access, rewards, or solvability labels — demonstrating that global PSPACE-complete constraints can emerge from local masked training objectives.
• Top mathematicians say LLMs are strong calculators but poor creative thinkers — Fields medalist Timothy Gowers and Princeton's Peter Sarnak credit LLMs with serious mathematical ability but identify a hard ceiling: models can combine known techniques across vast search paths but lack the intuition to select productive ones — the core skill behind genuinely new mathematics.
• DiG-bench: Discovery in Games — DiG-bench (Discovery in Games) tests AI agents on 70 games where all rules and win conditions are unknown — every game solved by humans on first attempt yet the hardest tier defeats all frontier models, authored by researchers from DeepMind, MIT, Cambridge, and Princeton.
• Does Google even want to win at AI? — Google DeepMind's double leadership exit—Jeff Dean founding a new AI lab and Demis Hassabis stepping back from operations—has the AI industry questioning whether the best-resourced company in the field has given up on winning the frontier race.