AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。
Forget the agent that rewrites its own code. The most economically important self-improvement loop in AI is the post-training pipeline, and every frontier lab has been running it at industrial scale for two years. Last time I argued that the factory got automated before the design office, and that the line between them tracks whether the work comes with an answer key. This time I want to look inside the part of the factory that matters most, because there is a self-improvement loop running in there that gets almost no attention in the RSI conversation, even though it is the one actually producing the models. Here is the loop. A model writes many candidate answers to a problem. Something grades them. The good ones become training data. The model trains on them and gets slightly better at producing good ones. Repeat. That is it. It is called STaR in the 2022 paper that first stated it cleanly, RLVR in the 2025 vocabulary, and post-training in the org chart. It is the same loop, and it is the loop that took frontier models from chatbots to agents. Sourdough, not a screwdriver Read more