Eric Schmidt and Schmidt Sciences AI lead Suhas Mahesh argue that AI agents, not AlphaFold-style foundation models, will drive the next wave of scientific discovery. In an essay published August 10, the former Google CEO makes the case that the conditions producing AlphaFold — a 170,000-structure Protein Data Bank assembled over 53 years at roughly $21B of experimental cost — are too rare to reproduce across most scientific fields. Instead, they argue, reasoning agents that mimic how human researchers actually work will compress timelines from decades to months.
The framing pushes back on a widely-held assumption after Demis Hassabis and John Jumper of Google DeepMind shared the 2024 Nobel in chemistry for AlphaFold. Hassabis called AlphaFold "the template for how AI can accelerate all of science to digital speed," and billions of dollars have flowed to startups building foundation models for biology, chemistry, and materials discovery on that premise. Schmidt and Mahesh say the template is real but narrow — and that waiting for AlphaFold-scale datasets in every field is the wrong bet.
Their central number is the Protein Data Bank itself. Fifty-three years of international cooperation and $21B in wet-lab work produced the training corpus that made AlphaFold possible. Over 25 Nobel Prizes have relied on protein crystallography, the technique underneath those measurements, which is unusually replicable — most experimental science is not. Cell lines drift, chemicals carry contaminants, humidity moves numbers. Building a comparable dataset in a typical biology or chemistry subfield would require new measurement standards that do not yet exist.
Key facts
- 01Eric Schmidt argues the Protein Data Bank required 53 years and roughly $21B in experimental work — a template few fields can replicate.
- 02Google's AI Co-Scientist, announced in May, reproduced an antibiotic-resistance finding that took Imperial College London researchers a decade of wet-lab work.
- 03Over 25 Nobel Prizes have relied on protein crystallography, the unusually replicable technique that made AlphaFold's training set possible.
- 04In 2024, Demis Hassabis and John Jumper of Google DeepMind won part of the Nobel in chemistry for AlphaFold.
- 05Schmidt co-founded Schmidt Sciences in 2024 with Wendy Schmidt to fund unconventional science and tech research.
A handful of fields do meet the AlphaFold conditions — weather forecasting, much of genomics, limited slices of chemistry — and Schmidt and Mahesh expect breakthroughs there. But for most open questions, the essay argues that the constraint is not compute or model architecture. It is the physical, expensive, decades-long process of generating consistent measurements.
The alternative they lay out is agentic AI. "Simply put, an agent is an AI reasoning engine that has been given access to tools—digital or physical—and the capabilities to use them," the pair write. The claim is that large language models finally allow software to do what practicing scientists have always done under uncertainty: combine imperfect tools, weigh each method's failure modes, and revise as evidence arrives.
The case study is Google's AI Co-Scientist, announced in May. Given a one-page brief on how antibiotic resistance spreads between bacterial species, the system spun up sub-agents to draft hypotheses, critique them, run tournaments to rank candidates, and refine the winner. It concluded that resistance genes hitchhike on bacterial viruses that ferry them between hosts. Researchers at Imperial College London had reached the same conclusion through a decade of wet-lab work, and their paper — unseen by Co-Scientist — was still in peer review when the system produced the answer.
“While tools like AlphaFold apply a powerful approach to a limited question, agents are inherently generalists.”— Eric Schmidt, co-founder of Schmidt Sciences and former Google CEO
Schmidt and Mahesh point to three compounding effects if agents become routine. First, reproducibility: agents automatically log every action they take, creating exact records that human researchers have long resisted producing manually. Second, institutional memory: a lab's full experimental history gets captured in a standardized repository instead of scattered across graduate students' notebooks. Third, and the one they emphasize most, speed.
The economic argument is that when running an experiment costs less than debating it, debate collapses. An agent that can read a thousand papers in an hour, design 500 candidate molecules, and iterate overnight changes what questions researchers are willing to ask. The essay claims this will pull scientists toward stranger, riskier hypotheses they would not have spent months chasing under current cost structures.
The caveats are substantial and the authors name them. Agents still hallucinate. Their judgment is inconsistent run to run. Memory windows and input limits constrain how long they can operate autonomously. Schmidt and Mahesh bet those technical barriers fall, but they concede the timeline is unproven, and the essay does not engage with recent findings on multi-step agent reliability — an area where even frontier labs post modest numbers on long-horizon tasks.
“In any field, when testing an idea takes less time than arguing about it in a meeting, people stop debating and just run the test.”— Suhas Mahesh, AI for Science lead at Schmidt Sciences
The piece also arrives against a specific commercial backdrop. Foundation-model startups in biology, chemistry, and materials have raised billions on the AlphaFold template Schmidt is now downgrading. If agentic pipelines built on general-purpose reasoning models — the kind OpenAI, Anthropic, and Google already sell — can do most of what specialized science foundation models promised, the addressable market for vertical science-model companies narrows considerably. That is not a neutral read from a Google co-founder whose philanthropic vehicle funds AI-for-science work.
The framing matters for how research dollars get allocated over the next several years. If Schmidt is right, the highest-leverage AI investments in science are agent infrastructure, tool access, and evaluation harnesses — not another decade-long push to assemble a Protein Data Bank for every subfield. That thesis is testable. Co-Scientist reproducing a decade of Imperial College work is one data point; the field needs many more before betting the funding stack on it. But the bet is now on the table, from someone who ran Google from 2001 to 2011 and is putting Schmidt Sciences money behind the answer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
