The Creativity Code: How AI is learning to write, paint and think

The challenge of creating a machine that could learn to play Go still felt like a distant dream. They set their sights at first on a seemingly less cerebral goal: playing 1980s Atari games. Atari is probably responsible for a lot of students flunking courses in the late 1970s and early 1980s. I certainly remember wasting a huge amount of time playing the likes of Pong, Space Invaders and Asteroids on a friend’s Atari 2600 console. The console was one of the first whose hardware could play multiple games that were loaded via a cartridge. It allowed a whole range of different games to be developed over time. Previous consoles could only play games that had been physically programmed into the units.

One of my favourite Atari games was called Breakout. A wall of coloured bricks was at the top of the screen and you controlled a paddle at the bottom that could be moved left or right using a joystick. A ball would bounce off the paddle and head towards the bricks. Each time it hit a brick, the brick would disappear. The aim was to clear the bricks. The yellow bricks at the bottom of the wall scored one point. The red bricks on top got you seven points. As you cleared blocks, the paddle would shrink and the ball would speed up to make the game play harder.

We were particularly pleased one afternoon when we found a clever way to hack the game. If you dug a tunnel up through the bricks on the edge of the screen, once the ball made it through to the top it bounced back and forward off the top of the screen and the upper high-scoring bricks, gradually clearing the wall. You could sit back and watch until the ball eventually came back down through the wall. You just had to be ready with the paddle to bat the ball back up again. It was a very satisfying strategy!

Hassabis and the team he was assembling also spent a lot of time playing computer games in their youth. Their parents may be happy to know that the time and effort they put into those games did not go to waste. It turned out that Breakout was a perfect test case to see if the team at DeepMind could program a computer to learn how to play games. It would have been a relatively straightforward job to write a program for each individual game. Hassabis and his team were going to set themselves a much greater challenge.

They wanted to write a program that would receive as an input the state of the pixels on the screen and the current score and set it to play with the goal of maximising the score. The program was not told the rules of the game: it had to experiment randomly with different ways of moving the paddle in Breakout or firing the laser cannon at the descending aliens of Space Invaders. Each time it made a move it could assess whether the move had helped increase the score or had had no effect.

The code implements an idea dating from the 1990s called reinforcement learning, which aims to update the probability of actions based on the effect on a reward function or score. For example, in Breakout the only decision is whether to move the paddle at the bottom left or right. Initially the choice will be 50:50. But if moving the paddle randomly results in it hitting the ball, then a short time later the score goes up. The code then recalibrates the probability of whether to go left or right based on this new information. It will increase the chance of heading in the direction towards which the ball is heading. The new feature was to combine this learning with neural networks that would assess the state of the pixels to decide what features were correlating to the increase in score.

At the outset, because the computer was just trying random moves, it was terrible, hardly scoring anything. But each time it made a random move that bumped up the score, it would remember the move and reinforce the use of such a move in future. Gradually the random moves disappeared and a more informed set of moves began to emerge, moves that the program had learned through experiment seemed to boost its score.

It’s worth watching the supplementary video the DeepMind team included in the paper they eventually wrote. It shows the program learning to play Breakout. At first you see it randomly moving the paddle back and forward to see what will happen. Then, when the ball finally hits the paddle and bounces back and hits a brick and the score goes up, the program starts to rewrite itself. If the pixels of the ball and the pixels of the paddle connect that seems to be a good thing. After 400 game plays it’s doing really well, getting the paddle to continually bat the ball back and forward.

The shock for me came when you see what it discovered after 600 games. It found our hack! I’m not sure how many games it took us as kids to find this trick, but judging by the amount of time I wasted with my friend it could well have been more. But there it is. The program manipulated the paddle to tunnel its way up the sides, such that the ball would be stuck in the gap between the top of the wall and the top of the screen. At this point the score goes up very fast without the computer’s having to do very much. I remember my friend and I high-fiving when we’d discovered this trick. The machine felt nothing.

By 2014, four years after the creation of DeepMind, the program had learned how to outperform humans on twenty-nine of the forty-nine Atari games it had been exposed to. The paper the team submitted to Nature detailing their achievement was published in early 2015. To be published in Nature is one of the highlights of a scientist’s career. But their paper achieved the even greater accolade of being featured as the cover story of the whole issue. The journal recognised that this was a huge moment for artificial intelligence.

It has to be reiterated what an amazing feat of programming this was. From just the raw data of the state of the pixels and the changing score, the program had changed itself from randomly moving the paddle of Breakout back and forth to learning that tunnelling the sides of the wall would win you the top score. But Atari games are hardly on a par with the ancient game of Go. Hassabis and his team at DeepMind decided they were ready to create a new program that could take it on.

