Yet Another Great Victory For Humanity
As I have written in the past, the improvements in AI research have gone from linear to parabolic. The recent AlphaGo’s victory against 18-time world champion Lee Se-dol last week is just another clear signal of it. Go has officially fallen to machine, just as Jeopardy did before it to Watson (2010), and chess before that to Deep Blue (1997). Even if it got an unprecedented media coverage for an AI milestone, I would like to take a few minutes to reflect on this event and share my thoughts.
Firstly, just as a quick reminder, when asked to estimate when a computer would defeat a prominent Go player, Rémi Coulom would reply “Maybe in another ten years”. This interview was given only 2 years ago and at that time, Coulom’s computer program, Crazy Stone, was the very best AI at playing Go. It is easy to smile now at Coulom’s estimates but one should understand the greatest of the challenge first. Google’s DeepMind artificial intelligence team likes to say that there are more possible Go boards than atoms in the known universe. But as Ken Jennings a former Jeopardy champion, underlines it, that vastly understates the computational problem. According to him:
“There are about 1⁰¹⁷⁰ board positions in Go, and only 1⁰⁸⁰ atoms in the universe. That means that if there were as many parallel universes as there are atoms in our universe (?!), then the total number of atoms in all those universes combined would be close to the possibilities on a single Go board.”
That rough estimate alone makes us realize that any brute-force approach to scan every possible move would be hopeless. And that is the greatest of the deep neural networks approach as it gets around that barrier in the same way our own minds do, by learning to estimate what feels like the best move. Before the game AlphaGo had been analyzing millions of professional games and playing itself millions of time.
This leads to my second observation. From my point of view, the series of Go matches between AlphaGo and Lee Sedol was the best public demonstration that machine is much better at learning from experience than humans are. The reason is knowledge sharing vs human experience which is often a candle that only lights oneself. How many times parents tell their kids not to do something and yet they have to experience it themselves to then realize it was a bad idea? At the contrary, Google self-driving cars are said to mutualized their experience with every other self-driving car on the planet.
Thirdly, some of AlphaGo’s moves have been described not just as “beautiful”, but moves that no human could understand, and much less anticipate. The combination of deep learning with reinforcement learning, pitting a machine against itself and thus creating new moves from others and that are then fed back into the system have been brilliantly demonstrated.
Fourthly, as always this AI event generated various fear reactions all around the web. It is important to remind everyone that AlphaGo is a task-specific AI. It is exceptionally good at doing exactly one thing but that is pretty much it. When people discuss the fear around AI they usually talk about something very different: an artificial general intelligence (AGI). An AGI would be a machine that is equally capable of learning Go or learning chess, understanding politics (including Donald Trump), the rules of rugby, and any other aspect of human life. It would be able to grow and adapt to new challenges rather than just being limited to a specific task. This kind is the one on which concerned was raised by Nick Bostrom, Stephen Hawking, Elon Musk, Bill Gates, and others. Fortunately, we do not really know what an AGI would look like since we don’t yet know how to build one. What we do know, however — what AlphaGo has so beautifully demonstrated — is that before we get to AGIs, we are going to have more and more highly capable task-specific AIs at our disposal.
That being said, beating expert humans at chess, Jeopardy, or the old Atari games is undoubtedly amusing but not so damaging. What the bigger picture tells us is that AIs are soon going to be beating us at just about any task you can imagine, whether it’s working or playing. Very soon we will see AIs taking strategic business and public decisions such as fixing the Central Banks interest rates, tax policies, or how much of a pension we receive. It would hopefully be for the best. We simply need to try to wrap our heads around it and think ahead as it would imply a full redesign of some of our society pillars: wealth distribution, the role of humans, the development of society, among others.
To be honest, I was surprised by the confidence of many Go experts, including Mr. Lee himself, predicting that the machine would be defeated 5–0 or, if it did well, 4–1. I am absolutely not familiar with the Go community but I was not expecting trash talking to be that high, so I can only imagine how much of a shockwave it has been for them. Luckily, having a new master is a fantastic news for them as it will bring the game to yet a higher level by learning from it, as the chess community did with Deep Blue these last two decades.