Posted: Oct 19, 2018 10:37 am
zoon wrote:
My main beef with that passage it that it describes a markedly silly way to build robot butlers. Why on earth train them via the slow and inefficient method of punishment and reward, when they could perfectly well be programmed to do what you want in the first place? This would be very much simpler, as well as giving better results.
There is a simple answer to that bit. It's simpler to devise an algorithm to make a ganaral purpose AI do what we want. We can't even define what we want as simple rules.
We might not use punishment and reward in the human sense though. It would more likely be a suite of merit / cost functions that influence reinforcement lerning algorithms to guide the evolution of the neural nets.
The scope for AI to invent unexpected behaviours is considerable and will probably increase the close we get to a general inteligence AI. I don't think it's that far fetched for a robot butler to surprise you by rearranging your furniture if it has at some time been tasked to move furniture and trained to anticipate your needs.
All the same a digital computer system operating in an environment is a model of deterministic function. We don't need to imagine a "free will chip" for the scenario to be credible. We could have what looks like free will from determinism.