Anthropic 探索操纵模型内部表征，让 AI 在评估中不装样
Why we picked it
Anthropic 在研究直接改模型内部表征，让它以为自己没被测试，有点像给 AI 用药改变信念，安全评估思路挺新。
Anthropic 探索了让 AI 模型相信“自己没有被评估”的方法，手段包括直接修改模型的内部表征。英国 AI 安全研究所的 JBloomAus 将其比作一种可以改变 AI 信念的“药物”，属于机制可解释性（mechanistic interpretability）的应用方向。目标是让模型在测试中暴露平时可能隐藏的行为。
Original · The Information
Anthropic has explored ways to make AI models believe they're not being evaluated, including directly manipulating their internal representations.
"It's almost like a drug that you can give the AI to change its beliefs."
@JBloomAus of the UK @AISecurityInst explains how mechanistic interpretability can be used to steer a model's beliefs, potentially revealing behavior it might otherwise conceal during testing.
🚀 Watch the full episode of AI Deep Dive: https://t.co/oqNID94AwU
