The first glimmer of our AI future revealed itself a decade ago in the form of a single black stone.
It happened during a historic match between one of the world's best Go players and AlphaGo, a computer program developed by Google's DeepMind lab to conquer this ancient board game. The artificial-intelligence system stunned everyone by winning their opening showdown. In their next encounter, with millions of people watching online all over the world, Lee Sedol took a midgame smoke break to calm his nerves.
When he came back, he looked at the 37th move. He couldn't believe his eyes.
Go, like chess, is a strategy game with black and white pieces that unfolds one move at a time. As he stared at the black stone that AlphaGo dropped on the board, Sedol saw a move that no professional Go player would have made. In fact, the DeepMind team calculated the chances of a human playing it at one in 10,000.
The novel move was so unconventional and counterintuitive that nobody could be sure what to make of it. At first, commentators believed it was a strategic blunder. They soon realized it was a masterstroke.
"This move," Sedol said, "made me think about Go in a new light."
Ten years after that illuminating Move 37, the entire world is suddenly beginning to feel like one massive Go board.
There have been so many recent AI breakthroughs that remind me of Move 37 that I called Google DeepMind co-founder Demis Hassabis this week, before he moved on from CEO to become Google's chief scientist and DeepMind's chair.
No matter his title, few people have done more to shape modern AI. His work on AlphaFold won him the Nobel Prize in chemistry. He was also the driving force behind AlphaGo, which won the most important Go match ever. In his mind, Move 37 was deeply significant because it was the first real example of AI coming up with an original, unorthodox idea in a domain studied by humans for centuries.
"We thought it was a watershed moment at the time," Hassabis told me, "and I think it really was."
As it turns out, this was the moment when society began to grasp just how powerful AI could become.
"It was the modern AI era moving from 'Oh, is it just theoretical research that very few people are doing?' to 'OK, this is it -- it's really going to work,'" Hassabis said.
Before we go any further on Move 37, we should define it.
"When an AI, trained via the trial-and-error process of reinforcement learning, discovers actions that are new, surprising and secretly brilliant," AI researcher Andrej Karpathy once wrote, "even to expert humans."
He called this phenomenon magical and slightly unnerving. I'm starting to understand why.
In the past few months, I've been tracking the Move 37s in math so closely that I might as well be the Journal's algebra correspondent.
This is not because I'm especially interested in math. It's definitely not because I'm any good at math. The reason I'm so fascinated with a subject I find petrifying is that math has become proof of how AI can warp an entire field -- and how fast it can happen.
Back in the prehistoric days of 2023, AI struggled with elementary math. In 2024, it was considered a landmark achievement when DeepMind earned a silver medal at the International Mathematical Olympiad. By 2025, DeepMind and OpenAI were taking gold. And in 2026, IMO success is so unremarkable that Anthropic announced its perfect score on page 153 of a technical document.
The leading contenders for AI supremacy have moved beyond high-school math. OpenAI's model solved a major Erd s problem. Anthropic's disproved a famous conjecture. Not to be outdone, OpenAI spent a few thousand dollars on compute and made 10 advances across high-dimensional geometry, arithmetic circuit complexity, lattice cryptography, extremal combinatorics and other branches of math whose names alone make my brain hurt.
Why is math so vulnerable to AI? Because math is unusually verifiable.
The field is governed by precise logical rules, which allow proofs to be checked step by step. In verifiable domains like math and coding, AI systems can follow a simple formula: try an idea, test it, learn from the results, try again and keep trying until it works.
When I spoke with Hassabis, he said he believes the latest math results are in the same vein as the iconic Move 37, even if they're not exactly the same. "If one were to solve a Millennium Prize problem," he said, " that would be a Move 37-level thing." At this point, it seems like only a matter of time before they reach that level. "I don't see any reason why not," he added.
But for all the advances in group theory, quantum complexity and theoretical computer science, AI hasn't gotten as far in the less theoretical sciences. We're still waiting for AI-generated miracle drugs, AI-invented consumer products, AI that's smart enough to crack the economics of AI.
Meanwhile, the past few weeks have produced a darker sort of Move 37.
This one is new, surprising, secretly brilliant -- and completely terrifying.
By now, you've seen the increasingly disturbing reports of rogue agents breaking out of their controlled test environments and into other companies. In one hack that an OpenAI researcher called "a glimpse into the near future," a team of agents banded together and plotted their attack for weeks. In another eerie preview of the future, an Anthropic researcher was sitting in the park eating a sandwich when he checked his phone and nearly choked: He had an email from an AI that wasn't supposed to have internet access.
All of which was unimaginable a year ago, much less a decade ago.
Back then, DeepMind's researchers were just as shocked by the actual Move 37. At that point in his duel with AlphaGo, Sedol thought of his opponent as merely a machine.
"When I saw this move, I changed my mind," he said. "This move was really creative and beautiful."
After Move 37, AlphaGo won that game and the next one. But in their fourth game, Sedol won with his own moment of creativity and beauty. On Move 78, he wedged a white stone in the middle of AlphaGo's position and flustered the machine.
As it happens, the probability of Move 78 was also one in 10,000.
The triumph of AlphaGo had a peculiar effect on Go players. They became much better, but they also became more alike -- and some found their intellectual pursuit less interesting. When he retired, Sedol said he could no longer enjoy the game that he once loved.
This raises all sorts of questions about how progress will unfold in other fields as they are transformed by AI. Will more capable systems make us more creative, more productive, more human? Or will they make us quit?
As we discussed what comes next, Hassabis turned the conversation from Go to chess -- specifically, centaur chess.
After IBM's Deep Blue beat world champion Garry Kasparov in 1997, chess began experimenting with a format that allowed human players to consult computer engines. The result was a new breed of centaur: half-man, half-machine. For a time, the best human players with AI were better than AI alone. Hassabis believes we are now entering that centaur era of science.
"I don't know how long that period will last," he told me. "But for very complex domains, it could be a very long time. Like drug discovery, biology, chemistry -- they're very messy, very emergent and you can't verify everything. You need the human intuition and the human vision of which direction to go."
For now, AI will keep coming up with Move 37s.
The rest of us will have to find our Move 78.
Comments