AI Agents Will Transform Science More Than AlphaFold
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- AlphaFold's success depended on the Protein Data Bank, which took 53 years and roughly $21 billion of experimental work to assemble approximately 170,000 validated protein structures — conditions the authors say are unreplicable in most fields.
- AlphaFold-style foundation models require standardized, replicable datasets that exist in only a handful of areas such as weather forecasting, genomics, and limited corners of chemistry, making the approach impractical for most of science.
- AI agents digitally model how scientists actually work — combining multiple tools, reasoning under uncertainty, and revising conclusions as new evidence arrives — without the need for massive domain-specific training datasets.
- Google's AI Co-Scientist, announced in May, demonstrated the approach by spinning up sub-agents that hypothesized how antibiotic resistance spreads between bacterial species, reaching the same conclusion Imperial College London researchers arrived at over a decade of wet-lab work.
- Current agents still hallucinate, show inconsistent judgment, and face memory and input constraints that limit how long they can run autonomously — limitations the authors expect to fall away.
- Agent-driven workflows could structurally address science's reproducibility crisis by automatically logging every step, and could build standardized repositories of institutional knowledge that outlast individual researchers.
Why it matters: The authors — former Google CEO Eric Schmidt and Schmidt Sciences AI lead Suhas Mahesh — frame agentic AI as a 'calculus-tier' tool that transforms every scientific field at once, rather than the narrow, data-bound breakthroughs AlphaFold represents. For funders, this reframes the bottleneck from dataset-building to reasoning-architecture development; for working scientists, it points to a near-term path where experimentation costs collapse and reproducibility is automated.
Ask SkimNews
