Thirty-seven moves into one of the most historic games ever played, the machine did something incredible.
In March 2016 more than 60 million people around the world watched an artificial intelligence programme called , one of the most fiendishly complex board games conceived by humanity. On the 37th move in the second of five games, AlphaGo unleashed a move no human would ever play, stunning experts and fans and utterly wrong-footing world champion Lee Sedol.
Go experts were puzzled by AlphaGo's moves until they realised that where a human might play to win by a wide margin, the AI didn't care if it won by a single point. So it pursued only the slimmest advantage necessary to carry the game.
As compelling and dramatically human as a sports movie, documentary "AlphaGo" goes behind the scenes of the match in Seoul, South Korea. It puts you in the room to see how even the AI's creators, Google-backed British startup DeepMind, didn't fully understand what AlphaGo was doing with manoeuvres like move 37.
Making the movie
The movie premiered at the Tribeca film festival earlier this year, and is expected to be available on streaming services soon. I met with the film's director, Greg Kohs, to find out what it was like to witness the historic match.
How did the film come about?
Greg Kohs: It had very modest beginnings. A couple members of Google's creative lab that I'd worked with before gave me a ring and said we'd have access behind the curtain with [DeepMind founder and CEO] Demis Hassabis and his team. So I jumped on board with the expectation we would just film what happens for archival purposes and then put it on a shelf on a hard drive and that would be the end of it.
But then along the journey, in Seoul, there was a moment when it became apparent to myself that it felt like this is a really beautiful story, a story with a really beautiful beginning, middle and end.
With the secrecy around DeepMind, were you able to film everything?
Greg Kohs: Demis and DeepMind were completely transparent. It was just myself and my sound person, and camera assistant. I was doing the filming, so it was just a very small and trusted unit. We had the run of the place. Demis was mostly interested in making sure that his team took a look at the screens that were being shown to make sure they were shareable.
Did the match feel like a publicity stunt to you, or did it feel like a genuine milestone for AI?
Greg Kohs: There was a lot of hype and spectacle around it, but it didn't ever feel like a stunt to me. Early in my film making career I spent ten years making movies about American football, and I was always drawn to the pageantry and spectacle of a game more than the game itself -- the goosebumps, the jet flyovers, all that stuff. And this had all that, in its own way. Or more like a boxing match in a Vegas hotel.
The moment where it felt like history was the first move, when everyone was waiting for AlphaGo to make that first move. Once it made that move, I felt inside, holy cow, it's on. This is important, this is gonna be history now, whatever the outcome.
It was fascinating to see the DeepMind team watching AlphaGo play and they weren't sure why it was choosing the moves it did. How did that feel, to see that play out?
Greg Kohs: I was looking for cues from [the DeepMind team], and when they say they're puzzled by that move, it's not logical, it was definitely interesting … It was exciting because things were happening no-one quite understood. Afterward, they were so excited. It was electric to listening to them talk, especially after they lost.
Man vs. machine
In October DeepMind revealed that the matchwinning AlphaGo programme had been superceded. The programme seen in the film learned the game by studying matches played by humans, but a new version calledsimply by playing against itself millions of times in a few weeks, without needing to learn anything from human players.
The spectacle of man versus machine is something that humans can easily identify with. But should we be thinking of ourselves as being in competition with AI?
Greg Kohs: I suppose our intent with the film is it's just one moment within the development and the life of overarching artificial intelligence. I wouldn't expect someone to say "Well this is all you need to know about AI." it's just one sentence or one word in a very big volume of AI research.
People have said man versus machine; it's really man with machine ... I loved getting a deeper appreciation for machine learning, and the whole idea of it being assistance [for humans], as opposed to replacement ... The Go community is thrilled by AlphaGo coming about, to help them learn more, to unlock these interesting moves. I take hope from that. I hope that in the future, the medical community would look at it the same way. A doctor on his 12th hour in the ER might have some tool to assist them in their diagnoses.
Do you think DeepMind's goal of making a programme to play a board game was perhaps naive given the possible wider implications of AI's rise?
Greg Kohs: I feel that from the beginning, they took the responsibility extremely seriously. They've now created an ethics board which shows they're thinking about checks and balances.
I learned a lot from [speaking to experts for the film]. Here are some of the deepest thinkers in the world, that are comfortable with where we're at. That gave me comfort.
What was it like to see totally unexpected strategies like move 37 that human players typically wouldn't play?
Greg Kohs: AlphaGo is reminding us it's not always about getting more, it's just about having what you need. When I learned that was what was happening, I immediately couldn't wait to get on the phone with my kids. I love that lesson that it was reminding us, it wasn't about just more and more and more. Do you need more? What do you need truly to be happy? I couldn't wait to call my kids.
They were like, "whatever Dad." [Laughs.]
Tech Culture: From film and television to social media and games, here's your place for the lighter side of tech.
Tech Enabled: CNET chronicles tech's role in providing new kinds of accessibility.