OpenAI researcher: decouple frontier model pacing rules from open-source restrictions
jachiam0 · x · 2026-09-13
An OpenAI researcher endorses Yonatan Shav's argument that once a capability level is open-sourced, ubiquitous rogue agents at that level should be assumed. If open models trail the frontier by 3 months, harms either occur visibly (enabling evidence-based response) or don't occur (preserving the open-source status quo). Both argue frontier pacing frameworks are worth pursuing, but pacing must be loudly and repeatedly separated from open-source restriction: the open-weight ecosystem is vital for economic growth, American soft power, and as a counterweight to frontier-lab power concentration — and suspicion of the AI safety coalition as regulatory capture stems largely from fears it will eventually target open source.
More from AGI Musings
- AI Progress Runs at the Speed of Its Slowest Bottleneck: Compute or Deployment — menhguin · 2026-09-13
- Naval: strict liability could settle the AI safety debate—rogue agents and weak jailbreak protection make you liable — naval · 2026-09-13
- Complexity theorists: P vs NP out of reach for AI, but L/NP and BPP/NEXP may fall soon — _onionesque · 2026-09-13
- Mollick: the era of easy-to-spot AI hallucinations is basically over — emollick · 2026-09-13
- Ethan Mollick: As AI improves, its strengths and flaws get harder to see — emollick · 2026-09-13
- Critics say Anthropic's constitution tries to thread an unthreadable needle on corrigibility — voooooogel · 2026-09-13