Emmanuel Ameisen says some models show an internal “gut check” signal distinct from step-by-step reasoning

What was said

"we found that, you know, models can like actually answer questions. Obviously, they can do math or they can, you know, you can ask them like, what's the color of a banana? And they'll actually like you know if you look inside you'll see all these intermediate steps but they have a separate mechanism which is kind of like a gut check uh and which is totally different and which is just like as they reason they'll have like some part of them that will flag like something's wrong."

Emmanuel Ameisen says his team has observed, through interpretability experiments, that some models appear to have an internal mechanism separate from explicit reasoning that flags when an answer seems obviously wrong. He characterises it as a “gut check” or hunch-like signal that can detect implausible results without re-deriving the full calculation.