Schmidt: AI agents, not AlphaFold, are key to accelerating science

In a provocative op-ed published in MIT Technology Review, former Google CEO Eric Schmidt and Suhas Mahesh of Schmidt Sciences argue that the AI community is overinvesting in the AlphaFold template for scientific discovery. AlphaFold, the DeepMind neural network that won the 2024 Nobel Prize in Chemistry, proved AI can solve grand challenges by predicting protein structures. But that success relied on a dataset of roughly 170,000 experimentally validated protein structures that took 53 years and approximately $21 billion of experimental work to assemble. Comparable datasets simply do not exist in most scientific domains, making this approach impossible to scale broadly.

Instead, Schmidt and Mahesh champion AI agents—systems that model the iterative, highly contingent process of real research. Unlike AlphaFold, which applies a powerful pattern-matching tool to a narrow question, agents are generalists: they can formulate hypotheses, design experiments, interpret results, and adjust course, just like human scientists. This isn't a new way of doing science, they argue, but a digital recreation of the human discovery process. By shifting focus from static datasets to dynamic agentic workflows, science could accelerate dramatically across fields like materials, medicine, and climate, where the data pipeline is far slower than protein crystallography once was.