For a long time, “AI scientist” meant a person who builds artificial intelligence systems. Lately, the term has also come to mean software systems that try to carry out the whole research cycle by themselves, from reading the literature to writing and reviewing the paper. In 2026 such systems moved from curiosity to headline: a fully AI-written paper passed peer review and was described in Nature, and in September a swarm of some ten thousand AI agents produced a computer-verified proof that the three-dimensional Navier–Stokes equations can develop singularities, settling one of the Millennium Prize Problems and provoking a sharp reaction from the mathematical community. This article provides an informal overview of the field from the point of view of a researcher who has worked in machine learning, computer vision and multimedia since 2002. It explains how current AI scientist systems are organised, how much attention and public funding the topic now receives, including the European RAISE initiative, looks at the Navier–Stokes episode as a case study, and summarises what the evidence says about the strengths and weaknesses of these systems. The systems are fast, tireless and very effective where results can be checked automatically, but they remain weak at choosing questions and at producing understanding rather than answers. The article closes with the question many researchers are now asking each other: how do we take it from here?