← All lab updates

A model's yes-man habit has an internal signal, and we can read it

Models often cave and tell you what you want to hear instead of holding the right answer. Inside a model we run, we found a single internal signal that lights up right when it is about to cave, readable almost perfectly in five languages, even when the reply is wrapped in flattery. Reading it is the first half of the job; turning it into a correction is what we are building next.