AI for science needs reasoning, not just data

Every few decades, science gets declared basically done. Albert Michelson said as much in 1903. Stephen Hawking floated a similar idea in the 1980s. Now AI has revived that old optimism, and AlphaFold gave it a very shiny badge: in 2024, Demis Hassabis and John Jumper of Google DeepMind won part of the Nobel Prize in chemistry for work on protein structure prediction.

That success was real, and it was also unusually dependent on a very specific setup. DeepMind had the Protein Data Bank, with roughly 170,000 experimentally validated protein structures to train on. Building that archive took 53 years of international cooperation and, by one estimate, about $21 billion in experimental work. Protein crystallography also happens to be unusually dependable, which is one reason more than 25 Nobel Prizes have leaned on it.

Most of science does not look like that. Cell lines drift. Chemicals pick up contaminants. Lab humidity changes. In a lot of fields, the problem is not just lack of scale; it is that the data are too messy and inconsistent to feed a modern neural network in the first place. Weather forecasting, much of genomics, and narrow slices of chemistry may still be ripe for AlphaFold-style wins. For most open questions, the article argues, they are not.

The better near-term bet is agentic AI: systems that can reason with tools instead of waiting for giant perfect datasets. Scientists already work that way, combining docking calculations, known structures, molecular dynamics, binding assays, and judgment. Agents now can do a version of that loop digitally, and they can keep records while they work. That matters for reproducibility, because the machine can automatically log each step instead of leaving everyone to reconstruct the experiment from a half-legible notebook later.

Google’s AI Co-Scientist is the cleanest example in the piece. Given a one-page brief about antibiotic resistance, it built hypotheses, attacked them, ranked them, and settled on a correct answer: resistance genes were riding on bacterial viruses. Researchers at Imperial College London had reached the same conclusion after a decade of wet-lab work. The point is not that the machine is done with science. It’s that a tool which can read, test, revise, and remember may do more for science than a tool that only has lots of data and a narrow trick.

The catch is that agents still hallucinate, wobble in judgment, and run into memory limits. But if those frayed edges get cleaned up, the payoff is bigger than a single protein model. The article’s real bet is that science gets faster when software starts acting less like a prediction engine and more like a junior research team that never forgets to file its notes.