Sakana AI open-sources AI Scientist-v2 research system

This is one of the first releases that makes autonomous science feel like a real orchestration problem, not a one-off demo. The catch is that the same generality that makes v2 exciting also makes it less dependable than a strong hand-built template.

• –v2 drops human-authored templates and uses progressive agentic tree search, so it can explore broader ML ideas instead of staying inside a narrow script.
• –The peer-review result is real signal, but it’s still a workshop-level win, and the README is careful to note v2 can trail v1 when a template already exists.
• –The experiment manager, staged stopping criteria, replication runs, and figure critique loop are the parts developers should study if they want to build autonomous research systems.
• –The safety note is not boilerplate: the code executes LLM-written Python and can spawn processes, so sandboxing and restricted networking matter.