It was at this moment that Hassabis decided to sell the company to Google. ‘We weren’t planning to, but three years in, focused on fundraising, I had only ten per cent of my time for research,’ he explained in an interview in Wired at the time. ‘I realised that there’s maybe not enough time in one lifetime to both build a Google-sized company and solve AI. Would I be happier looking back on building a multi-billion business or helping solve intelligence? It was an easy choice.’ The sale put Google’s firepower at his fingertips and provided the space for him to create code to realise his goal of solving Go … and then intelligence.

First blood

Previous computer programs built to play Go had not come close to playing competitively against even a pretty good amateur, so most pundits were highly sceptical of DeepMind’s dream to create code that could get anywhere near an international champion of the game. Most people still agreed with the view expressed in The New York Times by the astrophysicist Piet Hut after DeepBlue’s success at chess in 1997: ‘It may be a hundred years before a computer beats humans at Go – maybe even longer. If a reasonably intelligent person learned to play Go, in a few months he could beat all existing computer programs. You don’t have to be a Kasparov.’

Just two decades into that hundred years, the DeepMind team believed they might have cracked the code. Their strategy of getting algorithms to learn and adapt appeared to be working, but they were unsure quite how powerful the emerging algorithm really was. So in October 2015 they decided to test-run their program in a secret competition against the current European champion, the Chinese-born Fan Hui.

AlphaGo destroyed Fan Hui five games to nil. But the gulf between European players of the game and those in the Far East is huge. The top European players, when put in a global league, rank in the 600s. So, although it was still an impressive achievement, it was like building a driverless car that could beat a human driving a Ford Fiesta round Silverstone then trying to challenge Lewis Hamilton in a Grand Prix.

Certainly when the press in the Far East heard about Fan Hui’s defeat they were merciless in their dismissal of how meaningless the win was for AlphaGo. Indeed, when Fan Hui’s wife contacted him in London after the news got out, she begged her husband not to go online. Needless to say he couldn’t resist. It was not a pleasant experience to read how dismissive the commentators in his home country were of his credentials to challenge AlphaGo.

Fan Hui credits his matches with AlphaGo with teaching him new insights into how to play the game. In the following months his ranking went from 633 to the 300s. But it wasn’t only Fan Hui who was learning. Every game AlphaGo plays affects its code and changes it to improve its play next time around.

It was at this point that the DeepMind team felt confident enough to offer their challenge to Lee Sedol, South Korea’s eighteen-time world champion and a formidable player of the game.

The match was to be played over five games scheduled between 9 and 15 March 2016 at the Four Seasons hotel in Seoul, and would be broadcast live across the internet. The winner would receive a prize of a million dollars. Although the venue was public, the precise location within the hotel was kept secret and was isolated from noise – not that AlphaGo was going to be disturbed by the chitchat of the press and the whispers of curious bystanders. It would assume a perfect Zen-like state of concentration wherever it was placed.

Sedol wasn’t fazed by the news that he was up against a machine that had beaten Fan Hui. Following Fan Hui’s loss he had declared: ‘Based on its level seen … I think I will win the game by a near landslide.’

Although he was aware of the fact that the machine he would be playing was learning and evolving, this did not concern him. But as the match approached, you could hear doubts beginning to creep into his view of whether AI will ultimately be too powerful for humans to defeat it even in the game of Go. In February he stated: ‘I have heard that DeepMind’s AI is surprisingly strong and getting stronger, but I am confident that I can win … at least this time.’

Most people still felt that despite great inroads into programming, an AI Go champion was still a distant goal. Rémi Coulom, the creator of Crazy Stone, the only program to get close to playing Go at any high standard, was still predicting another decade before computers would beat the best humans at the game.

As the date for the match approached, the team at DeepMind felt they needed someone to really stretch AlphaGo and to test it for any weaknesses. So they invited Fan Hui back to play the machine going into the last few weeks. Despite having suffered a 5–0 defeat and being humiliated by the press back in China, he was keen to help out. Perhaps a bit of him felt that if he could help make AlphaGo good enough to beat Sedol, it would make his defeat less humiliating.

As Fan Hui played he could see that AlphaGo was extremely strong in some areas but he managed to reveal a weakness that the team was not aware of. There were certain configurations in which it seemed to completely fail to assess who had control of the game, often becoming totally delusional that it was winning when the opposite was true. If Sedol tapped into this weakness, AlphaGo wouldn’t just lose, it would appear extremely stupid.

The DeepMind team worked around the clock trying to fix this blind spot. Eventually they just had to lock down the code as it was. It was time to ship the laptop they were using to Seoul.

The stage was set for a fascinating duel as the players, or at least one player, sat down on 9 March to play the first of the five games.

‘Beautiful. Beautiful. Beautiful’

It was with a sense of existential anxiety that I fired up the YouTube channel broadcasting the matches that Sedol would play against AlphaGo and joined 280 million other viewers to see humanity take on the machines. Having for years compared creating mathematics to playing the game of Go, I had a lot on the line.

