What made move 37 so interesting is that no one expected it. It was early in game two of the million-dollar Google DeepMind Challenge Match, and AlphaGo, an artificial intelligence (AI) system developed by Google, placed its 19th stone on a part of the game board that no human Go master would have considered. Some called it a “mistake.” Others called it “creative” and “unique.” But considering that AlphaGo went on to win its third game in a row against one of the strongest Go players in the world, the move should probably have been called what it really was: “intuitive.”
Note: The Google DeepMind Challenge Match was completed on March 15, 2016. Final score: AlphaGo 4, Lee Sedol 1.
Update: AlphaGo vs Ke Jie was a three-game Go match between AlphaGo and current world No. 1 ranking player Ke Jie played on 23, 25, and 27 May 2017. AlphaGo won 3 games to 0. Reflecting on the match, Ke Jie said, “It was like playing a Go god.”
Turing Would Love This
In 1959 Arthur Samuel began to teach a computer to play checkers, thinking that it was a good model for rudimentary problem solving. He defined machine learning as “a field of study that gives computers the ability to learn without being explicitly programmed.”
Back then, Samuel’s definition of the verb “to learn” was operational, not cognitive. But that subtlety is usually lost in translation. People always argue about whether or not computers can think. It’s the wrong argument. Paraphrasing from Alan Turing’s famous paper, “Computing Machinery and Intelligence,” let’s not ask the question, “Can machines think?” Let’s ask, “Can machines perform the way we (who can think) do?” (For more, see Can Machines Really Learn?)
Which brings us to the current challenge. We’ve seen computers beat humans at several contests of “human” intellect. Back in 1997, IBM’s Deep Blue supercomputer beat world chess champion Gary Kasparov in a very public match. In 2011, IBM’s Watson AI system beat Brad Rutter and Ken Jennings on the television game show Jeopardy! But the game of Go is different. A chess player may have to contemplate 20 to 35 moves per turn. A Go player is faced with 10 times that number. Numerically, the possible board combinations in an average 150-move game are vast (on the order of 10170). Google says that is greater than the number of atoms in the universe. (I’ve been told by several people that there are an estimated 1080 atoms in the universe.)
It is the exceptionally large number of possible moves that sets Go apart from other gameplay-based demonstrations of AI. Aside from logic, the world’s best human Go players win by using a combination of strategy, instinct and intuition. Go has so many moves, a computer cannot win by calculating all of the them – it must learn to “perform the way we (who can think) do.”
Back to Game 2, Move 37
How did AlphaGo “decide” to make this unusual move? AlphaGo lead project manager David Silver said that AlphaGo’s policy network has a model of what humans would do in this situation. After evaluating the high-probability moves, it starts to consider less probable moves, thinking ahead and considering potential futures. When asked if AlphaGo has a “human bias,” Silver went on to say, “AlphaGo will explore the human probability moves more thoroughly; this is its bias and it uses this to guide it toward its initial estimate. We train our neural networks on human data, so that does provide a bias, but that bias is a guide for the search. … It can always overwhelm that bias by searching more deeply and analyze things in an introspective way.”
In other words, AlphaGo doesn’t play by trying every combination (it can’t; there are too many possible moves). AlphaGo thinks, tries stuff, plays by feel and learns from its mistakes — it’s “thinking” more like us than any machine has thought before.
AlphaGo vs. You
AlphaGo has demonstrated a huge leap forward in AI and machine learning. The speed at which this two-year-old team evolved this system is truly awe-inspiring. Where does it lead? Well … if you can teach AlphaGo to be almost unbeatable (by a human) at Go, imagine what else you might be able to teach AlphaGo to do.
If you analyze reports for a living, move numbers from one cell in Excel to another, play “What if,” project manage or evaluate productivity in almost any way, a system with AlphaGo’s capabilities is going to learn how to do your job. It will be better at it than you could ever be, which leads to only one logical conclusion: your job function will become a computer function – a couple of clicks on a screen, and AlphaGo will do the rest.
I Think for a Living!
Yes, you do! But so does AlphaGo, and soon a purpose-built version will be able to do almost every low-level, most mid-level and some high-level white-collar jobs. Importantly, this type of AI will always outperform its human competition. Of course AlphaGo can lose, underperform or make a subjectively or objectively “bad” decision. But the future is clear — no white-collar job is safe. Not yours, not mine, not anyone’s.
This kind of AI can read, write, recognize natural language, recognize pictures, pattern match, simulate, optimize – in fact the only good news is that no one has any idea how to transfer neural network capabilities between disciplines. AlphaGo is dangerous to 9-dan Go masters, but harmless to people who optimize media purchases. But AlphaMedia (hypothetically) would always out-optimize them. That said, according to DeepMind founder Demis Hassabis, Google’s goal is to develop a generalized AI system – a system that could build on its knowledge and apply its learning to anything. This is an awesome goal, as in, it should fill you with awe!
Thrilled and Scared
I am thrilled by the success of the AlphaGo team and I am absolutely humbled by the power of what they have created. And it really, really scares me. Not because I don’t understand it, but because I do.
AI is a powerful tool. With this breakthrough, we are getting close to the dividing line between raw computing power and cognition – between craft and creativity – between machine and human. Ray Kurzweil has predicted a “Singularity” (where men and machines merge) for quite some time – and, as predicted by Kurzweil’s often-quoted Law of Accelerating Returns, we are closer to it than ever.
Thinking machines will have the capacity to heal the sick, feed the hungry and help us predict and survive natural disasters. They will make amazing lawyers, accountants, doctors, researchers, managers, writers and hundreds of other kinds of workers. They will also make exceptional productivity partners for work, entertainment and the doing of life. Like all technologies, they will ultimately be used more creatively than we can currently imagine.
But here I have to urge caution. Thinking machines will also learn to fight, they will learn to create computer viruses unlike any the world has ever known, they will level the playing field between good guys and bad guys in ways no one can really predict and they will impose symmetry on warfare that is currently asymmetrical – which is what scares me the most.
Congratulations to Google, DeepMind and the incredible team of engineers, scientists and coders who have just changed the world. Alan Turing and Arthur Samuel would be proud.