AI for science needs reasoning, not just data

Every few decades, someone declares science complete — Albert Michelson in 1903, Stephen Hawking in the 1980s — and the rise of AI has revived that belief. In 2024, Demis Hassabis and John Jumper of Google DeepMind won part of the Nobel Prize in Chemistry for AlphaFold, a neural network that predicts protein structures from thousands of experimentally measured examples. AlphaFold solved a problem that had resisted systematic attacks for 50 years, and DeepMind called it '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 raised billions on the strength of that promise.

But the article argues this template is fragile. AlphaFold's success depended on the Protein Data Bank, a dataset of about 170,000 experimentally validated protein structures that took 53 years of international cooperation and an estimated $21 billion of experimental work to assemble. Efforts of that scale are extremely difficult to fund, coordinate, and execute, and they often fail. Moreover, even well-organized fields face the scientific impossibility of generating comparable data: protein crystallography is an unusually replicable technique (25+ Nobel Prizes relied on it), whereas most experimental science sees results that vary — cell lines drift, measurements fluctuate.

Given these constraints, the article contends that AI will bring profound changes but AlphaFold is not the blueprint. The near-term acceleration of science will come from AI agents that reason, not just AI systems trained on massive static datasets.