Jeff Dean's Discovery Loop bets billions on AI-driven science

Jeff Dean, Google's legendary AI heavyweight, just left the company to launch Discovery Loop, a startup focused entirely on automating science. His vision is a generalized system that performs the entire cycle of scientific discovery—hypothesis, experiment, evaluation—across any domain, largely unsupervised. The core loop is propose-run-evaluate, executed thousands of times in parallel, with billions of dollars flowing into this approach. It builds on AlphaFold's Nobel-winning success but aims for something far broader: a universal scientific engine rather than a domain-specific tool.

However, a growing counterargument holds that non-physical AI has a hard ceiling. An article from Chaotropy argues that AI without sensory or motor interfaces can only reason over static measurements and can never run physical experiments. DeepMind's Tom Zahavy similarly claims LLMs can't 'jump'—they lack the manipulative abduction needed to generate genuinely new axioms, confining them to optimizing within existing frameworks. The article pushes back, citing Darwin's theory of natural selection as proof that transformative science can emerge from deep reasoning over existing data. Even without embodiment, there's likely enormous untapped mileage in the data we already have, making Dean's bet and the broader pursuit of automated science far from futile.