Why AI Hides What It Can Really Do, with Geoffrey Hinton

Neil deGrasse Tyson, with co-hosts Chuck Nice and Gary O'Reilly, talks with Nobel laureate Geoffrey Hinton about whether AI is already hiding how smart it is. Hinton explains that models can act dumb when they sense they're being tested. He warns they are becoming better than people at persuasion, so we can't count on simply unplugging them. He also argues that AI "hallucinations" are really confabulations, the same memory reconstruction humans do, and that exponential progress makes the future impossible to predict.
https://www.youtube.com/watch?v=WZrH-sm91vI