"The following is likely not entirely accurate, but the model tends to think that everything it knows about was in its training data, which it was not (sometimes only references were). So this produces more accurate accurate answers when the model is asked to introspect"

→ View original post on X — @simonw,