It took Google's Deep Mind project, based in England, months and months to train the first version of the Go program, feeding it huge quanties of data about actual human Go games. And it won the match against the greatest human champion.
The second version was given no data, just the rules of the game (and a vast memory with the ability to compare everything it saw with everything it had seen before). It trained for just three days, playing 4.9 million games against itself. It then promptly beat the first version of the program 100 to 0.
Amazing story!
To some friends I earlier sent around a long story on this from the Financial Times, but I see it’s now behind a paywall. This story is so incredibly amazing that some of you might want to do some deep googling.
This computer is an idiot savant. It only does one thing: play Go. But it has invented strategies never seen before in thousands of years of human play, and it will reliably (and swiftly) slaughter any human Go player on the planet. That’s an amazing—and terrifying—accomplishment.
Update 1: I have the impression most readers don’t know about Go and have no clue whatsoever about how hard it is to get a computer to play top-level Go. I invite you to do some reasearch on that. It’s a genuinely amazing story, both on Go and its development over many centuries, and on how the computer has conquered all of that in a few days.