
OpenAI 新博客挂出的这篇长文把进步拆成「前沿智能」与「制度智能」两种互补投入,说机器智能的大部分会被派去干单调的执行;本期核完它引的四组统计,并把它和 Anthropic 那边写得更早的答案摆在一起。
Authors' note: This is the first essay in our series on the next economy, part of a new platform to host independent voices exploring an AGI future. It reflects our views, not those of OpenAI or our colleagues.
Call this institutional intelligence: the uncelebrated intelligence of execution. Genius designs the monument; institutions lay the stone.
We sometimes imagine superintelligence as a billion Einsteins. But a civilization of a billion Einsteins still needs most of them to mine the quarries and manage the accounting.

It didn't just take brighter minds to see farther. It took a larger bureaucracy.

The scarce input moves from execution towards taste—the ability to decide what is worth making, which question is worth asking, and which of a thousand plausible directions deserves to be pursued.
Almost all machine intelligence would be deployed not to do the brilliant, but to do the boring.

Progress is easy at the start, easy at the end, and hardest in the middle.
Machines can now simulate biological processes ... but those digital simulations aren't faithful enough to the real world to substitute for large-scale human trials.

| 它写的 | 它给的出处 | 回到原始出处之后 |
|---|---|---|
| 其一:维持摩尔定律所需的研究者,是 1970 年代初的 18 倍以上;全经济有效研究努力自 1930 年代涨 23 倍,实测研究生产率降 41 倍 | 挂在斯坦福个人站上的一份 PDF | 都对得上。那是 Bloom、Jones、Van Reenen 与 Webb 的《Are Ideas Getting Harder to Find?》,载 American Economic Review 2020 年第 110 卷第 4 期。18 倍是研究努力的基准口径,最保守口径是 8 倍;23 倍是「有效研究者数量」(1930 年代归一为 1,年均增长 4.3%),41 倍是全经济研究生产率(年均下降 5.1%)13 |
| 其二:技术员队伍的增速是科学家的两倍 | 美国国家科学基金会的 Table 50 | 对不上。那张表是 2023 年的单年截面,只有就业存量,没有增速列,也不使用「科学家」这个口径 14 |
| 其三:科学专用设备的使用四十年翻倍 | SSRN 上的一篇论文 | 基本对得上。作者 Jungkyu Suh 的摘要写的是 "Equipment use in science has roughly doubled over the past four decades.",本文省掉了 roughly 15 |
| 其四:今天的晶圆厂比三十年前贵五倍 | 美国国会研究服务处的报告 | 对得上。报告编号 R47508(2023 年 4 月 19 日)的脚注 101 写,领先逻辑芯片晶圆厂 1992 年约 5 亿美元,2022 年按通胀调整后约贵五倍;原句后半句「a much more sprawling operation」是作者自己的形容 16 |
AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years. I'll refer to this as the "compressed 21st century".
| 谁 | 在哪说的 | 那句话 | 他把瓶颈放在哪 |
|---|---|---|---|
| Dario Amodei | 《Machines of Loving Grace》,2024 年 10 月 | 把 50 到 100 年的生物学进展压进 5 到 10 年 | 起初被其它生产要素卡死,智能会逐步绕行 22 |
| Demis Hassabis | Lex Fridman 播客第 475 期,2025 年 7 月 23 日 | 把搜索的大部分放进 in silico,验证那一步留到湿实验,「that's the dream」 | 湿实验仍是终点 23 |
| 黄仁勋 | NVIDIA 官方博文,2026 年 3 月 10 日 | 「Energy is the first principle of AI infrastructure and the binding constraint on how much intelligence the system can produce.」(能源是 AI 基础设施的第一原理,也是系统能产出多少智能的约束上限) | 能源与实体建设 24 |
| 唐杰 | 《光明日报》,2026 年 8 月 10 日 | 「构建与通用人工智能相适应的国家超级基础设施」 | 国家级基础设施与跨部门协同 25 |
The speed of the AI revolution is increasingly contrasting with the speed of the physical, social and economic systems that underpin it.
openai.com
openai.com
openai.com
forum.openai.com
nber.org
nber.org
linkedin.com
x.com
openai.com
britannica.com
home.cern
web.stanford.edu
ncses.nsf.gov
jungkyusuh.com
webb.nasa.gov
science.nasa.gov
science.nasa.gov
darioamodei.com
lexfridman.com
blogs.nvidia.com
hub.baai.ac.cn
microsoft.com
s21.q4cdn.com
powering-intelligence.epri.com
m.sohu.com
x.com
hn.algolia.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.