Levie 刚刚给出了我见过的关于 AI 丰裕（AI abundance）最好的论证之一……但他仍然低估了真正的瓶颈。 他用杰文斯悖论（Jevons Paradox）来框定问题是对的。 当某样东西变得更便宜时，需求会扩张，以吸收所有效率提升。煤炭、算力、云软件，都是如此。这个模式一再成立。 是的，AI Agent 会像 SaaS 降低确定性工作的成本一样，大幅降低非确定性工作的成本。 但真实数据揭示的是另一件事：根据 BCG 2024 年 AI 采用研究，74% 的公司在从 AI 中获取并规模化价值方面遇到困难。这不是成本问题。任务成本已经下降了。 这是一个人类监督（human oversight）的问题。 真正的瓶颈在于：仍然需要有人去明确 AI 应该做什么、评估它是否做对了、把输出整合进工作流，并且做出最终判断。AI 让执行变得廉价，但规格定义、评估和整合仍然受限于人类。 德勤（Deloitte）的企业研究发现，企业抵制 AI 的原因来自：不熟悉、技能缺口、缺乏变革管理 只有 三分之一的公司把培训列为优先事项。而 2025 年的数据更具说明性：只有不到 5% 的企业报告因 AI 出现净裁员；相反，37% 的企业报告了任务重新分配，员工正在转向监督、整合和决策角色。 这意味着，知识工作的杰文斯悖论形态，与 Levie 描述的并不完全相同。 对 AI 任务执行的需求必然会爆炸式增长。但对人类判断、情境设定和质量验证的需求，增长得会更快。 一个 10 人规模的服务型公司，现在几天内就能做出原型。但公司里仍然必须有人知道该做什么、原型是否有效、以及如何部署。这些能力，比它们所驱动的具体任务要稀缺得多。 真正的竞争在于：谁能解决“监督瓶颈”，谁能构建起在规模化条件下进行规格定义、评估和整合的组织能力， 谁就能从廉价的 AI 执行中攫取大部分价值。 约束已经从：“我们负担得起做这个任务吗？” 转变为：“我们有没有人能判断这个任务是否被正确完成？” Levie just made the best case for AI abundance I’ve seen… and he’s still underselling the real bottleneck. His Jevons Paradox frame is right. When you make something cheaper, demand expands to absorb the efficiency. Coal, compute, cloud software. The pattern holds. And yes, AI agents will drop the cost of non-deterministic work the way SaaS dropped the cost of deterministic work. But here’s what the data actually shows: 74% of companies struggle to achieve and scale value from AI, according to BCG’s 2024 adoption study. That’s not a cost problem. Task costs already dropped. It’s a human oversight problem. The bottleneck is that someone still has to specify what the AI should do, evaluate whether it did it correctly, integrate the output into a workflow, and make the final call. AI makes the execution cheap. But specification, evaluation, and integration remain human-bound. Deloitte’s enterprise research found that resistance to AI adoption stemmed from unfamiliarity, skill gaps, and lack of change management. Only a third of companies prioritized training. And the 2025 data is even more telling: fewer than 5% of firms reported net workforce reductions from AI. Instead, 37% reported task redistribution, with workers shifting toward oversight, integration, and decision-making roles. This means the Jevons Paradox for knowledge work has a different shape than Levie describes. Demand for AI task execution will absolutely explode. But demand for human judgment, context-setting, and quality verification will explode faster. The 10-person services firm can now build a prototype in days. But someone at that firm still has to know what to build, whether the prototype works, and how to deploy it. Those skills are scarcer than the tasks they enable. The real trade is that the companies who solve the oversight bottleneck, who build organizational capacity to specify, evaluate, and integrate at scale, will capture most of the value from cheap AI execution. The constraint moved from “can we afford to do this task?” to “do we have anyone who can tell if this task was done correctly?” Jevons works. But the coal equivalent here is the human attention required to make that output useful.