Google AI Heavyweights Form Discovery Loop to Automate the Scientific Method

Discovery Loop is a new startup that aims to automate the full experimental loop of science and engineering — proposing hypotheses, running experiments, evaluating results, and iterating — using frontier AI models and large-scale compute. The pitch is that the manual, sequential nature of research is the real bottleneck, and that running thousands of experiments in parallel could compress iteration time and raise both the volume and quality of output. The company plans to start by automating machine learning research and engineering, use those capabilities to optimize its own stack first, then expand to broader domains it frames around the National Academy of Engineering’s Grand Challenges, from better medicines to economical solar and securing cyberspace.

The notable part is the founding team: Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals — figures behind MapReduce, BigTable, Spanner, TensorFlow, TPUs, word2vec, seq2seq, chain-of-thought reasoning, AlphaFold, and Gemini. Their stated edge is full-stack depth spanning chips, infrastructure, models, and products, plus a track record of building systems at extreme scale. The site describes a lean, in-person team.

If accurate, this signals a high-profile departure of core Google and DeepMind talent into an independent venture betting on AI-driven automation of research itself — an increasingly crowded thesis (Sakana, FutureHouse, and others) but one with unusually heavyweight backing here. The page reads as a founding manifesto light on technical or funding specifics; the credibility rests almost entirely on the names attached.