This Is Why A Computer Winning At Go Is Such A Big Deal

People didn't think this would happen for at least 10 years; it's a sign of how far artificial intelligence has come.

For the first time in history, a computer has beaten the human world champion at Go.

AlphaGo / YouTube / Via youtube.com

Go is an ancient Chinese game in which you place stones on a 19 by 19 board, and capture your opponent's stones by surrounding them. The rules are very simple, but they give rise to a complex, subtle game.

This morning, AlphaGo, a computer designed by the Google-owned, London-based company DeepMind, defeated Lee Sedol, the reigning Go world champion, in the fifth game of a five-game series. AlphaGo beat Lee 4-1 overall, with Lee taking the fourth game, when the series was already lost.

Here's why that's a big deal. First, Go is incredibly complicated – millions upon millions of times more complex than chess.

youtube.com

"It's sheer mathematics," Professor Murray Shanahan, an AI researcher at Imperial College London, told BuzzFeed News. "The number of possible board configurations in chess, of course, is huge. But with Go, it's enormously larger."

In chess, there are on average about 35 to 38 moves you can make at any point. That's called the "branching factor". In Go, the branching factor is about 250. In two moves, there would be 250 times 250 possible moves, or 62,500. Three moves would be 250 times 250 times 250, or 15,625,000. Games of Go often last for hundreds of moves.

It's sometimes said that in chess there are more possible games than there are atoms in the observable universe. In Go, by one estimate, there are something like a trillion trillion trillion trillion trillion trillion trillion more than that. To write the total number out, you'd need to put a 1 followed by 170 zeroes. That's why, nearly 20 years after computers became better at chess than humans, they've only just caught up at Go.

That means that a computer can't just look at every single possible move and pick the best one.

That's called "brute force" processing. "You simply can't use brute force for Go," says Shanahan. "You can't with chess either, but you can tackle it that way a bit, use brute force to search ahead through many, many possibilities. But with Go the number of possible board combinations is enormously larger."

The branching factor means that even a few turns ahead, the number of possibilities becomes too huge for even the fastest computer to search through.

That means that AlphaGo's victory isn't simply a product of computers getting faster and more powerful. Computers will never be powerful enough to brute-force Go. Software is always more important than hardware.

"The general rule of thumb in these areas is that hardware counts for an enormous amount, but software counts for more," Eliezer Yudkowsky, an AI researcher and co-founder of the Machine Intelligence Research Institute (MIRI) in California, tells BuzzFeed News. "If you have a choice between using software from 2016 and hardware from 1996, or vice versa, and you want to play computer chess or Go, choose the software every time."

And that means that AlphaGo has had to use learning techniques that are more like human intuition.

Human players don't follow every possible branch that the game could go. They look at the board and see patterns. "The way that human players play chess or Go or any game like that," says Shanahan, "is that we get to recognise what a good board pattern looks like. There's an intuitive feel for what's a good strong position versus what's a weak one.

"Human players build that up through experience. What DeepMind have managed to do is capture that process using so-called deep learning, so it can learn what constitutes a good board configuration."

Its creators fed it hundreds of thousands of top-level Go games, and then, after it had learned from them, let it play against itself, millions of times.

Lee Sedol. Pic by Reuters

And from that huge dataset, it was able to pick out the deep rules of Go that the top players know intuitively (but often can't explain). It started out as a not-very-good player, and learned from its own mistakes.

"The core of it is having one system play itself, and improve itself to a superhuman level, without specific tweaking from its designers," says Yudkowsky.

Unlike humans, because it learns from a vast number of games, it can't actually learn very much from any single one. "Each game is only a drop in the ocean of data," says George van den Driessche, one of the AlphaGo researchers. "It contributes only a tiny amount to the eventual model, so attempting to incorporate our games against Lee Sedol into our model would make no noticeable difference."

So its own designers probably don't know, really, how it works.

Kasparov v Deep Blue in 1997. Stan Honda / AFP / Getty Images

They understand the principles behind its learning, and its overall structure, but not the methods it's used to defeat Lee. That makes it entirely different to IBM's Deep Blue, the chess-playing computer that beat the world No 1 Garry Kasparov in 1997.

"Deep Blue was special purpose," says Yudkowsky. "Its designers tweaked it as it went along; people kind of understood how it worked. From the outside it looks like the people who did AlphaGo don't know how it works."

Shanahan agrees: "I don't suppose anyone in DeepMind understands quite how AlphaGo beat Lee Sedol."

AlphaGo's victory has come as a major shock to the artificial intelligence community.

"People weren't expecting computer Go to be solved for 10 years," says Yudkowsky.

Even the AlphaGo team were shocked. Van den Driessche says: "We certainly were. We went very quickly from 'Let's see how well this works' to 'We seem to have a very strong player on our hands' to 'This player has become so strong that probably only a world champion can find its limits'."

While the way it learns is somewhat similar to how humans do, there are subtle but important differences.

AlphaGo / Nature / Via nature.com

"It's called a 'neural network', so that sounds very brain-like," says Shanahan. "And each of the little 'neurons' in these networks sort of resembles a neuron [nerve cell] in the human brain.

