Nick Thorp

HOW AN AI's FORBIDDEN MOVE TRANSFORMED A 2,500-YEAR-OLD GAME OVERNIGHT March 2016. Seoul. Lee Sedol sits down at the board. Eighteen international titles. One of the great fighting players Go has ever produced. Before the match, he's asked how many games he expects to lose to this Google machine. His answer, essentially: one. At most. Nobody thought that was arrogance. Go was supposed to be untouchable. More legal positions than atoms in the observable universe. Too intuitive, too pattern-driven, too human for silicon. Chess fell in 1997, sure. But Go is different, everyone said. Give it another decade. Months earlier, this same AI had quietly beaten the reigning European champion 5-0, and most of the world shrugged. Fine for Europe. Seoul was a different weight class. Game 1, AlphaGo wins. Uncomfortable, but survivable. Great players lose games. Game 2 begins, and for a while it looks like ordinary Go. Then move 37. AlphaGo slides a stone onto the fifth line, a loose shoulder hit out near the corner. It does nothing immediately. It looks wrong the way a misplaced word looks wrong in a sentence you've read a thousand times. Quick primer, because the wrongness only makes sense with the context. For most of its 2,500 years, Go has been one long argument: territory versus influence. Take small, countable, safe points now — or play looser, higher moves that promise more later if the position breaks your way. Two and a half millennia of play distilled into a rough consensus: the third and fourth lines are the efficient real estate. The fifth line, that early, in a match of this level? Somewhere between a joke and an insult. The professional commentating the broadcast does a double take. Viewers flood the chat: mistake, bug, glitch. Lee studies the board a long, long time. Then he leaves the room. Fan Hui — the European champion who had served as AlphaGo's training partner and lost to it 5-0 — watched the move come down from the commentary room. His reaction got written down, and it's the entire story in three sentences: "It's not a human move. I've never seen a human play this move. So beautiful." Then came the number that broke everyone's brain. DeepMind said AlphaGo estimated the probability a human would ever play move 37 at about 1 in 10,000. Not brilliant. Not creative. Practically a typo. And it played it anyway, because under the move was something cold and simple: it had calculated the win probability, and the fifth-line shoulder hit was the best option it could find. It wasn't trying to look smart. It was trying to win. It just wasn't using human habit as the definition of what's allowed. Lee resigned Game 2. AlphaGo took the match 4-1, in front of roughly 200 million viewers. The one game humans did win — Game 4, Lee's famous "divine move" — only sharpened the point. It took one of the greatest players alive producing one of the greatest games of his career to pin this thing once. The aftermath is the part I think about most. Within months, fifth-line invasions started showing up in professional title matches. Opening theory got re-sorted almost overnight — centuries of "this move is not played" quietly reclassified as "this move was never explored." Students of the game stopped asking "is this how a human plays?" and started asking "does it win?" Territory versus influence didn't get a final verdict. It got a new frontier. A 2,500-year-old game grew a fresh branch in a week. And remember what the argument was before March 2016: machines can calculate, but creativity is human. AlphaGo was dismissed as brute force with better pruning. Move 37 is where that argument died in public. Creativity turned out to have no mystical sparkle in it at all. It's what happens when a search runs deeper than habit. Here's why this eats at me as someone who builds AI systems for a living. AlphaGo's first versions were trained to imitate humans — millions of expert games, predict the move the master would pick. Imitation made it strong. That's what got it off the ground. But the move that made history came after its makers let it play itself. Millions of games with nobody left to copy. Exploration past the point where human habit runs out. Now look at what most of us build with today. Large language models are imitation engines. They've read everything we've written — which means they've inherited our ruts, our fourth-line safety, our 1-in-10,000 moves we never even consider. A system trained only to imitate people can never, by construction, play move 37. Its ceiling is the average of whatever it copied. Which is why my first instinct when a model does something off-script — must be a bug, must be a bad prompt, shove it back toward the average — deserves a second look. Sometimes it is a bug. Sometimes it's the fifth line. "This isn't how a human would do it" and "this is wrong" are two different claims. One is a measurement. The other is a reflex. The commentators in Seoul had about an hour of confidence before the reflex cost them the game. So when your model hands you an answer that looks wrong, test it before you correct it. You might be holding a move 37. What about you — when did a model hand you a "mistake" that turned out better than what you'd have written yourself? Tell me in the comments. I read every one. And if you're building with AI, follow along. I write the weird moves down so you don't have to find them the hard way.

RIP Vibe coders, learn marketing 🤣😂🤣

Fable 5 already carried the same “new standard” framing in June and still lost on price-to-performance against open coding models on many real repo tasks. A 5.5 release does not erase the 5–10× effective cost advantage of local Qwen/DeepSeek-class models for the same SWE-bench-style work.

Will this one work as intended?

People talking about the X algo drop like it's inscirbed on tablets dropped from the heavens from the most high. Spoiler Alert: Small accounts have to reply to big accounts for 365 consecutive days in order to grow.

Try not to get too hyped about Jack Dorseys AI company after "revealing" Dessa. Unless I get hands on, I'm not buying the hype.

Interesting, but a few sample recordings and no way to test it myself is a red flag.