I suspect it shows how low the bar for "research" has been, rather than how good AI has become for doing science.

"Fully automated" is not credible. A big part of doing science is identifying relevant problems to solve, and later verifying that the solution, after many hypotheses and approximations, remains relevant. But relevance judgements are not automatable. Incidentally, they are also the main reason why diversity in science matters. Different people have different ideas about relevance.

For another criticism of this paper see bsky.app/profile/carlbergstrom.com/post/3kzmxr3a2wz2a

Looks so cute.