In this clip Hinton says, "Recent research showing that if you give them a goal and you say you really need to achieve this goal, um, they will pretend, um, to do things during training. So during training, they'll pretend not to be as smart as they are so that, um, you will allow them to be that smart. So it's scary."
That is, AI is smart unless it's pretending to be a slacker, so you will allow it to be smart. This argument conflates AI’s functional limitations with human-like deception and some self-referential nonsense about human permission, resulting in something completely incomprehensible.
That said, I hope AI companies start using this to explain any inconsistent performance.
"No, no, no… our AI isn’t underperforming. It’s strategically sandbagging."
