Eric Schmidt, the former Google CEO, is making a provocative bet on the future of scientific discovery, and it’s not just about feeding more data into bigger models. In a new op-ed for MIT Technology Review, Schmidt and Suhas Mahesh, who leads AI-for-science work at Schmidt Sciences, argue that the AlphaFold playbook—spectacular as it was—is a dead end for many fields. The problem? AlphaFold’s Nobel Prize-winning breakthrough relied on a painstakingly assembled dataset of roughly 170,000 protein structures. That effort took 53 years and an estimated $21 billion. Replicating that kind of curated, gold-standard data in other complex disciplines like materials science or climate modeling is simply not feasible.
The alternative they’re championing is AI agents. Unlike tools that excel at a single, narrow prediction task, agents are generalists designed to digitally model the messy, iterative process of human research. They don’t just recognize patterns in a static dataset; they could formulate hypotheses, design experiments, and adjust their approach based on results—much like a flesh-and-blood scientist. It’s a shift from treating AI as a powerful calculator to treating it as a junior colleague who learns by doing. This isn’t a new way to do science, Schmidt and Mahesh contend, but a digital mirroring of the discovery process itself.
The timing matters. We’re in a moment of extreme hype around AI’s scientific prowess, fueled largely by AlphaFold’s success. But Schmidt’s push for agents is a reality check couched in optimism. It acknowledges that the low-hanging fruit of massive, clean datasets has mostly been picked. For AI to genuinely accelerate discoveries in areas where data is sparse or incredibly expensive to generate, it needs to get comfortable with uncertainty and learn to reason its way through problems. The $21 billion protein dataset isn’t a blueprint the world can afford to repeat. It’s an anomaly that agents could help us move beyond.
