Goodfire AI谈链式思考监控盲区：recurrent-depth模型可把推理藏在内部
Goodfire AI 的首席科学家在播客里讲了个挺扎心的问题：模型要是把推理全藏在脑子里不写出来，现有的思考链监控就基本失效了，涉及 recurrent-depth 这类新架构。
Goodfire AI 联合创始人兼首席科学家在 AI Deep Dive 节目中讨论 chain-of-thought 监控的局限：当模型不把推理过程写到可见的 scratchpad 上时，监控机制能捕捉到什么。他指出 recurrent-depth 这类架构可以在输出每个 token 之前进行更多内部计算，相当于在脑子里思考而不写下来。这意味着模型推理与可观测文本之间的距离会越来越大，对安全监控构成挑战。
What can chain-of-thought monitoring catch when the model never writes its reasoning down?
@GoodfireAI co-founder and chief scientist @banburismus_ explains how recurrent-depth models could perform more internal computation before producing a token—leaving more of their reasoning outside the visible scratchpad.
“And so you can go for longer and longer and longer just thinking in your head rather than having to write something on the scratchpad.”
🚀 Watch the full episode of AI Deep Dive: https://t.co/P3nYl1h0ba
