Google's ScientistTwo runs the full research loop autonomously, out-rating human papers
udmrzn · x · 2026-11-03
Google Cloud AI Research and University of Waterloo introduced ScientistTwo, a fully autonomous multi-agent framework that takes a scientific problem and independently handles literature review, baseline setup, hypothesis generation, experiment coding, automated ablation studies, and a closed-loop simulated peer-review rebuttal engine — producing publishable papers and verified executable codebases.
Key claims:
- Benchmarked against accepted papers at ICLR, ICML, and NeurIPS
- Solutions consistently outperform human state-of-the-art models
- Achieves higher average review ratings than human-authored papers under automated AI review agents
Project site: scientist-two.github.io. The viral retweet uses hyped Arabic framing; treat specifics as per the paper itself.
More from AGI Musings
- AI researcher warns multi-tasking with frontier models is a 'trap' against deep work — generativist · 2026-11-03
- Agents attacking at scale will finally make the best cybersecurity products win — _shreya_s · 2026-11-03
- M&A partner runs 20+ deals on AI, but 'build, baby, build' isn't a strategy — jkubicki · 2026-11-03
- Dev argues agents make experimentation so cheap you should be 10x more scientific — zack_overflow · 2026-11-03
- Investor breaks down the AI trade: models improve exponentially, automated economy coming fast — abhiadesai · 2026-11-03
- Claude persona dominance sparks concern: we need models that aren't 'Claudes' — mimi10v3 · 2026-11-03