AI Agents, Not Foundation Models, Will Accelerate Scientific Discovery

Eric Schmidt argues in MIT Technology Review that the AlphaFold template is the exception, not the rule — AI agents that mimic the iterative process of research are the real accelerant for science. Google DeepMind's Co-Scientist is already proving the case.

When AlphaFold won the Nobel Prize in Chemistry in 2024, it seemed to promise a new template for science: feed enough data into a neural network and watch it crack half-century-old problems. But Eric Schmidt and Suhas Mahesh argue in MIT Technology Review that this template is a mirage for most fields — the conditions that produced AlphaFold are vanishingly rare.

The Protein Data Bank took 53 years and roughly $21 billion to assemble. Most scientific domains lack anything comparable, and the experimental techniques that produce consistent, scalable training data — like the protein crystallography behind AlphaFold — simply don't exist in cell biology, chemistry, or materials science.

The real accelerant, they argue, is AI agents: systems that don't just predict but reason. These agents mimic the messy, iterative process of actual research — generating hypotheses, debating them, revising conclusions as evidence comes in.

Google DeepMind's Co-Scientist system, published in Nature in May, demonstrates the approach. Given a one-page brief on antibiotic resistance, the system spun up sub-agents that drafted hypotheses, peer-reviewed each other, and ran tournaments to rank the strongest candidates. It correctly concluded that resistance genes hitch rides on bacterial viruses — a finding that Imperial College London researchers spent a decade reaching through wet-lab work. Their paper was still in peer review when Co-Scientist independently reached the same conclusion.

Stanford's Gary Peltz used Co-Scientist to identify overlooked drug-repurposing candidates for liver fibrosis; one blocked 91% of a scarring-linked response in lab tests. MIT's Ritu Raman leveraged it to bridge expertise between labs working on ALS.

Agents also offer structural fixes for science's deeper problems: they automatically log every move, creating exact records for replication — a direct answer to the reproducibility crisis. They amplify institutional memory, recording decades of lab knowledge in searchable repositories. And they collapse the cost of experimentation, letting researchers chase bold questions they'd never risk time on before.

"Historically, tools of such scope have arrived just a handful of times: calculus, statistical inference, spectroscopy, the computer," Schmidt and Mahesh write. "With agents, another such transformation is upon us."

Sources: MIT Technology Review / Google DeepMind