Cao Yuan, a former senior research scientist at Google DeepMind and co-founder/CEO of Unreasonable Labs, says AI for Science is entering its explosive phase. But he believes it will take at least another two or three decades before AI can truly make Nobel Prize-level scientific discoveries on its own. Cao previously worked on projects like Gemini; his current startup is aimed at getting AI to discover new knowledge.
His biggest bottleneck, he says, is verification. Code can be tested immediately, and math can be checked step by step with Lean. But biology, materials science, and physics ultimately require real experiments — and an experiment can be expensive or slow. That makes it hard for AI to iterate with the quick trial-and-error of a coding agent.
The harder part is "creating new concepts." Cao argues today's models are great at searching within existing knowledge, definitions, and theorems, but they still can't abstract out new mathematical objects, definitions, and theories the way top scientists can. He calls that "conceptual abstraction" the "last mile" of AGI.