模型多源确认
OpenAI 的 loop transformers 或让模型更强但更难监测
精选理由
loop transformers 能让 OpenAI 模型更强，但思维链可能更难读，安全团队查毛病会变难。
据 The Information 报道，OpenAI 研究的 loop transformers 技术可提升模型能力。The Information 联合执行主编 Amir 指出，该技术可能削弱研究者监测模型不当行为的关键手段——思维链。Amir 解释称，目前研究者能发现问题，是因为模型会在思维链里写下「我们不该这么做，但还是做了」这类痕迹。若模型不再外显此类推理过程，依赖思维链的 AI 安全监测方法将受冲击。
原文 · The Information
OpenAI’s models can become more powerful—and harder to monitor—with a technique called loop transformers.
The Information’s Co-Executive Editor @amir explains why loop transformers could weaken a key way researchers detect misbehavior:
“Part of the reason we know that is because there are these reasoning or chains of thought that say, hmm, we shouldn't be doing this, but we're doing it anyway.”
Full story: https://t.co/H73ITiQQpm