Lee Sedol picked up a black stone and placed it on the board and then waited for the response. Aja Huang, a member of the DeepMind team, would play the physical moves for AlphaGo. This, after all, was not a test of robotics but of artificial intelligence. Huang stared at AlphaGo’s screen, waiting for its response to Sedol’s first stone. But nothing came.

We all stared at our screens wondering if the program had crashed! The DeepMind team was also beginning to wonder what was up. The opening moves are generally something of a formality. No human would think so long over move 2. After all, there was nothing really to go on yet. What was happening? And then a white stone appeared on the computer screen. It had made its move. The DeepMind team breathed a huge sigh of relief. We were off! Over the next couple of hours the stones began to build up across the board.

One of the problems I had as I watched the game was assessing who was winning at any given point in the game. It turns out that this isn’t just because I’m not a very experienced Go player. It is a characteristic of the game. Indeed, this is one of the main reasons why programming a computer to play Go is so hard. There isn’t an easy way to turn the current state of the game into a robust scoring system of who leads by how much.

Chess, by contrast, is much easier to score as you play. Each piece has a different numerical value which gives you a simple first approximation of who is winning. Chess is destructive. One by one pieces are removed so the state of the board simplifies as the game proceeds. But Go increases in complexity as you play. It is constructive. The commentators kept up a steady stream of observations but struggled to say if anyone was in the lead right up until the final moments of the game.

What they were able to pick up quite quickly was Sedol’s opening strategy. If AlphaGo had learned to play on games that had been played in the past, then Sedol was working on the principle that it would put him at an advantage if he disrupted the expectations it had built up by playing moves that were not in the conventional repertoire. The trouble was that this required Sedol to play an unconventional game – one that was not his own.

It was a good idea but it didn’t work. Any conventional machine programmed on a database of accepted openings wouldn’t have known how to respond and would most likely have made a move that would have serious consequences in the grand arc of the game. But AlphaGo was not a conventional machine. It could assess the new moves and determine a good response based on what it had learned over the course of its many games. As David Silver, the lead programmer on AlphaGo, explained in the lead-up to the match: ‘AlphaGo learned to discover new strategies for itself, by playing millions of games between its neural networks, against themselves, and gradually improving.’ If anything, Sedol had put himself at a disadvantage by playing a game that was not his own.

As I watched I couldn’t help feeling for Sedol. You could see his confidence draining out of him as it gradually dawned on him that he was losing. He kept looking over at Huang, the DeepMind representative who was playing AlphaGo’s moves, but there was nothing he could glean from Huang’s face. By move 186 Sedol had to recognise that there was no way to overturn the advantage AlphaGo had built up on the board. He placed a stone on the side of the board to indicate his resignation.

By the end of day one it was: AlphaGo 1 Humans 0. Sedol admitted at the press conference that day: ‘I was very surprised because I didn’t think I would lose.’

But it was game 2 that was going to truly shock not just Sedol but every human player of the game of Go. The first game was one that experts could follow and appreciate why AlphaGo was playing the moves it was. They were moves a human champion would play. But as I watched game 2 on my laptop at home, something rather strange happened. Sedol played move 36 and then retired to the roof of the hotel for a cigarette break. While he was away, AlphaGo on move 37 instructed Huang, its human representative, to place a black stone on the line five steps in from the edge of the board. Everyone was shocked.

The conventional wisdom is that during the early part of the game you play stones on the outer four lines. The third line builds up short-term territory strength on the edge of the board while playing on the fourth line contributes to your strength later in the game as you move into the centre of the board. Players have always found that there is a fine balance between playing on the third and fourth lines. Playing on the fifth line has always been regarded as suboptimal, giving your opponent the chance to build up territory that has both short- and long-term influence.

AlphaGo had broken this orthodoxy built up over centuries of competing. Some commentators declared it a clear mistake. Others were more cautious. Everyone was intrigued to see what Sedol would make of the move when he returned from his cigarette break. As he sat down, you could see him physically flinch as he took in the new stone on the board. He was certainly as shocked as all of the rest of us by the move. He sat there thinking for over twelve minutes. Like chess, the game was being played under time constraints. Using twelve minutes of your time was very costly. It is a mark of how surprising this move was that it took Sedol so long to respond. He could not understand what AlphaGo was doing. Why had the program abandoned the region of stones they were competing over?

Was this a mistake by AlphaGo? Or did it see something deep inside the game that humans were missing? Fan Hui, who had been given the role of one of the referees, looked down on the board. His initial reaction matched everyone else’s: shock. And then he began to realise: ‘It’s not a human move. I’ve never seen a human play this move,’ he said. ‘So beautiful. Beautiful. Beautiful. Beautiful.’

