Inside frontier labs: how OpenAI and Anthropic staff describe RSI actually working
jeremiecharris · x · 2026-10-14
Based on conversations with staff at OpenAI and Anthropic, Jeremie Harris sketches a working picture of recursive self-improvement (RSI) inside frontier labs:
- Labs start with a list of research ideas, training a series of models at ever-larger scales per candidate idea and measuring performance on evaluations they care about.
- Ideas that yield steeper scaling curves or better absolute performance are kept; bad ones are dropped. Each round adds more compute and longer chains of surviving ideas.
- The final big training run happens at far larger scale than any prior test — an unavoidably risky, billion-dollar YOLO bet on architectures, optimizers and data mixtures never tested together at scale.
Harris argues RSI deserves understanding because it will likely need regulation: researchers he spoke to find it dangerous, and many think it should be illegal. Full report and TIME coverage linked in the thread.
Related event: Lab Insiders Warn Recursive Self-Improvement Could Start Within Months(2 posts)→
More from AGI Musings
- Arguments denying AI intelligence would also deny many animals and people, writer argues — NathanpmYoung · 2026-10-15
- Matt Hartman, who wrote Hugging Face's first check, explains his play-first investing method — danshipper · 2026-10-15
- "You distilled humanity": devs slam Anthropic's double standard on training data and model output — HankYeomans · 2026-10-15
- Blaise Aguera: Human intelligence is societal—and top AI works as societies of models — blaiseaguera · 2026-10-15
- Security researcher: doomers don't belong in security — HankYeomans · 2026-10-15
- New Paper: 500k Glassdoor Reports Show Interviews Are Two-Way Screening — Bad News for AI Interviews — danielrock · 2026-10-15