Every few decades experts declare that science has run its course. In the early 20th century Albert A. Michelson wrote that much of physical science’s facts had already been discovered; later, Stephen Hawking speculated about the possible end of theoretical physics. The arrival of powerful artificial intelligence has revived such sentiments, now amplified by Nobel-recognized work.
In 2024, Demis Hassabis and John Jumper of Google DeepMind received part of the Nobel Prize in Chemistry for AlphaFold, a neural network that predicts three-dimensional protein structures by learning from tens of thousands of experimentally determined shapes. The problem AlphaFold addressed had resisted systematic solutions for decades, and the model’s success led many to view its approach as a template for accelerating science. Its impact also spurred a wave of startups building foundation models for biology, chemistry, and materials discovery, attracting billions in funding.
Why AlphaFold succeeded — and why that success is rare
A primary condition enabling AlphaFold was the existence of the Protein Data Bank, an archive of roughly 170,000 experimentally validated protein structures that provided training data. The Protein Data Bank did not appear overnight: building it required 53 years of international scientific cooperation and, by a recent estimate, about $21 billion worth of experimental work. Projects of that scale are notoriously difficult to fund and coordinate, and frequently fail to materialize.
Even where cohesion and resources exist, another obstacle is less often discussed: the scientific impracticality of producing comparable data in many fields. Protein crystallography, the key experimental method for many structures, is unusually reproducible and dependable — more than 25 Nobel Prizes have relied on it. But results in much of experimental science vary: cell lines drift, chemicals contain trace contaminants, lab humidity shifts. Creating datasets that are consistent, accurate, precise, and scalable enough to train modern neural networks across most areas of biology or chemistry would demand new measurement types and standardization practices — none of which are likely to appear quickly.
Fields where AlphaFold-like breakthroughs may come first
A few domains already meet these data requirements: weather forecasting, much of genomics, and narrow areas of chemistry. These fields may see AlphaFold-style advances soon, if they have not already. Government support for producing and coordinating such datasets will be critical, as the US National Security Commission on Emerging Biotechnology has argued.
A complementary path: reasoning agents
For most open scientific problems, a different near-term plan is required. Scientists have always reasoned under uncertainty: identifying drug targets or other phenomena involves combining docking calculations, known structures, molecular dynamics, a handful of binding assays, and expert judgment. The craft of science lies in synthesizing many imperfect tools and updating conclusions as new evidence arrives. Until recently, no software could emulate that process end-to-end.
That is changing with the emergence of AI agents. An agent is an AI reasoning engine given access to tools — digital or physical — and the capability to use them. A recent architectural shift around large language models has enabled these programs to proliferate rapidly and reduced the need for domain-specific, large-scale datasets. For science, agents represent a foundational change: they can digitally model the iterative, contingent workflow of actual research. Unlike specialized models such as AlphaFold, agents are generalists that mimic human processes of discovery.
Example: Google AI Co-Scientist
Google’s AI Co-Scientist, announced in May, illustrates the potential. Researchers provided a one-page brief and a goal: determine how antibiotic resistance spreads between bacterial species. The system spun up sub-agents: one drafted hypotheses from the literature, another critiqued them like a peer reviewer, a third ran tournaments to rank candidates, and a fourth refined the winning hypothesis. The agent concluded that resistance genes hitchhiked on bacteriophages — viral vectors that can move genes into new hosts. That hypothesis proved correct; researchers at Imperial College London had reached the same conclusion after roughly a decade of wet-lab work, and their paper (unknown to Co-Scientist) was still in peer review.
Technical challenges and structural benefits
Agents remain novel and face real challenges: they can hallucinate, their judgment can be inconsistent, and memory and input limits constrain how long they can run autonomously. These technical barriers are likely to diminish over time, however, and as they do we should expect compounding effects on the reliability, consistency, and speed of scientific work.
Structurally, agents offer remedies for the reproducibility crisis. For decades the scientific community has urged researchers to share raw data and exact code so experiments can be standardized and repeated, but researchers often resist the tedious post-hoc documentation. Agents automatically log every action they take, producing exact records of methods that enable precise replication.
Agents also amplify scientific memory. Knowledge transfer between researchers is often messy: unless learned through years of apprenticeship, trainees must sift through incomplete lab notebooks. As agents become integral to laboratory work, a lab’s entire scientific history can be captured in a centralized, standardized institutional repository.
Most importantly, agents accelerate experimentation. When testing an idea becomes faster than debating it, researchers will test rather than argue. An agent that can read a thousand papers in an hour, design hundreds of candidate molecules, and learn from failed tests overnight lowers experimental costs and changes research tempo. It will free scientists to pursue bolder, riskier questions and open doors to previously unimagined problems.
Conclusion: agents will complement, not replace, dataset-driven breakthroughs
AlphaFold-style solutions will remain crucial for specific breakthroughs, but by themselves they will not bring about a final era of science. The rise of agentic AI, however, could be a rarer kind of transformation: a toolset that integrates across fields and reshapes how science is practiced. Historically, similarly sweeping tools — calculus, statistical inference, spectroscopy, the computer — revealed new classes of problems and redefined entire disciplines. With AI agents, another such transformation appears to be beginning.
Authors and affiliations
Eric Schmidt was CEO of Google from 2001 to 2011. In 2024 he and his wife Wendy co-founded Schmidt Sciences, a philanthropic venture funding unconventional exploration in science and technology. Suhas Mahesh leads AI for Science work at the AI Center of Schmidt Sciences and specializes in AI for materials discovery. Additional research was provided by Maya Levin, associate and sciences lead, Office of Eric Schmidt.
