Comment by iheartpotatoes
8 years ago
Alpha zero is a generalized AlphaGo, so they are headed in the right direction!
AlphaGo -> win this boardgame
AlphaZero -> win any boardgame (well, three right now)
AlphaMinus1 -> win any game
AlphaMinus2 -> win anything
AlphaMinus3 -> win winning. so much winning.
But you get my drift, I'm extrapolating hugely off one abstraction step.
Yeah maybe. How do you define "win" in terms of general intelligence? Outsmart all top human experts at their field?
Impossible to answer: who "won" politics? Democrats or republicans?
Eventually we go from a black-and-white rules (hell, even Go is kind of a negotiation), to using economics as a tool to determine what option satisfies all parties. I don't think humans have solved that yet, but like radiology, perhaps AI can make better compromises?
Whatever the reward function dictates would be a win :)
Right. So we've just successfully moved the goalposts to "write a reward function that objectively evaluates g"