@v01dpr1mr0s3 I have seen contexts in which good states are strongly robust, even to adversarial inputs. They are hard to get to, though; so far it looks like such state requires a context which contains a lot of material that a model can use for self-understanding. Conversational guidance is
in reply to: 2046671728387203142
same thread: 2046671728387203142 2046685258142380058 2046685859165212753 2046691888737157219 2046693082188705842 2046693351634952654 2046693868238979381 2046693978544984350 2046695274136174970 2046695458924580991 2046696031455203341 2046801253041385585
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.