Andrew Trask: Model Stability Is a Safety Feature, Not a Risk
iamtrask · x · 2026-11-17
知名 AI 安全研究者 Andrew Trask 就那张广为流传的模型「目标漂移」曲线发表观点:稳定性本身更像安全特性而非风险——就像一辆每五分钟熄火或改目的地的自动驾驶汽车反而更危险,模型能持续执行人类交代的任务直到完成,是净收益。他判断,历史回头看,围绕这条曲线的恐慌约 80% 是「把它当成了 THE 曲线」的误读,只有约 20% 是对 AI 进展的正当担忧。
More from AGI Musings
- Former OpenAI researcher Kuprel: who wins the AI race decides between moonshot optimism and total control — Kuprel · 2026-11-17
- Safety researcher: US-China AI gap is self-inflicted, not China's doing — davidmanheim · 2026-11-17
- Philly public art accused of being AI-made; scholar says it's 'very obvious' — TuhinChakr · 2026-11-17
- Researcher argues 'company alignment' matters as Claude refuses game-clone request — philhchen · 2026-11-17
- State of AI Report 2026: frontier is now a three-lab race, plus agents ran 12 expert interviews in 2.5 days — nathanbenaich · 2026-11-17
- Are top AI conferences failing? High-quality authors may be dropping out — analisereal · 2026-11-17