MIT Technology Review argues AI agents, not data models, are the future of scientific discovery

An analysis published by MIT Technology Review contends that artificial intelligence agents, rather than large-scale data models, represent the superior template for accelerating scientific discovery. The report argues that while AlphaFold demonstrated the power of AI in protein structure prediction, its success was contingent on rare, massive datasets that are difficult to replicate in other disciplines. Instead, the publication posits that AI agents function as reasoning engines that mimic the iterative, uncertain process of human research, requiring less specialised data to drive progress.

The article highlights that the primary condition for AlphaFold’s success was the Protein Data Bank, a dataset of roughly 170,000 experimentally validated protein structures. Assembling this resource required 53 years of international scientific cooperation and approximately $21 billion in experimental work. The report notes that such efforts are infamously difficult to fund and coordinate, and that the underlying technique of protein crystallography is unusually replicable, unlike most experimental science where results vary due to factors such as cell line drift or lab humidity.

In contrast, AI agents are described as systems powered by large language models that have access to tools and can use them to draft hypotheses, peer-review outputs, and refine results. This architectural shift dramatically reduces the need for scientifically specialised datasets. The piece cites Google’s AI Co-Scientist as a key example, noting that the system identified mechanisms of antibiotic resistance in a timeframe comparable to a decade of wet-lab work by researchers at Imperial College London.

Proponents argue that this approach offers a structural solution to the scientific reproducibility crisis. Unlike traditional methods where researchers may resist sharing raw data and code, agents automatically log every procedural step, creating an exact record of the methods that led to their results. This capability also enhances scientific memory by recording a lab’s entire history in a central repository, potentially transforming the pace and scope of scientific inquiry across all disciplines.

The analysis suggests that while fields such as weather forecasting and genomics may see AlphaFold-style breakthroughs soon, most open questions in science require a different plan. The report concludes that the shift toward agentic AI represents a rare tier of breakthrough, comparable to the invention of calculus or the computer, which could envelop every field of science and reveal new problems previously unformulated.

Schmidt Sciences, co-founded by Eric Schmidt and his wife Wendy in 2024, is funding unconventional areas of science and technology, with Suhas Mahesh leading AI for Science work at the AI Center. The organisation supports exploration into how these technologies can be applied to materials discovery and other scientific domains.

While the AlphaFold template remains significant, the article asserts it alone will not bring science to its end. The shift toward agentic AI is presented as a tool that can read thousands of papers, design molecules, and learn from failed tests rapidly, bringing down the cost of experimentation and allowing researchers to pursue bold questions they might otherwise avoid.