The AI Scientist is published in Nature, with evidence that its papers improve as base models improve

Sakana AI/The AI Scientist

Sakana, UBC, the Vector Institute and Oxford published The AI Scientist as an open-access Nature paper. It describes the full automated pipeline and reports that paper quality rises as the underlying language models get better. The authors also list its limits: naive ideas, weak rigour and hallucinated citations.

• Pipeline: ideas, literature search, tree-search experiments, LaTeX write-up.
• Finding: better foundation models give better generated papers.
• Limits named by the authors: shallow ideas, complex code, wrong citations.
• Code for v1 and v2 is open source on GitHub.