← Back to feed

AI-Scientist-v2

GitHub Repo Pretty sure · cherry-picked success story
https://github.com/SakanaAI/AI-Scientist-v2

Autonomously generates ML papers via tree search and LLM agents—genuinely ships research, but the 'workshop paper' narrative obscures that it's still mostly prompt engineering at scale with heavy human scaffolding.

25%
20%
55%
Slop 25%Signal 20%Science 55%

Real contribution: agentic tree search + iterative refinement actually generates valid experiments and produces peer-reviewed output. That's non-trivial. But the framing is doing work: v1 'works best with templates,' v2 is 'lower success rates' exploratory—basically admitting v2 is the more interesting version but less reliable. The workshop paper was accepted peer review, which is signal, but one paper from a well-resourced org with unlimited compute proves concept, not maturity. Heavy relia...

2827 stars Python 2025-12-19 353 days old

Become a MFer to rate — log in