你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。
最容易做错的，是Zig Creator Calls Spade a Spade, Anthropic Blows Smoke；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：AI代码最贵的是评审分钟。。
My conservative take: the most expensive part of AI code is reviewer minutes, meaning the human time spent checking whether a change should be merged. That is why the line "Zig Creator Calls Spade a Spade, Anthropic Blows Smoke" sticks with me more than another polished 演示（demo）.
The Zig case makes it concrete. Andrew Kelley said one-off AI-assisted submissions were net negative while Zig had about 200 open PRs, because the real bill was reviewer attention, not token spend [S001].
A July 2026 paper found the same pattern across 294 open-source repositories: PR volume rose while merge rates fell as generated code got easier to produce [S005]. In plain English, writing got cheaper, screening got harder.
That does not make AI coding useless. It draws a boundary. AI on a private branch inside a team is one problem. Public inbound PRs with weak ownership are another. A product update is not worth your time because of how many features it lists. It is worth your time if it changes your next decision.
So the next metric I would watch is not tokens or lines generated. It is reviewer minutes per accepted change. If your team is still treating AI output as free throughput, share this with the person who owns the review queue.
真正该讨论的是：Zig Creator Calls Spade a Spade, Anthropic Blows Smoke