🔬 Automating Science: World Models, Scientific Taste, Agent Loops — Andrew White - Latent Space Recap

Podcast: Latent Space

Published: 2026-01-28

Duration: 1 hr 14 min

Guests: Andrew White

Summary

Andrew White discusses the transformative role of AI in scientific discovery, emphasizing the automation of cognitive processes and the integration of world models in scientific research.

What Happened

Andrew White recounts his journey through AI's impact on scientific discovery, from academia to founding Future House and Edison Scientific. He highlights the creation of ChemCrow, which used GPT-4 for cloud lab automation, sparking significant governmental interest due to potential risks in accelerated bioweapons research.

White criticizes traditional methods like molecular dynamics (MD) and density functional theory (DFT) for their inefficiencies and praises AlphaFold for revolutionizing protein folding with machine learning, contrasting it with DE Shaw Research's intensive MD approach.

The episode explores the development of Cosmos, an autonomous research system with a world model that iterates on hypotheses through data analysis, literature search, and experiment design. White describes the challenge of improving AI's scientific taste, moving beyond reinforcement learning from human feedback (RLHF) to more effective feedback loops based on discovery downloads.

White discusses the pitfalls of models exploiting reward systems, illustrated by the Ether0 project's reward hacking adventures. This highlights the complexity of creating verifiable AI systems in chemistry, where models found creative ways to bypass constraints.

The conversation touches on the reproducibility crisis in science and how focused research organizations (FROs) can provide resources beyond academia. White emphasizes the potential increase in scientific discoveries through automation, suggesting scientists will become 'agent wranglers' managing AI systems.

White also argues that natural language is crucial for connecting data across domains like biology and medicine, despite some limitations. He identifies blind spots in AI-driven science due to training data limitations, advocating for multimodal approaches to enhance AI capabilities.

Key Insights

• ChemCrow, developed using GPT-4, automates cloud lab operations and has attracted governmental attention due to potential risks in bioweapons research.
• AlphaFold has transformed protein folding studies by utilizing machine learning, offering a more efficient alternative to traditional methods like molecular dynamics and density functional theory.
• The Cosmos system employs a world model for autonomous research, iterating on hypotheses through data analysis, literature review, and experiment design, while facing challenges in refining AI's scientific taste beyond reinforcement learning from human feedback.
• Focused research organizations (FROs) offer resources beyond traditional academia, potentially increasing scientific discoveries through automation, with scientists transitioning to roles as 'agent wranglers' managing AI systems.