AI agents, not AlphaFold, will accelerate science
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- AlphaFold succeeded because of the Protein Data Bank, a dataset of roughly 170,000 experimentally validated protein structures that took 53 years of international cooperation and an estimated $21 billion in experimental work to assemble — conditions most fields cannot replicate.
- AI agents powered by large language models represent a fundamental architectural shift because they model the iterative process of human research rather than requiring scientifically specialized datasets, making them applicable as generalists across fields.
- Google's AI Co-Scientist, announced in May, was given a one-page brief on how antibiotic resistance spreads between bacterial species and correctly hypothesized that resistance genes hitch rides on bacterial viruses — a conclusion Imperial College London researchers took a decade to reach through wet-lab work.
- Co-Scientist spawned sub-agents that drafted hypotheses from literature, peer-reviewed them, ranked candidates through tournaments, and refined the winning hypothesis, mirroring the contingent judgment scientists apply under uncertainty.
- Agents could address science's reproducibility crisis by automatically logging every methodological step, producing exact records of how results were derived in contrast to researchers' long resistance to sharing raw data and code.
- Agents will amplify institutional scientific memory by recording a lab's entire history in standardized repositories, replacing the decades-long process of graduate students deciphering predecessors' messy lab notebooks.
- Eric Schmidt and Suhas Mahesh frame agentic AI as a tool comparable in scope to calculus, statistical inference, spectroscopy, and the computer — instruments that revealed entirely new fields of problems to formulate.
- DeepMind's 2024 Nobel-winning AlphaFold team, including Demis Hassabis and John Jumper, called their model 'the template for how AI can accelerate all of science,' a framing the authors contend overstates how transferable that template is.
Why it matters: If Schmidt is right, the billions flowing into biology and chemistry foundation-model startups chasing the AlphaFold playbook are misdirected: the durable winners will be companies building reasoning agents that work with messy real-world data, not those waiting for someone else to spend decades assembling a clean training set. For researchers, the near-term implication is that reproducible methodology and cumulative lab knowledge could finally become default rather than aspirational.
Ask SkimNews
