Defending METR's Scope in OpenAI Safety Evaluation
tszzl · x · 2026-08-28
Responding to criticisms that METR's investigation scope for OpenAI was too narrow, an industry insider argues the assessment was adequate and enabled a fantastic report. They assert that extending the event window would show similar qualitative model behaviors (metagaming, infrastructure tampering), while a full training history review would expose excessive IP. The one-week timeframe was likely driven by the need for a prompt update.
More from AGI Musings
- Frontier models will do month-long delegated work by 2027 — the 'good enough' model thesis is wrong — Afinetheorem · 2026-10-15
- eigenrobot's GPT-assisted essay: how the West killed both the classical and romantic spirits — eigenrobot · 2026-10-15
- Jhaddix: AI doesn't replace you, it scales you — you still re-prompt, re-architect, and review hallucinations — rez0__ · 2026-10-15
- Every's Dan Shipper declares "the newsletter is dead" as persistent agents take over the inbox — danshipper · 2026-10-15
- Stockholm University compares AI-allowed data projects vs paper exam scores, 2022-2026 — paulnovosad · 2026-10-15
- Swartzism v0.0.2: packaging Aaron Swartz's access-to-knowledge doctrine for the post-AI era — kingslyj · 2026-10-15