> It is by definition incapable of trying something entirely new Only if the AI doesn't see any usage. But when you start using it, you get AI+human generating

> It is by definition incapable of trying something entirely new Only if the AI doesn't see any usage. But when you start using it, you get AI+human generating things that can be outside its training scope. Humans can input novel ideas into the prompt or include new information. Then AI gets to retrain on the data generated with human in the loop and improve. There is a chance for humans and AI to explore new directions. If you got a simple way to test, like the DeepMind geometry problem solving model, then AI models can improve on their own by doing massive search and validation. Kind of similar to the scientific method - formulate hypothesis, validate, observe outcomes, and repeat. Works for code too, or any AI output that can be "executed" and generate a validation signal.