"But they're an approximation. It's loosely inspired by what happens in the human brain, and the way the networks are connected together is loosely inspired, but that's all."

The main difference, says Shanahan, is that when your brain performs some action – orders your hand to swing a tennis racket, or retrieves a memory – a certain pattern of neurons will fire. When the same pattern fires repeatedly, the connections between those neurons get stronger, so the pattern gets fixed and the action gets easier.

AlphaGo also has patterns of connections between its neurons. But instead of its patterns getting stronger or weaker as they fire, it looks at the outcomes it wants to achieve, then uses an algorithm to adjust the strengths of the various connections to best achieve them. It's a technical-sounding difference, but it means that at a deep level, AlphaGo thinks in a very different way to human players.

The way AlphaGo learns means that it has applications outside playing games.

By Goban1 - Own work, Public Domain / Via commons.wikimedia.org

The methods the team has used – the deep learning and the self-improvement – can be used in areas other than Go. "It's a more significant milestone than chess in 1997," says Shanahan, "because the techniques they've applied to this have quite general applications."

Demis Hassabis, the founder of AlphaGo, has talked about medical applications – using the deep learning techniques to create an AI that can help doctors make diagnoses, for instance. The AlphaGo team has published its research in the journal Nature.

Artificial intelligence researchers say that this is a "sign of how far AI has come".

Computers like AlphaGo and Deep Blue, the machine that beat Garry Kasparov in 1997, are artificial intelligences, but they are intelligent in a highly specific way. The goal of some researchers is to develop an all-purpose intelligence, capable of solving all kinds of problems, as human brains are. That goal is known as "artificial general intelligence" (AGI).

AlphaGo's victory is a step along that road, says Shanahan, because of the generalisable way that it learns. He thinks that the techniques the AlphaGo team have used are the most promising route to AGI.

It's also a demonstration of just how powerful AI is now, and how quickly the field is moving, says Yudkowsky. "I'm not saying that AlphaGo in and of itself is going to lead to robots in 10 years," he says. "We just don't know about that. But AlphaGo is a sign of how far AI has come."

Although they warn there's a long way from here to true, human-level intelligence.

"Go is a tremendously complex game," says Shanahan. "But the everyday world is very, very much more complex." After all, he says, the real world contains Go, and chess, and driving cars. "The space of possible moves in the real world is truly huge. For example, the space of possible moves includes becoming a champion Go player."

He thinks that AGI is extremely unlikely in the next 10 years, but possible by 2050 and pretty likely by the end of the century. "This is not just a fantasy," he says. "We're talking about something that might actually affect our children, if not ourselves." Van den Driessche agrees, saying this is a "major milestone", but warning that human-level AGI is "still decades away".

"Nobody knows how long the road is [to AGI]," says Yudkowsky. "But we're pretty sure there's a long way left."

Still, they say, AlphaGo has shown that surprises happen. And AGI has the potential to be a big enough problem that it's worth paying attention now.

Handout / Getty Images

"People think that because it's not right around the corner, that means there's nothing to worry about," says Yudkowsky. "People cannot pry these ideas apart."

He thinks that humanity has already dropped the ball somewhat on thinking about what will happen when we build a machine that's as smart as us. The risk, he says, is that an intelligent machine that can rewrite its own code could improve itself very rapidly, and become far more intelligent than us. If it's not built with humanity's best interests at its core, it could end badly for us.

"We should have been thinking about this 30 years ago." He and the philosopher Nick Bostrom have been thinking for a while about how to minimise the risks AI poses to humanity, but, he says, we should be much further along that road already. "We should be on the technical stuff, the nitty gritty."

"The scenarios that are discussed by Bostrom and Yudkowksy are legitimate and need to be taken seriously," says Shanahan. "When we get to the point of building AGI I think we will quite quickly get to a superintelligence. We need to be absolutely sure it's safe."

And AlphaGo has shown, too, that there's no reason to think that any future artificial intelligences need to be anything like us.

"In 1997, Garry Kasparov said that he sensed a kind of alien intelligence on the other side of the board," says Shanahan. "And I've noticed in the commentary on AlphaGo that some of the commentators thought that it had made some weak moves earlier on, but now they're not sure if it wasn't some clever plan for the end game. Sometimes an AI might solve things in a way that's quite different from how we might tackle things."

A future AGI, in a much more dramatic way, might not be "human" either. There's no reason to think it would have our desires, or even things that we'd call desires at all, says Shanahan. Bostrom has pointed out that there's no reason to think it would even be conscious.

What's not clear, yet, is just how good AlphaGo is.

By Katpatuka. Wikimedia, CC BY-SA 3.0 / Via en.wikipedia.org

Obviously it's capable of beating the best human in the world, but how much better is it? Would it win every series?

"The games looked even," says Yudkowsky, "but is that because AlphaGo is an alien intelligence, or because they're actually close?" He says that in the fourth game, when it lost to Lee, some of its decisions looked more like mistakes, but there were some genuinely weird, superhuman moves in the second and third games – moves that looked wrong at the time, but which set up winning positions later in the game.

And it's worth remembering that there were two players in the series, and that Lee Sedol played extraordinarily well. "We're all honoured to have had the privilege to pit our creation against such a distinguished and capable opponent," says van den Driessche.