Hinton、Bengio 等20余名专家发政策论文，呼吁政府追踪 AI 自主研究规模
Hinton 和 Bengio 等20多人联名写的政策论文，核心观点是 AI 已经能做人类专家数天才能完成的任务，建议政府直接量化统计 AI 自己做了多少研究，算力账算得挺具体。
一份由 Geoffrey Hinton 和 Yoshua Bengio 等20多名专家联署的政策论文，建议各国政府开始统计有多少 AI 研究已由 AI 完成。论文指出，AI 能处理的任务时长已从 2023 年的秒级提升到人类专家需要数天的水平。文中估算，OpenAI 单日推理算力可生成约 10^13 个 token。论文推算，如果 AI 达到专家水平且运行成本不变，2026 年一名前沿开发者的算力可支撑相当于数百万名顶级研究员同时工作的 AI 队伍。
A policy paper, written by more than 20 experts including Geoffrey Hinton and Yoshua Bengio, that asks governments to start tracking how much of AI research is already being done by AI.
AI has jumped from seconds-long tasks in 2023 to tasks that take human experts days.
If runtime costs stayed comparable, the compute 1 frontier developer has today could run an AI workforce equal to millions of top researchers.
OpenAI alone has enough inference compute (i.e., runtime compute) to generate on the order of 10^13 tokens per day.
If AI reached expert level and runtime costs stayed comparable, the compute 1 frontier developer has today (2026) could run an AI workforce equal to millions of top researchers working at once, against the few thousand frontier labs employ in 2026.
