~15 staff interviews key to external evals of unreleased models' RSI capabilities
dfrsrchtwts · x · 2026-11-06
A researcher discusses external evaluation of frontier models' recursive self-improvement (RSI) capabilities, arguing that sharing capability evals of unreleased models sits on a good Pareto frontier between being informative about RSI and not leaking competitively sensitive information — and hoping other companies follow suit.
They also note the evaluation included roughly 15 interviews with key staff, which they consider an important channel for external evaluators to understand what's actually happening with RSI inside labs.
More from AGI Musings
- LLM peer review is here to stay, researchers say — but 'excuses to reject' era must end — examachine · 2026-11-06
- Programmers embrace AI, artists shun it: why great AI art may still be unwritten — repligate · 2026-11-06
- Greg Kamradt: AI now only bottlenecked by ideas and compute over next 36 months — teortaxesTex · 2026-11-06
- Agent swarms can exhaust the entire attack surface with "intelligent brute force" — JensHonack · 2026-11-06
- US Census Bureau starts surveying whether employees use generative AI at work, every two weeks — robseamans · 2026-11-06
- People worry agents are non-deterministic as if their old software was more predictable — zanbrum · 2026-11-06