OpenAI researcher: frontier AI ideas can't be patched for safety after the fact
willdepue · x · 2026-10-28
willdepue (OpenAI) argues frontier safety can't be patched at the end: "you can't leak the plans for Little Boy in 1944 and then lock down security in 1945." The theories, ideas and methods being developed today are the ones that lead to tomorrow's models — implying capability research spillover must be governed now.
More from AGI Musings
- tszzl: Early AGI signs were obvious to Scott Alexander and Gwern readers — tszzl · 2026-10-28
- Shkreli vs ChrisGPT: does AGI 'curing' diseases actually count as a cure? — ChrisGPT · 2026-10-28
- Anyone can imagine a language as fast as C and as dynamic as Lisp — until they sit down — math_rachel · 2026-10-28
- Rachel Thomas threads zetalyrae: why 'the ideas are mine, the writing is AI' is wrong — math_rachel · 2026-10-28
- Emmett Shear's Softmax essay argues trust is a property of relationships, not models — Hidenori8Tanaka · 2026-10-28
- Biotech founder: most 2023 customers are dead or pivoted to data generation — kenbwork · 2026-10-28