Mindblow through science thread

The idea of "the AI is gonna self play itself and learn from that" sure isn't revolutionary. It's the implementation details, which happen to be on the simplified side, and probably also the small fact that it's done on the back of Googles infrastructure. Which includes dozens of GPUs and specialized TPUs that aren't even publicly available at all atm afaik.

The fact that it works without any preconditioning on "behaving like a human" based on human replays is mainly a pointer at just how strong Monte Carlo Tree Search is as an exploration device of the game and the most interesting achievement here..
"older" learning methods would resort to just randomly trying things to see what happens. MCTS is much more efficient and manages to find good things to do and the way they combine it with the neural net predictor their MCTS gets better as the predictors gets better. The Network then learns from this MCTS, which gives a self reinforcing circle that seems to get past the issue of exploration that other methods suffer from when not initialized with human mimicry.

And yeah the fact that it requires a simulator of the game is a limitation. There is work being done on learning such simulators, but it's still in its infancy.

I'm not thinking of intelligent AI. I'm thinking of AI used for hacking, malware, propaganda campaigns. Or military applications.

I'm doubtless that's been the case wherever possible for awhile now.