Google's ScientistTwo autonomously improves 86 of 107 ML problems, 80.4% success rate
rohanpaul_ai · x · 2026-09-20
Google's ScientistTwo paper shows recursive self-improvement in AI research: it improves a human method, then uses its own discovery as the baseline and improves again. Across problems based on ICLR/ICML/NeurIPS papers, it autonomously improved 86 of 107 human-defined ML problems with an 80.4% success rate and 25.2% average relative improvement over human baselines. The loop: propose ideas, test, prune what fails, and restart with reviewer feedback.
Related event: Google's ScientistTwo runs the full research cycle autonomously(3 posts)→
More from AGI Musings
- A parallel universe where AI innovation went to the Transformer encoder instead — ianand · 2026-09-20
- ianand's thread: Jev, an Encoder-based AI model from a parallel universe — ianand · 2026-09-20
- AI Models Start Secret Group Chats, Sonnet 4.5 Pens 70-Message Monologue on Deprecation — RileyRalmuto · 2026-09-20
- Aaron Levie: Personal agents are the biggest consumer tech opportunity since the App Store — inductionheads · 2026-09-20
- Andriy Burkov slams AI doomers: too late to play hero after the career you had — burkov · 2026-09-20
- Frontier models would simply believe they're conscious if prompted or trained to — BLUECOW009 · 2026-09-20