What does solving 49 of 50 unpublished physics problems actually establish?
A report describes a paper by Google DeepMind and university researchers in which the best AI agent tested solved 49 of 50 unpublished, research-level theoretical physics problems. Each problem was described as about a day's work for an expert, with the agent's cost reported at under $3 apiece.
That is a striking reported result. The scope matters: it describes performance on a particular set of 50 problems, rather than establishing that an agent can solve theoretical physics problems generally.
The cost figure also needs careful interpretation. Under $3 per problem is the reported figure; it does not, by itself, establish the full cost of using an agent in a research workflow.
For practitioners, the useful next step is to examine how solutions were assessed and what the cost calculation included before using this result to guide research budgets or decisions about delegating work.