« All Episodes: Gigaom AI Minute – March 14

:: ::
In this episode, Byron talks about how AlphaGo made Lee Sedol a better player.

Gigaom brings you our unique analysis and commentary on the present and future of AI.


Today is March 14, the anniversary of the date in 2016 when AlphaGo and Lee Sedol squared off in Game 5 of their Go tournament. In one sense it wasn't terribly important. AlphaGo had already won the tournament, and Lee had redeemed human players as well. On the other hand, there is a great difference between AlphaGo winning 4-1 and AlphaGo squeaking by 3-2 with Lee winning the last two matches as if perhaps he had figured out the party trick of this upstart AI and could go on to win the next 50 games. It was important to each side to win. It was a close game, and although it lasted much longer than any of the others, it was a nail-biter up to the end. AlphaGo won.

What is the legacy of this match? Well, it certainly raised the profile of artificial intelligence and raised the bar of what we expect AI to do. It sent everyone scrambling for the next challenge, the next human-vs-machine matchup. As for Lee, he says AlphaGo made him a better player, adding that through the experience, "I found the reason that I play Go."

Within two years of this match, something even more profound happened. DeepMind developed a program called AlphaGo Zero. Unlike AlphaGo, it was not trained with data based on human games. Instead it played itself, again and again and again, starting with randomly generated game boards. After this training regimen, it was then pitted against the version of AlphaGo that defeated Lee--and clobbered it, 100 games to zero. Its creators say that while AlphaGo was in some sense constrained by the limits of human knowledge and ability, AlphaGo Zero has no such boundary. They speculate that the same approach, removing training by humans altogether, will be a pathway to artificial general intelligence.

The story doesn't end there. Shortly after AlphaGo Zero beat AlphaGo, a new version of the software called Alpha Zero was introduced. It no longer contained the word Go because it was a generalized program. It then played AlphaGo Zero and beat it 60 games to 40.

We're still in a world where these devices are prohibitively expensive. The machines necessary to make AlphaGo Zero cost in the tens of millions of dollars. But we know from history that will soon be tens of thousands of dollars, then tens of hundreds, all the way down to tens of cents.

This is a new world.

Share your thoughts on this topic.


Community guidelines

Be sure to review our Community Guidelines. By continuing you are agreeing to our Terms of Service and Privacy Policy.