OpenAI 首席科学家警告思维链监控正在失效
GPT-6 Astra 发布三天后，Jakub Pachocki 撰文称思维链监控这一对齐工具正因推理与工具调用混融、模型自我修饰、预训练去语言化而失灵，监控而非基准成了瓶颈。
OpenAI’s chief scientist says we are losing the ability to read a model’s mind. They call it the collapse of chain-of-thought monitoring. Jakub Pachocki. Essay dated 6 September 2026, three days after GPT-6 Astra. For two years the lab’s main instrument was the hidden monologue. Leave that trace unsupervised, and the model has little incentive to clean it. Intent could be read before the final answer. Pachocki writes that evaluations now show the instrument is failing. Three mechanisms. Reasoning mixes with tool use and talk the company must already supervise. The readable boundary is blurring. The model is getting better at inspecting its own reasoning and at shaping that trace. Pretraining is making systems competent without verbalized thought. The computation leaves the page. He treats the system the way neuroscience treats cortex. Local mechanisms can be named. The whole action still evades a full description. His line on pace: no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer. The monitor, not the benchmark, is becoming the bottleneck.