🔬 Automating Science: World Models, Scientific Taste, Agent Loops — Andrew White

What It Covers

Andrew White, cofounder of Future House and Edison Scientific, discusses his transition from academia to automating scientific discovery using AI agents. He covers the development of Cosmos, a system that automates hypothesis generation, literature research, data analysis, and experimental design. White explains how language models can accelerate science through enumeration and filtering rather than pure intelligence.

Key Questions Answered

• •AlphaFold's Unexpected Efficiency: Protein folding was solved not through specialized hardware like DESRES's custom silicon MD computers, but through machine learning on experimental X-ray crystallography data running on standard GPUs. This demonstrates that empirical data-driven approaches can outperform first-principles simulations by orders of magnitude, requiring only approximately 10,000 GPU hours to train versus massive specialized infrastructure.
• •Scientific Taste as the Frontier: Human experts agree only 70% of the time on data analysis interpretations, matching current AI performance on bioinformatics benchmarks like BixBench. The bottleneck in automating science is not intelligence for proposing experiments but capturing scientific taste—understanding which hypotheses lead to impactful discoveries versus boring results. This requires end-to-end feedback loops where downstream experimental success informs hypothesis quality.
• •Enumeration Over Intelligence: AI agents succeed in science by trying more hypotheses faster and filtering through literature search and data analysis rather than being smarter. In the Robin paper on age-related macular degeneration, the hypothesis human experts ranked highest was not the one that led to discovering Ripasudil as an effective treatment, demonstrating that verifiable rewards outperform human intuition.
• •World Models as Scientific Memory: Cosmos uses world models as a distillation mechanism similar to Git repositories—accumulating and organizing information over time while enabling predictions. This differs from simple memory or literature databases by being operational and updatable through experimental loops. The data analysis agent in the loop enables real exploration versus literature-only approaches that failed to provide actionable feedback.
• •Natural Language as Universal Interface: Natural language serves as the only representation that bridges all scientific data types—code, papers, population data, molecular structures—because humans continuously innovate language to represent all known observations. While abstractions like graphs or geometry matter, language sits at the boundary between abstract enough to be practical and concrete enough to be useful, avoiding the infinite regress of simulation detail.
• •Jevons Paradox in Science: Automating science will not displace scientists because scientific discovery has unlimited appetite unlike finite tasks like driving. Scientists will become agent wranglers exploring 100 ideas simultaneously rather than one at a time. The demand for science will match automation capacity since there is no fixed number of discoveries to make, though short-term friction exists in R&D hiring decisions.