Demis Hassabis is the CEO of the Artificial Intelligence company DeepMind. In a recent interview on the Possible podcast, as shared by Kara Swisher's On podcast, he was asked to talk about the significance of Move 37. This is a moment well-known in the AI community, but not widely discussed, understood, or even heard of anywhere else.
Before I explain it, I will tell you I was immediately drawn in by two things.
First, I love it when I encounter a new word or phrase, or a new way to use an existing word or phrase. (I've been accused of overusing “bruh” in my daily interactions, for example. Because I LOVE how that word functions to express different things at different times.) Wow that example makes me feel like I sound stupid.
"I am NOT stupid." - Alexander Hamilton as imagined by Lin Manuel Miranda.
I wish to defend myself by saying I was also recently accused of using the word “colloquialism” too many times in one conversation. Am I redeemed?
Second, esoteric language is fascinating. Every profession, skill, hobby, or other collection of related minds develops its own language that is a shortcut to larger ideas or concepts. The simplest way to explain this is that while you and I might call a bunch of different things “skateboarding”, skaters in fact have words that describe complex sequences of moves. They even have words for moves or actions that are indistinguishable from one another to the casual observer. For example, a “fakie” is the term for riding the skateboard in your usual position (determined by which foot is in front) but the tail of your skateboard is leading. Here's a fairly comprehensive list of skateboarding maneuvers and links to their meanings. I love esoteric terms because they reveal complexity disguised as simplicity.
They also help reveal why expertise is important to solving big problems. Big problems are comprised of dozens of small problems if you're building a stone wall, or millions of them if you're trying to slow the gradual planetary warming. And if you've never thought of the small problems you're unlikely to understand the bigger problem very well. That is, if you've never grappled with the concept of a "fakie", good luck skating that concrete half pipe, champ! Wear a helmet!
So when Move 37 was mentioned on the podcast, and everyone involved in the conversation seemed to lean in as Demis Hassabis started to discuss its significance, I was leaning in too.
To understand Move 37, you need a modest understanding of the game “Go.” In Go, players alternately lay black and white stones on a grid, attempting to capture territory by walling it off or capturing opponents' stones. It's played on a 19x19 grid board. The immense size of the board makes for a stunningly large number of possible moves and outcomes. For comparison, the game of chess, universally accepted as simple to learn but complex to master, has a board that is 8x8.
(Image of a Go board by Elena Popova on Unsplash.)
So when computers were initially being programmed to play chess, they used “brute force” computing. That is, they simply played out as many versions of the game as they could from each point, and made the move that put them in the best position. But with Go, the amount of compute required to solve the task by brute force was insurmountable. The amount of energy it took to play out thousands of scenarios on a 19x19 grid was essentially infinite. They needed to teach the computer how the game worked.
They thought they had cracked the code and were watching one sample game when the computer did something really weird. It's 37th move of the game didn't make sense. Humans watching immediately pointed out the weirdness of it. The developers of the AI examined the code to see what had gone wrong. The game was lost, or at least the computer had wasted a move for no good reason.
Or so they thought.
Roughly 80 moves later, the computer used the stone placed in Move 37 to give itself an insurmountable advantage and win the game. The decision made sense in the larger context of the game. For the first time in DeepMind's experience, the computer had “thought” in a way that a human would - dozens of steps ahead, with an understanding of long-term strategy.
Move 37 is now understood to be a pivotal moment in AI development, as a symbol for what was possible.