Ready, Set, Go!

“Yesterday I was surprised, but today it’s more than that — I am speechless,” said Lee in the post-game press conference. “I admit that it was a very clear loss on my part. From the beginning of the game, I did not feel like there was a point that I was leading.”

“I am in shock….” said Lee.

“Playing against a machine is very different from an actual human opponent,” (Lee) had told BBC. “Normally, you can sense your opponent’s breathing, their energy. And lots of times you make decisions which are dependent on the physical reactions of the person you’re playing against. With a machine you can’t do that.

It’s not hard for us to feel sympathy for Lee Sedol. Until March 15, he was able to legitimately call himself World Champion in the ancient Chinese game of Go.

Now, he must add the qualifier, “human” to his title.

As most of us know by now, Google’s AlphaGo defeated Lee 4–1 in a historic match in Seoul, Korea. He now joins an elite group of extraordinary humans who have been defeated in highly specialized games by artificial intelligence systems. First came IBM Deep Blue’s 1997 defeat of chess grandmaster Garry Kasparov, then IBM Watson’s trouncing of Jeopardy! champs Brad Rutter and Ken Jennings in 2011.

Chess and Jeopardy! were one thing, we were told before the Go match, but the Chinese game was several orders of magnitude more difficult for a machine to master. The prior two contests took place in domains of finite possibilities: build a searchable database of all possible chess moves in any specific board situation and chess becomes manageable for a modern computer; read all of Wikipedia (from which most Jeopardy! questions emanate) and winning becomes a simple matter of beating human reaction time to pressing the buzzer.

But Go, we learned, uses that holy grail of human decision-making: intuition.

Since the 19 x 19 square Go board enables 10¹⁷⁰ moves, a number vastly greater than the number of atoms in the universe (a paltry 10⁸⁰!) no “database” of possible moves could be created. Go champions often explain highly creative moves by stating simply, “it just felt right.”

Surely, no machine could match the elegance and subtlety of a champion of Lee’s stature.

So, how did this happen? And, what does it mean?

AlphaGo is a product of DeepMind Technologies, a company founded by Demis Hassabis in 2010 and acquired by Google in 2014 for $500 million. Hassabis describes the company as, “An Apollo Programme for AI.”

With the simplicity so often encountered in genius, Hassabis states DeepMind’s twofold mission as:

1. Solve intelligence
2. Use intelligence to solve everything else

Since its inception, DeepMind has focused on developing the two artificial intelligence components that led to its latest achievement: deep learning and reinforcement learning. The ingenious methods DeepMind’s scientists used to craft AlphaGo’s victory are described in detail in a video of a talk Hassabis gave at Oxford University two weeks prior to the match. A brief summary might go as follows: since it’s impossible to pre-program an AI to defeat a Go champion, Google’s engineers needed to find a way to teach AlphaGo to teach itself to play the game, which they did using a combination of neural networks and reinforcement learning.

OK, but, so what?

Why should anyone not vying for global Go dominance care about AlphaGo’s victory?

One reason is the look on Lee Sedol’s face in the photo shown above. The — what is it? — bewilderment? that we see feels like a preview of experiences many of us are likely to have over the next decade or so as one barrier after another between us and machines proves to be breachable.

This is not a comfortable prospect for most of us, but one for which we clearly need to prepare.

In his Oxford talk, Hassabis describes DeepMind’s next targets. Using the same fundamental techniques that defeated Go, his team will set their sights on some of humanity’s most essential cognitive and emotional capabilities: memory, attention, concepts, planning, navigation, and imagination. Doing so will enable them to apply these technologies to real world problems, starting with healthcare, robotics and smart assistants. More broadly, he sees “meta-solutions” emerging to tackle problems of information overload (in areas like big data, genomics, entertainment and personalization) and system complexity (to investigate climate, disease, energy, macroeconomics, and high energy physics).

And so, as disorienting as these moments of AI achievement will be, there’s little doubt that the resounding victory the world saw in Seoul will usher in an age of tremendous progress in tackling some of our most intractable problems.

In the process, however, Hassabis stressed that while human-level artificial general intelligence may be decades off, we must start a deep discussion of the technology’s enormous ethical implications.

Now.