Scale isn't the barrier you think
A search space larger than the universe sounds hopeless. It wasn't. The lesson isn't "compute harder" — it's that the right kind of learned judgment can navigate spaces no one could ever fully enumerate.
Go is one of humanity's oldest games and, for decades, artificial intelligence's most stubborn problem. The board is simple. The number of ways to fill it is not — it dwarfs the count of every atom in the observable universe. This is the story of why that made Go "unsolvable," and how a machine finally cracked it — not with raw calculation, but with something that looked a lot like intuition. Every number and quote here is real; sources are listed at the bottom.
Go was invented in China more than 2,500 years ago and is widely considered the oldest board game still played in its original form. Its first written mention appears in the Zuo Zhuan, a Chinese text from around 548 BCE. In imperial China it became one of the "four arts" every cultivated scholar was expected to master — alongside calligraphy, painting, and music.
The rules fit on a postcard. Players take turns placing black and white stones on the intersections of a 19×19 grid. Surround empty territory and capture enemy stones; whoever controls more of the board wins. That's essentially it.
Yet mastery can take a lifetime. Unlike chess, Go rewards a kind of whole-board feel — shape, influence, and balance that top players describe as intuition rather than calculation. Children who devote themselves to it spend years studying professional games the way musicians study scores.
Try it: this is a 9×9 practice board (beginners often start here before graduating to the full 19×19). Place a few stones to get a feel for the grid.
Here is where the simplicity ends. A full Go board has 361 intersections, and each can be empty, black, or white. The number of legal board positions was computed exactly in 2016 by John Tromp and collaborators: it is 208,168,199,381,979,984,699,478,633,344,862,770,286,522,453,884,530,548,425,639,456,820,927,419,612,738,015,378,525,648,451,698,519,643,907,259,916,015,628,128,546,089,888,314,427,129,715,319,317,557,736,620,397,247,064,840,935 — about 2.08 × 10170. Drag the board size below and watch that number sprint past everything else in existence.
To be clear about the comparison: scientists estimate the observable universe contains around 1080 atoms. The number of legal Go positions is 10170. That is not "a lot more" — it is the number of atoms in the universe, multiplied by itself, and then multiplied by ten billion more. And the number of possible games (every sequence of moves, not just resulting positions) is vastly larger still, with estimates running past 10700. You could give every atom in the universe its own universe of atoms and still not have enough labels to tag every Go position.
In 1997, IBM's Deep Blue beat world chess champion Garry Kasparov mostly by calculating — searching roughly 200 million positions per second and looking many moves ahead. That works for chess because the tree of possibilities, while huge, is narrow enough to prune. Go breaks it. On each turn a Go player has around 250 reasonable moves to consider; a chess player has around 35. Multiply that out over just a few moves and the search space explodes.
Even a computer checking a trillion positions every second would need far longer than the age of the universe to brute-force a single Go game to the end. Raw search alone was never going to be enough. Something had to judge a position the way a human does — at a glance, without reading every branch. That something turned out to be a neural network.
DeepMind's AlphaGo paired tree search with two neural networks — one to suggest promising moves, one to estimate who was winning — trained on human games and then on millions of games it played against itself. Instead of reading every branch, it developed a learned sense of which positions were good. In under two years it went from theory to beating the best humans alive.
AlphaGo placed a stone on the fifth line — a spot human masters almost never choose. Its own model estimated the odds a human would play it at about one in ten thousand. Commentators first assumed it was a mistake. It wasn't. The move reshaped the whole board and won the game, overturning centuries of received wisdom. Fan Hui, watching, called it simply "beautiful."
Down three games to none, Lee Sedol found a wedge between AlphaGo's stones so precise it was later nicknamed the "God's touch." It was, by the same measure, a one-in-ten- thousand move. AlphaGo's evaluation of the game collapsed; it spun into a cascade of weak replies, and Lee won. It remains the only game any human ever won against competition-strength AlphaGo.
The lasting shock of the match was not brute strength — it was that both sides produced moves that looked like inspiration. Move 37 showed a machine finding beauty no human had; Move 78 showed a human, at the absolute edge of the possible, finding what the machine had missed. Two one-in-ten-thousand moves, one from each kind of mind.
"It's not a human move. I've never seen a human play this move. So beautiful."
"It seems like a wall… I know AlphaGo is a computer, but if no one told me, maybe I would think the player was a real person — a very strong player."
"Even if I become the number one, there is an entity that cannot be defeated."
Three years after the match, Lee Sedol retired from professional play. He was the only human ever to beat AlphaGo, and he walked away saying the thing he had spent his life mastering now had, at its summit, an opponent that could not be beaten. It was a somber coda — but it also marked the moment a problem long dismissed as "impossible for machines" had, quietly and decisively, been solved.
Go is a vivid, self-contained lesson in how modern AI actually works — and why it matters far beyond a game board.
A search space larger than the universe sounds hopeless. It wasn't. The lesson isn't "compute harder" — it's that the right kind of learned judgment can navigate spaces no one could ever fully enumerate.
AlphaGo wasn't told the rules of good play. It learned a "feel" for strong positions from examples and self-play — the same broad recipe behind today's language and image models.
Many believed a computer beating a Go professional was ten years off — right up until it happened. Timelines for what AI can and can't do have a habit of collapsing.
The best moment wasn't the machine's win — it was Move 78, a human rising to meet it. The most interesting outcomes come from people and AI pushing one another, not one replacing the other.
The hard part of AI in a business is rarely the model. It's knowing which of your "impossible" problems are actually a Go board — enormous, but tractable with the right approach — and which are not. That's the conversation we have with leadership teams before they commit budget.