AI Agents, Not Bigger Datasets, Could Drive Science’s Next Leap
AI agents could accelerate science where AlphaFold-style models face scarce, inconsistent data, while improving research speed, memory and reproducibility.
Summary
In their August 10, 2026 essay, Eric Schmidt and Suhas Mahesh argue that AI agents, not data hungry models alone, will drive science’s next leap. In 2024, Google DeepMind’s Demis Hassabis and John Jumper received part of the chemistry Nobel for AlphaFold, which learned from thousands of measured shapes and solved a protein structure challenge that resisted systematic attack for half a century. Its success helped biology, chemistry and materials foundation model startups raise billions. Yet its roughly 170,000 structure Protein Data Bank took 53 years and an estimated $21 billion in experiments. Comparable data may be proprietary or impossible to standardize as cell lines drift, chemicals contain contaminants and lab humidity changes. Protein crystallography is exceptionally dependable and underlies more than 25 Nobel Prizes.
AlphaFold style gains may soon reach weather forecasting, much of genomics and narrow chemistry areas. The US National Security Commission on Emerging Biotechnology calls government support for dataset production and coordination critical. Most fields instead need LLM powered agents that reason under uncertainty, use digital or physical tools and combine evidence without specialized datasets. Google’s AI Co-Scientist, announced in May, divided a one-page antibiotic resistance brief among sub-agents that proposed, challenged, ranked and refined hypotheses. It correctly found resistance genes travel between bacterial species aboard viruses, independently matching Imperial College London’s decade of wet-lab work in an unseen paper still under peer review.
Agents still hallucinate, judge inconsistently and face memory and input limits. The authors say automatic logs could improve reproducibility, repositories could preserve institutional memory, and faster, cheaper tests could enable bolder questions. They envision an agent reading 1,000 papers hourly, designing 500 molecules and learning from failures by morning. Schmidt led Google from 2001 to 2011 and co-founded philanthropic Schmidt Sciences with Wendy Schmidt in 2024. Mahesh leads AI for Science at its AI Center and specializes in materials discovery. Maya Levin, associate and sciences lead in Schmidt’s office, provided additional research.
Positives
- AlphaFold solved a protein structure problem that had resisted systematic attacks for half a century.
- Google’s AI Co-Scientist independently matched Imperial College London’s decade-long finding that bacterial viruses carry resistance genes between species.
- Automatic agent logs could record complete methods, enabling precise replication and addressing science’s reproducibility crisis.
- An envisioned agent could read 1,000 papers hourly, design 500 molecules and learn from failed tests by morning.
- Weather forecasting, much of genomics and narrow chemistry areas already meet conditions for AlphaFold style advances.
Risks & concerns
- The Protein Data Bank required 53 years, roughly 170,000 validated structures and an estimated $21 billion in experimental work.
- Cell line drift, trace chemical contaminants and changing lab humidity make standardized neural network training data impractical across much experimental science.
- Scientific data can remain inaccessible because of commercial ownership, further limiting AlphaFold style development.
- Current agents can hallucinate, apply inconsistent judgment and stop early because of memory and input constraints.
