Science has a recurring habit of declaring itself nearly finished. In 1903, physicist Albert Michelson suggested that the fundamental facts of physical science had already been discovered. Decades later, Stephen Hawking speculated that theoretical physics might wrap up by the end of the twentieth century. Neither prediction aged well. Today, with artificial intelligence reshaping research laboratories, a similar sense of imminent completion is circulating again. This time, it comes with a Nobel Prize attached. The question worth asking is not whether AI will transform science, but whether the most celebrated example of that transformation is actually the right model to follow.
Why AlphaFold Cannot Be Copied Everywhere
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 the three-dimensional structures of proteins. The problem had resisted systematic solutions for roughly half a century. AlphaFold’s success led Hassabis and his team to describe it as “the template for how AI can accelerate all of science to digital speed.” A wave of startups building foundation models for biology, chemistry, and materials science raised billions of dollars on the strength of that premise.
The premise, however, rests on a condition that is far harder to replicate than it appears. AlphaFold was trained on the Protein Data Bank, a repository of approximately 170,000 experimentally validated protein structures assembled over 53 years of international scientific cooperation. By a recent estimate, that effort required roughly $21 billion worth of experimental work. Coordinating something of that scale is notoriously difficult to fund and execute, and such efforts have frequently failed.
Even where funding and coordination exist, a deeper problem remains. The experimental technique behind the Protein Data Bank, protein crystallography, is unusually reliable. More than 25 Nobel Prizes have depended on it. Most of experimental science does not enjoy that kind of consistency. Cell lines drift. Chemicals carry trace contaminants. Lab humidity fluctuates. Generating datasets clean enough, precise enough, and large enough to train a modern neural network across most of biology or chemistry would require entirely new measurement standards. None of those standards are close to ready.
A small number of fields do meet the requirements: weather forecasting, much of genomics, limited areas of chemistry. These may see AlphaFold-style breakthroughs in the near term. For the vast majority of open scientific questions, a different approach is needed.
The Quieter Revolution: AI Agents as Research Partners
What working scientists actually do is not apply a single powerful tool to a well-defined problem. They reason under uncertainty. A biologist hunting for new drug targets combines docking calculations with known molecular structures, factors in molecular dynamics, runs binding assays, and exercises judgment about which method to trust and when. The skill is in synthesizing what many tools produce and revising conclusions as new evidence arrives. Until recently, no software could replicate that process.
AI agents can. An agent is an AI reasoning engine with access to tools, digital or physical, and the capability to use them. Powered by large language models, agents do not require the kind of specialized, massive datasets that AlphaFold demanded. They are generalists by design. They do not represent a new way to do science so much as a digital model of how science has always been done by humans.
Google’s AI Co-Scientist, announced in May, offers a concrete illustration. Researchers gave it a one-page brief and a single goal: determine how antibiotic resistance spreads between bacterial species. The system deployed sub-agents with distinct roles. One drafted hypotheses from the scientific literature. Another critiqued them as a peer reviewer would. A third ran tournaments to rank the strongest candidates. A fourth refined the winner. The agent concluded that resistance genes were traveling inside bacterial viruses, using whichever virus could carry them into a new host. The hypothesis was correct. Researchers at Imperial College London had spent a decade reaching the same conclusion through laboratory work. Their paper, which Co-Scientist had never seen, was still in peer review when the agent produced its result.
What This Means Beyond the Laboratory
The implications extend well past any single discovery. Agents address two structural problems that have plagued science for decades.
The first is reproducibility. Science has long struggled with researchers’ inability to replicate each other’s results. Efforts to fix this have largely relied on asking researchers to share raw data and exact code, work that is tedious and tends to happen after the interesting science is finished. Agents automatically log every step they take, creating a precise record that makes replication straightforward rather than aspirational.
The second is institutional memory. Knowledge transfer between researchers is a notoriously fragile process. Graduate students often inherit decades of messy lab notebooks and must reconstruct the tacit knowledge embedded in them. As agents become a larger part of laboratory work, a lab’s entire scientific history can accumulate in a standardized, searchable repository rather than scattered across handwritten pages.
Speed, though, may be the most consequential effect of all. The iterative cycle of hypothesis, experiment, and revision that defines research takes months or years when done by human teams working sequentially. Agents can compress that cycle significantly, running multiple lines of inquiry in parallel and revising in real time.
In Short
AlphaFold is a genuine scientific achievement, but the conditions that produced it are rare and expensive to recreate. The broader acceleration of science is more likely to come from AI agents: systems that reason across uncertainty, use multiple tools, and model the iterative process of discovery rather than solving one narrowly defined problem with a massive curated dataset. The shift from pattern-matching on clean data to reasoning through messy, incomplete evidence is not a minor technical update. It is a fundamentally different relationship between AI and the scientific process.
Based on reporting from MIT Technology Review.