19世紀の経済学者ジェヴォンズは、蒸気機関の効率向上によって石炭の消費効率が上がると、かえって全体の消費量が増えることを見出しました。
用「杰文斯悖论」解释推理时间扩展（inference scaling）——这是一个绝妙的框架选择。效率提升→整体消耗增加，这正是 o1/R1 类推理模型出现后发生的事：单次推理更贵，但人们愿意为更难的问题付出更多算力。Sakana 用一个 19 世纪的经济学悖论，为 2026 年的 AI 产品战略提供了令人信服的理论背景——在技术营销中，历史类比是建立认知可信度的最有效工具之一。
https://en.wikipedia.org/wiki/Jevons_paradox
The increased efficiency in the use of a resource doesn't imply a decrease in the usage of that resource, but rather may cause in incommensurate increase in use because of decrease in cost.
The increased efficiency of the use of a resource may act as a catalyst for increasing usage.
When we throw in the fact that some of the parties with the deepest resources, expertise and capabilities in AI are the very same providers of the primitives – providers intent on growth and whose customers are struggling with the size of their respective product catalogs – it’s worth asking whether said providers may be coming around on the idea of a PaaS, but models based on AI rather than prescriptive curation and constraints. One thing, at least, is clear: if any of the above speculation proves true, the industry equilibrium is about to be punctuated.
This seems to make the argument that the biggest are likely to leapfrog past regular PaaS where devs choose some abstracted services, to some next thing where all the wiring of services together is provided automatically. A bit like "terraform, but AI"