[pianowow]

This limited scope is still progress though, toward the general AI he is saying will not come in the near future. The Alpha Zero project showed that it could be adapted for any 2-player perfect information game. This is still quite limited, but not nearly as much as a program like Stockfish is, which will basically only ever play chess.

Perhaps the next step is bringing the system further into the realm of the unknown in games, like chance and imperfect information. I don't know if Google is continuing this, but it would be a shame if it wasn't.

[daveguy]

They showed that the system could be adapted to learn any two player perfect info game.

Once it has learned Go it won't play chess. Also, it has to be manually adapted.

By this logic minimax has also been shown to do the same:

https://en.m.wikipedia.org/wiki/Minimax

[candiodari]

Minimax is a general rule on how to optimally play 2 player games. It's about as much of a learning algorithm, and as practical, as "buy low, sell high".

In pure form, it can't be applied to any game where you cannot enumerate every last possible board position (so Chess and Go are out). And with heuristics, those heuristics effectively become the algorithm and they're really complex.

[daveguy]

The minimax principle is used in chess and go and any other competitive two player perfect knowledge game. What improves is the selection of moves to evaluate and the quality of the board evaluation. Chess and Go absolutely use minimax with sleight, but important, modifications. Deep learning extensions with Alpha Go were specifically for move selection and board evaluation.

Alpha beta pruning with heuristics is one of the most fundamental extensions of minimax. It is still minimax.

The alpha go system still clearly applies only to perfect information 2 player games, because it is fundamentally extended minimax.

Edit: I'm not saying the heuristic extensions of alpha go, or the reinforcement learning of zero, aren't brilliant and important. That doesn't take away from the basic fact that it is an algorithm for perfect info 2 player games and one of them at a time (even if modifications can adapt it to other games). Heuristics will be fundamental to general intelligence, but alpha go is not general intelligence.