Beautiful and deadly it turned out to be. Not a mistake but an extraordinarily insightful move. Some fifty moves later, as the black and white stones fought over territory from the lower left-hand corner of the board, they found themselves creeping towards the black stone of move 37. It was joining up with this stone that gave AlphaGo the edge, allowing it to clock up its second win. AlphaGo 2 Humans 0.

Sedol’s mood in the press conference that followed was notably different. ‘Yesterday I was surprised. But today I am speechless … I am in shock. I can admit that … the third game is not going to be easy for me.’ The match was being played over five games. This was the game that Sedol needed to win to be able to stop AlphaGo claiming the match.

The human fight-back

Sedol had a day off to recover. The third game would be played on Saturday, 12 March. He needed the rest, unlike the machine. The first game had been over three hours of intense concentration. The second lasted over four hours. You could see the emotional toll that losing two games in a row was having on him.

Rather than resting, though, Sedol stayed up till 6 a.m. the next morning analysing the games he’d lost so far with a group of fellow professional Go players. Did AlphaGo have a weakness they could exploit? The machine wasn’t the only one who could learn and evolve. Sedol felt he might learn something from his losses.

Sedol played a very strong opening to game 3, forcing AlphaGo to manage a weak group of stones within his sphere of influence on the board. Commentators began to get excited. Some said Sedol had found AlphaGo’s weakness. But then, as one commentator posted: ‘Things began to get scary. As I watched the game unfold and the realisation of what was happening dawned on me, I felt physically unwell.’

Sedol pushed AlphaGo to its limits but in so doing he revealed the hidden powers that the program seemed to possess. As the game proceeded, it started to make what commentators called lazy moves. It had analysed its position and was so confident in its win that it chose safe moves. It didn’t care if it won by half a point. All that mattered was that it won. To play such lazy moves was almost an affront to Sedol, but AlphaGo was not programmed with any vindictive qualities. Its sole goal was to win the game. Sedol pushed this way and that, determined not to give in too quickly. Perhaps one of these lazy moves was a mistake that he could exploit.

By move 176 Sedol eventually caved in and resigned. AlphaGo 3 Humans 0. AlphaGo had won the match. Backstage, the DeepMind team was going through a strange range of emotions. They’d won the match, but seeing the devastating effect it was having on Sedol made it hard for them to rejoice. The million-dollar prize was theirs. They’d already decided to donate the prize, if they won, to a range of charities dedicated to promoting Go and science subjects as well as to Unicef. Yet their human code was causing them to empathise with Sedol’s pain.

AlphaGo did not demonstrate any emotional response to its win. No little surge of electrical current. No code spat out with a resounding ‘YES!’ It is this lack of response that gives humanity hope and is also scary at the same time. Hope because it is this emotional response that is the drive to be creative and venture into the unknown: it was humans, after all, who’d programmed AlphaGo with the goal of winning. Scary because the machine won’t care if the goal turns out to be not quite what its programmers had intended.

Sedol was devastated. He came out in the press conference and apologised:

I don’t know how to start or what to say today, but I think I would have to express my apologies first. I should have shown a better result, a better outcome, and better content in terms of the game played, and I do apologize for not being able to satisfy a lot of people’s expectations. I kind of felt powerless.

But he urged people to keep watching the final two games. His goal now was to try to at least get one back for humanity.

Having lost the match, Sedol started game 4 playing far more freely. It was as if the heavy burden of expectation had been lifted, allowing him to enjoy his game. In sharp contrast to the careful, almost cautious play of game 3, he launched into a much more extreme strategy called ‘amashi’. One commentator compared it to a city investor who, rather than squirrelling away small gains that accumulate over time, bet the whole bank.

Sedol and his team had stayed up all of Saturday night trying to reverse-engineer from AlphaGo’s games how it played. It seemed to work on a principle of playing moves that incrementally increase its probability of winning rather than betting on the potential outcome of a complicated single move. Sedol had witnessed this when AlphaGo preferred lazy moves to win game 3. The strategy they’d come up with was to disrupt this sensible play by playing the risky single moves. An all-or-nothing strategy might make it harder for AlphaGo to score so easily.

AlphaGo seemed unfazed by this line of attack. Seventy moves into the game, commentators were already beginning to see that AlphaGo had once again gained the upper hand. This was confirmed by a set of conservative moves that were AlphaGo’s signal that it had the lead. Sedol had to come up with something special if he was going to regain the momentum.

If move 37 of game 2 was AlphaGo’s moment of creative genius, move 78 of game 4 was Sedol’s retort. He’d sat there for thirty minutes staring at the board, staring at defeat, when he suddenly placed a white stone in an unusual position, between two of AlphaGo’s black stones. Michael Redmond, who was commentating on the YouTube channel, spoke for everyone: ‘It took me by surprise. I’m sure that it would take most opponents by surprise. I think it took AlphaGo by surprise.’