AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。
Here is an experience you have probably had by now. You set up an agent in the spring. Same model all quarter, weights frozen solid, not a single gradient step. And yet by summer the thing is noticeably better at your workflows. Fewer dumb mistakes, better tool choices, less babysitting. Nothing learned anything, in the technical sense of the word. So what improved? The answer is that improvement has three places to live, and only one of them is the model. Parts 1 and 2 of this series were about the bottom layer, the weights, where frontier labs run their industrial flywheel. This essay is about the other two layers, because that is where most of the user-visible improvement of 2026 actually happens, and because the mechanics up there are strange, underrated, and occasionally uncomfortable. Three layers, three very different update economics. The bottom layer gets all the press. The top layer gets all the updates. The employee with the immaculate notes Read more