AI for Science Needs Reasoning, Not Just Data

Author: Eric Schmidt and Suhas Mahesh (Schmidt Sciences) URL: https://www.technologyreview.com/2026/08/10/1141384/ai-agents-for-science/ One-line: AlphaFold is a profound achievement but a rare template — AI agents that model the iterative human process of research are the broader lever that will accelerate science across every field.

Key claims

• AlphaFold’s conditions are rare and not replicable everywhere. AlphaFold succeeded because the Protein Data Bank — 170,000 validated structures built over 53 years at a cost of roughly $21 billion — provided the consistent, scalable data needed to train a neural network. Most fields cannot produce comparable datasets, because experimental results vary (cell-line drift, chemical contaminants, humidity), measurement tools aren’t standardized, and coordination at that scale is “infamously difficult to fund.”
• AI agents are the generalist lever AlphaFold is not. An agent is an AI reasoning engine given access to tools and the ability to use them; it does not require a massive specialized dataset. It “digitally models the human process of discovery” — combining multiple imperfect methods, weighing their strengths, and revising as evidence accumulates — rather than applying a single powerful pattern to a narrow question.
• Google’s AI Co-Scientist demonstrates the gap. Given a one-page brief and the goal of understanding how antibiotic-resistance genes spread between bacterial species, Co-Scientist spun up sub-agents to draft hypotheses, peer-review them, run tournaments, and refine the winner. Its conclusion — that resistance genes hitch rides on bacterial viruses — matched a decade of wet-lab work at Imperial College London still in peer review at the time.
• Agents structurally fix the reproducibility crisis. Where researchers resist sharing raw data and code (tedious work done after the interesting science is complete), agents automatically log every move, creating an exact method record that enables precise replication.
• Agents will amplify scientific memory. Currently, knowledge transfer between researchers is “famously murky” — grad students pore through messy lab notebooks. As agents become integral to lab work, a central, standardized institutional knowledge repository emerges automatically.
• Speed is the most important impact. When testing an idea takes less time than debating it, people stop arguing and run the test. An agent that reads a thousand papers per hour, designs 500 molecules, and learns from failed tests by morning will “fundamentally change the pace at which science gets done” — and free researchers to chase bolder, stranger questions.
• Agentic AI is a historically rare tier of breakthrough. The authors compare it to calculus, statistical inference, spectroscopy, and the computer — tools that “envelop every field of science at once” rather than solving a specific problem.

Notable quotes

“AlphaFold had shown that the combination of AI and sufficient data could make groundbreaking discoveries (even if we did not understand the underlying mechanisms involved), and it seemed, once again, that a path through the rest of science was laid out before us.”

“Though it is a profound achievement, the conditions that produced the likes of AlphaFold are rare, and the time it will take to meet those conditions in other fields will be measured in decades, not years.”

“Scientists have always reasoned under uncertainty. Biologists working to identify new drug targets have never had perfect datasets. Instead, they combine docking calculations and known structures, factor in molecular dynamics, run a handful of binding assays, and use their judgment to weigh each method according to its particular strengths and points of failure. The skill of science is not in any single tool; it is synthesizing what many tools produce, and revising the results as the evidence comes in.”

“An agent that can read a thousand papers in an hour, design 500 molecules, and learn from its failed tests by morning will bring down the cost of experimentation and fundamentally change the pace at which science gets done.”

“Historically, tools of such scope have arrived just a handful of times: calculus, statistical inference, spectroscopy, the computer. Each revealed a world of problems no one had thought to formulate, and those problems, in turn, defined their fields anew.”

How it connects

• Eric Schmidt — co-authored via Schmidt Sciences, his philanthropic science-funding vehicle; this is his public argument for why agentic AI matters more than foundation models for most of science.
• Deep Learning — AlphaFold is the canonical win for the deep-learning-plus-large-dataset paradigm; the article argues that paradigm has limits outside rare high-data fields.
• Data Science — the piece draws a sharp contrast between the data requirements of pattern-matching models (AlphaFold) and the reasoning-under-uncertainty approach agents employ.
• AI Safety — agents hallucinate, lack consistency, and have memory constraints — the paper names these as barriers that “will fall away” but doesn’t deeply engage the safety dimension.
• Google — Google DeepMind built AlphaFold; Google released AI Co-Scientist, the agent example at the center of the article.