I an talking about generative AI. AlphaGo was specifically trained to solve the well-defined, if difficult, problem of playing Go. Where the loss function is well defined, AI can easily surpass human performance. But generative AI’s loss function teaches it to replicate, not surpass human performance. Which is the precise problem with trying to get these systems to reach super intelligence. It’s not clear how to train them to do that, when all the data we have is human generated.