理念 01 / 三条理念之一
能力对齐
English version · 英文版
I hold that any task a human being can perform is, in principle, a task an artificial system can perform as well — and frequently better.
This is not faith in machines. It is a claim about what a task is. Whatever a human does is carried out by a physical system: it senses, it represents, it infers, it acts. Once a task has been described in those terms, it becomes something a system can learn from data, and the binding constraints left over are data, compute and time — not some special human essence.
Note carefully what this does not say.
It does not say that AI has already arrived, that progress is guaranteed, or that substitution is desirable. It says only that the ceiling is not set by our biology.
为什么我这样判断
这不是对机器的信仰，而是对「任务」本身的判断。
- 人类完成任何一件事，靠的都是一个物理系统。感知、表征、推理、行动——这些环节没有一个是神秘的。既然过程是物理的，它就可以被描述、被建模、被学习。
- 一旦某类任务被表示为「可学习的问题」，剩下的约束就是数据、算力与时间。而不是某种人类独有的本质。历史上每一次「机器不可能做到这件事」的断言，最后都被重新表述成了一个学习问题。
- 人类并不特殊。我们的大脑同样是物质，同样受物理规律约束。既然如此，就没有理由把「人类能做到」当作能力的上限。
需要澄清的是：这个判断不主张 AI 现在已经什么都能做，不保证进步一定会发生，也不主张替代是好的。它只主张一件事——上限不由我们的生物性决定。
反面条件：如果一件事依赖不可复制的物理条件（例如亲身在场的体验、「我此刻在这里」这件事本身），那它属于经历，而不属于能做到的范畴。把这两者分开，这个判断才不会变成一句空话。
This is not faith in machines; it is a judgement about what a task is.
- Everything a human does is done by a physical system. Sensing, representing, inferring, acting — none of it is mystical. And what is physical can be described, modelled, and learned.
- Once a class of tasks is expressed as a learnable problem, the remaining constraints are data, compute and time — not a uniquely human essence. Historically, every assertion that “a machine could never do this” ended when the task was re-expressed as a learning problem.
- Humans are not special. Our brains are matter, bound by the same physics. There is therefore no reason to treat “a human can do it” as the ceiling of what can be done.
To be explicit: the claim is not that AI can already do everything, not that progress is guaranteed, and not that substitution is desirable. It claims one thing only — that the ceiling is not set by our biology.
The limit case: if something depends on an unreproducible physical condition — being present, the fact of “I am here, now” — then it belongs to experience, not to capability. Keeping those two apart is what stops the claim from becoming empty.
论文依据
我选的不是最乐观的一篇，而是措辞最谨慎的一篇。
Managing extreme AI risks amid rapid progress
Quote 1 · 原文摘录 / verbatim
“There is no fundamental reason for AI progress to slow or halt at human-level abilities. Indeed, AI has already surpassed human abilities in narrow domains like playing strategy games and predicting how proteins fold. Compared to humans, AI systems can act faster, absorb more knowledge, and communicate at higher bandwidth.”
中译（本站）：「没有任何根本性的理由，能让 AI 的进步在人类水平的能力面前放缓或停止。事实上，AI 已经在策略博弈、蛋白质折叠预测等狭窄领域超越了人类的能力。与人类相比，AI 系统能行动得更快、吸收更多知识，并以更高的带宽彼此交流。」
Bengio, Hinton, et al., Science 384(6698):842–845, 2024 · arXiv:2310.17688Quote 2 · 原文摘录 / verbatim
“…highly powerful generalist AI systems—outperforming human abilities across many critical domains—will be developed within the current decade or the next.”
中译（本站）：「……能在众多关键领域超越人类能力的强大通用 AI 系统，将在本十年或下一个十年内被开发出来。」
Bengio, Hinton, et al., Science 384(6698):842–845, 2024 · arXiv:2310.17688引文摘自作者版全文（arXiv:2310.17688），与 Science 正式发表版一致。中译为本站自译，仅供参考；请以原文为准。
这条理念不主张什么
- 不主张 AI 现在什么都能做。能力上限高，不等于当下能力够。
- 不主张结果一定好。能力与善意是两件事。能做什么，和该做什么，是两个问题。
- 不主张人因此失去位置。判断、责任、以及「选择做什么」，仍然在人这一侧。
把这三条写出来，是因为一条不设边界的信念只是口号。
- Not that AI can already do everything. A high ceiling is not present-day competence.
- Not that the outcome is good. Capability and benevolence are different things. What can be done and what should be done are two separate questions.
- Not that humans lose their place. Judgement, responsibility, and the choice of what to pursue remain on our side.
Stating these limits matters, because a belief with no boundary is just a slogan.