In March 2016, a computer made a move so strange that the human experts watching assumed it was a bug. The machine was AlphaGo, the game was Go, and the move has its own name now: move 37. A commentator said out loud that it had to be a mistake. The world’s top players could not place it. And then, a hundred or so moves later, it won the game. The story is a great one on its own, but it also quietly explains what people mean, and do not mean, when they say software is “intelligent.”
What actually happened on move 37
Go is a board game of such staggering complexity that there are more possible positions than atoms in the observable universe. For decades it was considered the one game computers could not crack, because you cannot brute-force it the way early machines brute-forced chess. AlphaGo, built by a team at DeepMind, played the legendary champion Lee Sedol in a five-game match. In game two it played a move on the right side of the board that no human would choose. The estimated chance a human player would have made it was about one in ten thousand. It looked like a blunder. It was, in hindsight, a move of real beauty that set up the entire rest of the game.
If you want the full account, the AlphaGo versus Lee Sedol match is one of the most documented moments in the history of artificial intelligence, and worth the read.
Why it looked like a glitch
Here is the part that matters. The move looked wrong because it did not match how humans had played Go for thousands of years. The machine was not following a textbook, because it had partly taught itself by playing millions of games against versions of itself. It had no human teacher telling it which moves were “proper.” So it found a move that worked but that no tradition had ever blessed. To the experts steeped in that tradition, anything outside it reads as an error. The glitch was not in the software. The glitch was in our assumption that the human way was the only good way.
What this means for the software you actually use
Most software is not AlphaGo, and that is worth saying plainly. The vast majority of apps, including the smart ones, are not teaching themselves anything. They follow rules a human wrote, very fast and very reliably. When people get nervous about “AI” in their product, they are often picturing move 37, a system doing things nobody can predict. That is almost never what a business app needs or should do. You do not want your checkout flow inventing creative new moves. You want it boring and correct, every single time.
The useful kind of AI in a normal product is narrow and pointed. It answers one question a human would otherwise answer slowly: which shade matches this photo, which support reply fits this message, which products go with this one. That is a feature, not a mind. Knowing the difference is most of deciding what is actually worth building, and it keeps you from paying for a moonshot when a simple tool would do.
The honest takeaway
Move 37 is thrilling because it hints at machines that surprise even their makers. But the lesson for anyone building a real product is calmer than the headline. Add intelligence where it removes a genuine bottleneck, and nowhere else. A pointed AI feature that answers one question well will do more for your customers than a vague promise of a system that “learns.” If you are weighing where a smart feature earns its place, it helps to be ruthless about which AI features are worth the cost before you build a single one.
There is one more quiet lesson hiding in move 37. AlphaGo was not a general genius. It could not drive a car, write an email, or play checkers. It was extraordinary at exactly one thing because that is all it was ever built to do. The systems that feel magical are almost always the narrowest ones, aimed with total focus at a single hard problem. When a product tries to be clever about everything at once, it usually ends up clever about nothing. Narrowness is not the compromise. It is the secret.
So yes, a computer once made a move so good it looked broken, and it changed how the world thinks about what machines can do. Just remember that the magic in that story came from a system built for exactly one purpose, pushed as far as it could go. The same principle holds for the far humbler software most of us need. Pick one job. Do it better than a human could. That is the whole game.
Prefer to watch?
We turned this into a quick Short. Here it is, or watch it on YouTube.
Leave a Reply