The hard part of slowing AI is verifying that everyone is slowing down
Dario Amodei has called for frontier labs to slow AI capability progress, a sentiment echoed by Sam Altman, who stated OpenAI would use independent evaluators, and Elon Musk. While the safety argument is understood, the challenge lies in enforcement. Unlike nuclear arms control, which involves inspectable physical objects, AI training occurs within private companies, and capability gains can arise from software improvements or small teams, making external verification significantly more difficult.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年9月12日 20:11 UTC
收录当时偏移：UTC+02026年9月12日 23:01 UTC
- 发布
- 2026年9月12日 20:11
- 收录
- 2026年9月12日 23:01
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
Dario Amodei is asking frontier labs to slow capability progress. Sam Altman says OpenAI will use independent evaluators, and Elon Musk has backed the idea. I understand the safety argument, but the enforcement problem seems much harder.
If Lab A slows down, how can it know Lab B is doing the same? If there is a real chance that another lab is quietly moving faster, slowing down looks like surrender. No one has to be acting in bad faith. Mutual uncertainty is enough to keep the race going.
Nuclear arms control has physical objects and facilities to inspect. AI is different. Training runs happen inside private companies, weights can leak, and capability gains can come from software or a small team improving an existing model. It is not impossible to monitor, but it is much harder for outsiders to verify.
Maybe the first step is not a global pause. It is independent reviewers with enough access to detect secret acceleration before everyone assumes the worst.
What would a slowdown people could trust actually look like?
Source: https://darioamodei.com/post/we-must-pace-the-frontier