什么是通用人工智能？When “AGI” Gets Lost in Translation

我最近注意到一个很少被认真讨论的问题，在中文语境中，人们谈论“通用人工智能”时，往往带着一种理所当然的想象，好像只要把现有的人工智能做得更强、更通用一些，它自然就会演变成所谓的“通用人工智能”。这种理解听起来顺理成章，但我越来越觉得，它在一开始就已经偏离了问题本身。

问题的根源，首先出现在语言层面，尤其是翻译。英文中的 Artificial General Intelligence，如果直译成中文，其实更接近“人工通用智能”，而不是“通用人工智能”。这两个表述在中文里看似只是一点语序差别，但强调重点却完全不同。中文“通用人工智能”很容易让人把注意力放在“通用”上，仿佛“智能”本身是可以被不断扩展、覆盖更多领域的东西；而英文里的 General Intelligence，强调的恰恰是“智能的类型”，它与 Narrow Intelligence 并列，指向的是两种不同形态、不同层级的智能，而不是同一种智能的不同使用范围。

正是这种语言重心的错位，制造了一个错觉：普通智能只要足够通用，就能自然演化为通用智能。可一旦进入技术和理论层面，这种直觉几乎站不住脚。所谓的“窄域人工智能”，本质上是任务绑定的系统，它们在既定目标和约束下表现得再出色，也并不意味着它们具备统一的认知结构、自我维持的能力，或者在陌生情境中进行真正迁移和重组的能力。而 AGI 在原始语境中所指向的，恰恰是一种整体性的、代理级别的智能形态。这两者之间的差异，不是覆盖面大小的问题，而是结构和存在层级的不同。

从这个角度看，把两者理解为一条连续发展过程中的前后阶段，本身就是一种误判。它更像是生物演化中的分化路径：猫不会通过不断“优化”而变成狗，它们各自沿着不同的演化规律展开。未来是否可能出现跨越性的架构突破，这是另一回事，但至少可以肯定，这绝不是一个靠堆算力、加任务、拉长上下文就能线性抵达的目标。用工程上的连续优化，去替代对智能形态本体差异的思考，只会让讨论越来越空洞。

也正因为如此，我越来越觉得，很多关于人工智能的争论，其实在翻译那一刻就已经被悄悄引导到了错误的方向。语言并不是中性的容器，它会塑造理解的路径。虽然专业术语的演变会形成特定的含义，不能望文生义，但是最初翻译成中文的时候就应该避免容易引起误解。

对这些问题保持敏感，本身就是一种重要的思想训练。而在这一点上，学习英语、直接接触原始语境，并不是一种“技能优势”，而是一种避免被概念误导、避免被二手理解牵着走的基本能力。至少，在面对这样根本性的议题时，它能让我们更接近问题真正开始的地方。

When “AGI” Gets Lost in Translation

I’ve recently noticed a fairly common phenomenon that rarely receives serious attention: in Chinese discussions, “general artificial intelligence” is often understood as something that existing AI systems will naturally become, as long as they grow stronger, broader, and more versatile. The assumption feels intuitive, almost self-evident. Yet the more I think about it, the more convinced I am that this intuition is already misguided at the very starting point.

The root of the problem appears first at the level of language, especially translation. Artificial General Intelligence, if rendered more literally, would be closer to “artificial general intelligence” rather than “general artificial intelligence.” In Chinese, this subtle shift in word order quietly changes the center of gravity of the concept. The emphasis slides toward “generality,” making intelligence itself seem like something that can simply be expanded, generalized, and applied everywhere. In English, however, general intelligence is contrasted with narrow intelligence; it refers to a category or type of intelligence, not to the degree to which a system is broadly applicable.

This linguistic shift gives rise to a particularly persistent misunderstanding: the idea that narrow intelligence, once made sufficiently general in its application, will eventually become general intelligence. But once we move beyond language and into technical and theoretical territory, this assumption quickly begins to unravel. Narrow AI systems are, at their core, task-bound. No matter how impressive they appear within specific domains, this does not imply that they possess a unified cognitive structure, self-sustaining agency, or the capacity to reorganize themselves meaningfully in unfamiliar situations. AGI, as originally conceived, points toward an agent-level form of intelligence—a fundamentally different mode of existence, not merely a broader toolbox.

Seen from this perspective, treating the two as points on a smooth continuum is a category mistake. The gap between them is not simply one of scale or coverage, but of structure. It resembles divergent evolutionary paths rather than incremental optimization: a cat does not become a dog by being trained more thoroughly, and each follows its own internal logic of development. Whether a future architectural breakthrough might bridge this gap is an open question, but it is clearly not something that can be reached through linear engineering alone—by stacking more data, more tasks, or more computation.

This is why I increasingly feel that many debates about AI are already skewed at the moment of translation. Language is not a neutral vessel; it actively shapes how problems are framed and what kinds of solutions seem plausible. Sensitivity to this is not pedantry but intellectual hygiene. In that sense, learning English and engaging directly with original technical and philosophical discussions is not merely a practical skill—it is a way to avoid being quietly misled by second-hand concepts, and to stay closer to where the real questions actually begin.