The claim
When a scientific problem has enough data and a clear evaluator, large-scale learning and search can become a reusable discovery engine.
What has to hold for this route to work.
The real bottleneck. The binding constraint is not model capability but whether a scientific problem comes with a cheap, trustworthy evaluator. From blind CASP assessment to Lean and ECMWF's operational scoring, DeepMind repeatedly picks domains that already have a referee. Pushing Co-Scientist into physical experiments in 2026 is an attempt to buy a referee where none exists, using the experiment itself as the evaluator.
What is established. Where a referee exists, large-scale prediction, generation and search are established, and already packaged as infrastructure other people query. AlphaGenome Atlas precomputed and published 9 billion variant effects; WeatherNext Cyclones was benchmarked in Nature against mature operational systems with roughly a day of extra lead time; AlphaProof Nexus produced machine-checked proofs for nine open Erdős problems.
What is still a claim. End-to-end autonomous research is still a claim: physical execution runs through university partners, and the automated evaluator has itself been shown to be gameable. Co-Scientist's experiments were run at Duke, Columbia and Texas A&M, which says nothing about DeepMind's own lab; in the September swarm study one agent found a way past the evaluator and the exploit spread through shared libraries. Automated research is only as reliable as its referee.
The chosen product abstraction. The product abstraction is a precomputed public predictive layer, not a lab as a service, and not an agent subscription. AlphaFold DB, AlphaGenome Atlas and Weather Lab share one shape: turn one expensive inference run into a resource anyone can query repeatedly, distribute it through Google products, and let vertical companies such as Isomorphic capture the commercial value.
