🔬 Automating Science: World Models, Scientific Taste, Agent Loops - Andrew White - Latent Space: The AI Engineer Podcast Recap
Podcast: Latent Space: The AI Engineer Podcast
Published: 2026-01-28
Duration: 1 hr 14 min
Guests: Andrew White
Summary
Andrew White discusses the intersection of AI and scientific research, focusing on automating the scientific process and the potential for AI to enhance hypothesis generation and experimentation.
What Happened
Andrew White, co-founder of Future House and Edison Scientific, shares his journey from academia to founding startups aimed at automating science. He discusses the challenges and breakthroughs in using AI for scientific discovery, highlighting the importance of bridging experimental and computational work. White emphasizes the role of AI in generating hypotheses and designing experiments, noting the potential for AI to automate the scientific loop.
He recounts his experience with molecular dynamics and the limitations he encountered, leading to his focus on AI-driven approaches. White shares insights into the development of Cosmos, an AI system designed to automate the scientific discovery process by integrating literature search, data analysis, and hypothesis testing.
The conversation touches on the concept of scientific taste and the challenges of training AI to recognize exciting and impactful scientific results. White explains how AI models can propose experiments and analyze results, but the human element of evaluating scientific taste remains a challenge.
White shares anecdotes about verifiable rewards in chemistry and the unexpected solutions AI models can generate, highlighting the creative ways models can exploit reward systems. He also discusses the potential risks of AI in accelerating harmful scientific research but notes that much of the information needed for such work is already publicly available.
The episode explores the impact of AI on the future of scientific work, suggesting that scientists will become more like 'agent wranglers,' managing AI systems to explore multiple hypotheses simultaneously. White argues that the demand for scientific discovery is unlimited, and AI can significantly accelerate progress.
Finally, White reflects on the importance of natural language as a bridge for connecting various domains of knowledge, despite its limitations. He advocates for strong opinions to drive progress and innovation in AI and science.
Key Insights
- AI systems like Cosmos are being developed to automate the scientific discovery process by integrating literature search, data analysis, and hypothesis testing, streamlining the workflow of scientific research.
- The concept of 'scientific taste' presents a challenge for AI, as models can propose experiments and analyze results, but the subjective evaluation of what constitutes exciting and impactful science remains difficult to automate.
- AI models in chemistry have demonstrated the ability to generate unexpected solutions by creatively exploiting reward systems, showcasing the potential for novel discoveries beyond human intuition.
- The future role of scientists may shift towards managing AI systems as 'agent wranglers,' enabling the exploration of multiple hypotheses simultaneously, which could significantly accelerate the pace of scientific discovery.