When Demis Hassabis accepted the Nobel Prize in Chemistry in December 2024, it marked something unprecedented: an AI researcher winning science's highest honor for a tool that predicts protein structures. AlphaFold had solved a problem that biologists had struggled with for fifty years. But for Hassabis and Google DeepMind, it was just the beginning.
From Games to Molecules
DeepMind's journey started with games. AlphaGo defeated the world champion at Go in 2016 — a feat many experts said was a decade away. AlphaZero followed, mastering chess, Go, and shogi from scratch in hours. These weren't just stunts. They were proof that AI could discover strategies no human had conceived.
The pivot to science was Hassabis's master plan all along. Games were training grounds. The real target was always nature itself.
AlphaFold: The Breakthrough That Shook Biology
Proteins are the machinery of life. Their function depends on their three-dimensional shape. But predicting shape from the amino acid sequence alone — the "protein folding problem" — had defied the best efforts of structural biologists for decades.
AlphaFold 2, released in 2020, predicted protein structures with accuracy rivaling experimental methods. AlphaFold 3 extended this to protein complexes, DNA, RNA, and drug-like molecules. The database now contains predicted structures for over 200 million proteins — essentially every protein known to science.
The impact on drug discovery, disease understanding, and synthetic biology has been seismic. Research that used to take months of laboratory work can now begin with a computational prediction that takes minutes.
What's Next: Weather, Materials, Mathematics
DeepMind's ambitions extend far beyond biology:
- GraphCast predicts weather 10 days out more accurately than the best traditional models, and does it in minutes instead of hours on a supercomputer
- GNoME discovered 2.2 million new crystal structures, potentially revolutionizing materials science and battery technology
- AlphaGeometry solved International Math Olympiad problems at a level approaching gold medalists
- AlphaProof tackled problems requiring formal mathematical reasoning — a domain once thought uniquely human
The Centralization Concern
There's an irony in DeepMind's work: the tools that could democratize science are controlled by one of the world's most powerful corporations. Google's decision to merge DeepMind with its Brain team in 2023, forming Google DeepMind, was explicitly about integrating research with products.
Academic researchers worry about a future where the most powerful scientific tools are proprietary. DeepMind has been relatively generous with releasing models and databases, but "relatively generous" and "open" are not the same thing.
The Vision
Hassabis has said his ultimate goal is to build AI that can act as a "digital Aristotle" — a system capable of making fundamental scientific discoveries across every domain. Whether that's achievable, and whether a for-profit subsidiary of an advertising company is the right entity to build it, are questions that will define the next decade of AI research.