[SciShow intro Plays]

Hank: For fans of artificial intelligence, the past few weeks have been very exciting. Last October, for the first time ever, an AI built by Google, known as AlphaGo, beat a professional human player at the incredibly complex board game Go. Then, in this last week, it went on to beat one of the best players in the world four times. Which is a really big deal. And let us explain to you why.

When engineers talk about artificial intelligence, they don’t mean sentient humanoid robots conversing with us, talking to us, becoming feeling, then taking over the world destroying us. AI is really just a way of programming computers to do things that humans normally do. See, computer programs follow a specific list of instructions to complete tasks, so they struggle with situations that have lots of options and require decision-making on the fly. Including playing complicated board games, like Go.

The basic objective of Go is to get as many points as possible, either by capturing the other player’s pieces or by claiming areas of the board with black or white stones. And a big part of the winning strategy for any game is the ability to think a few moves ahead, taking into account what the other player might do each time you make a move. We humans can use our past experiences with the game to figure out the best moves, and how our opponents might respond to them. But it’s hard to build that kind of pattern recognition into a computer program.

So generally, computers play board games -- like chess -- by searching through all the possible combinations of moves to find the one that means it’s most likely to win. The problem is, Go is played on a huge 19-by-19 grid -- and there are hundreds of possible moves that a player might make in every turn. In fact, there are more ways for a game of Go to play out than there are atoms in the universe. So if it’s that difficult to program an AI to beat humans at Go, how did engineers teach AlphaGo to do it?

Well, instead of having the AI search through all the possible combinations of moves they tried to help it understand the difference between a good move and a bad one. To do that, the engineers first fed AlphaGo 30 million combinations of moves, taken from real games with expert human players. Then, AlphaGo played thousands of matches against itself, to learn new strategies.

All that knowledge, plus some clever programming, helps the AI decide on the best next move. Instead of having to consider every single possible move, which would take a really long time, it can quickly narrow down the few, most relevant options. And so far, this strategy has worked really well against us humans. A few months ago, AlphaGo played a 5-match tournament against Fan Hui, the European Go champion, and won every single game -- the first time a computer had ever won against a professional Go player at all.

So then, Google decided that it was time to test the AI against Lee Sedol, a South Korean who’s been the top Go player in the world for the past decade. They livestreamed all five matches, and uploaded 15-minute summaries -- which are all linked in the description below. And -- spoiler alert -- AlphaGo won the first three, so best out of three of five: it won the tournament.

So it was pretty clear that the way the engineers programmed and trained the AI did work. But AlphaGo and Sedol were still set to play the last two games in the series. And Sedol actually won the fourth game -- which was a huge deal, because it showed that the AI still isn’t perfect at choosing the best moves. A big turning point in this game was when Sedol played a move known as a wedge -- one that has lots of possible responses. Basically, he was trying to confuse the AI by giving it too many options to explore. And that seems to have worked, even after all that training. After the wedge, the game went downhill for AlphaGo, until its internally-calculated chances of winning went below 20% -- at which point it’s programmed to resign.

Game 5 turned out to be a very close game, with AlphaGo making a mistake early on, but eventually winning. So AlphaGo might not be the best Go player in the world right now -- but it’ll just keep getting better from here, so odds are, eventually, it will be. Either way, it’s a huge step forward for artificial intelligence.

Thank you for watching this episode of SciShow News, and thank you especially to all of our patrons on Patreon who make this all possible. Thank you so much for being a patron if you are. If you want to become one of those people you can go to patreon.com/scishow­ there’s a bunch of cool things you can get there as well. And don’t forget to go to youtube.com/scishow and subscribe!