AI Learned Chess in Four Hours. Then It Started Playing Like an Alien.

AlphaZero wasn’t given chess books, famous games or centuries of human strategy. It was given the rules and told to play itself.

Chess has had centuries to figure itself out. Players have written books about openings, studied famous games and passed strategy from one generation to the next. Grandmasters can recognize positions they have seen hundreds of times before.

Then DeepMind gave an AI called AlphaZero a chessboard and almost none of that knowledge. It wasn’t shown grandmaster games, given an opening book or taught which moves humans considered smart or stupid. It got the rules and started playing itself.

After about four hours of self-play, AlphaZero had already become strong enough to beat Stockfish, one of the best chess programs in the world. That was impressive, but the stranger part was how it played.

It Learned Chess From Scratch

At first, AlphaZero played terrible chess because it knew almost nothing. Then, by playing millions of games against itself, it started rediscovering ideas humans had spent centuries developing: protect your king, control space, build strong pawn positions and don’t waste moves.

But it didn’t stop there. AlphaZero sometimes made moves that looked wrong to human players. It was willing to give up pieces for advantages that might not become clear until much later. It cared less about how many pieces it had and more about whether the position was slowly becoming impossible for the other side.

Sometimes it even gave up its queen, which can look insane to a human player because the queen is the strongest piece on the board. But AlphaZero wasn’t thinking about the queen the way people do. It was asking a simpler question: does this help me win?
That difference mattered.

Then Go Got Its Own Alien Move

A few years earlier, DeepMind had already shocked another group of experts. Its AI AlphaGo was playing one of the world’s best Go players, Lee Sedol, when it made a move that became famous almost immediately: Move 37.

Professional players watching the match struggled to understand it at first. It landed in a place where experienced players normally would not play that early, and some observers thought the machine had made a mistake.

It hadn’t. The move helped AlphaGo build a winning position.
That moment mattered because the AI wasn’t simply copying old human strategy faster than humans could. It had found a good move that human players had mostly ignored.

Then They Took Away the Game Board

The obvious question was whether this only worked in games. Chess and Go are perfect for AI because the rules are clear and the machine always knows whether it won or lost.

So researchers started turning other problems into games.
In 2022, DeepMind built a system called AlphaTensor and pointed it at a math problem used constantly by computers.

The exact mathematics is complicated, but the basic challenge was simple: can you do the same calculation using fewer steps?

Mathematicians had been working on versions of this problem for decades. AlphaTensor found new shortcuts, and for some calculations it discovered methods that beat the best known human ones.

The important part wasn’t the math itself. It was the pattern beginning to appear: give the AI a clear goal, let it experiment enough, and sometimes it finds a route humans missed.

Then AI Started Rewriting Computer Code

A year later, DeepMind tried the same basic idea on something more practical: sorting. Computers constantly sort information, including numbers, files, search results and database records. Programmers had already spent decades making these routines extremely fast, so there shouldn’t have been much easy improvement left.

DeepMind built an AI called AlphaDev and told it to look anyway.
Instead of writing code the way a human programmer normally would, AlphaDev experimented with tiny low-level computer instructions. Some of the patterns looked strange, but they worked.

For small sorting jobs, some versions were dramatically faster.
Then the really important part happened. Some of AlphaDev’s discoveries were added to a widely used C++ software library.
The strange move had left the chessboard and entered real software.

Then It Started Finding New Math

In 2023, DeepMind introduced another system called FunSearch. The idea was surprisingly simple: an AI suggests a solution, a computer checks whether it works, bad ideas get thrown away and better ones survive long enough to be changed again.

Researchers pointed FunSearch at a difficult math problem people had studied for years, and it found better answers. They also tested it on a practical problem called bin packing, which is basically the challenge of fitting things into limited space as efficiently as possible.

Think shipping boxes, computer storage or scheduling. Again, the system found new strategies. By then, the pattern was becoming hard to miss. AI had found a strange Go move, strange chess strategies, better math shortcuts, faster computer code and new answers to problems researchers already knew very well.

The Weird Part Isn’t That AI Can Beat Us

Computers beating humans at chess isn’t the interesting part anymore. We knew machines would eventually calculate faster than people ever could.

What made AlphaZero different was that it started with almost none of our strategy, rediscovered much of it by itself and then sometimes kept going into places humans hadn’t explored.
That doesn’t make the machine magical. Humans still build the system, choose the problem and decide what counts as a good answer.

But once the search begins, the AI doesn’t care whether an idea looks weird, ugly or completely wrong according to tradition.
It doesn’t care whether a grandmaster says nobody plays that move or whether programmers have written code the same way for twenty years. It only cares whether the idea works.

Human Knowledge Can Also Box Us In

Human knowledge is powerful because we don’t have to start over every generation. A chess player doesn’t need to rediscover basic strategy, a programmer doesn’t need to reinvent sorting and a mathematician can begin where someone else stopped. That’s how civilization moves forward.

But there is a downside. Once everyone agrees that something is normally done a certain way, people naturally spend less time exploring the stranger alternatives.

AI doesn’t have that problem in quite the same way. It hasn’t spent twenty years being told that a certain move looks stupid, so when it searches, it sometimes wanders into places humans stopped checking. Most of those places are useless. Occasionally, one isn’t.

The Next Alien Move Probably Won’t Be in Chess

Move 37 looked wrong until it worked. Some of AlphaZero’s chess looked wrong until grandmasters studied it. AlphaTensor found math shortcuts humans hadn’t found, AlphaDev discovered faster code and FunSearch pushed known problems a little further.

None of these systems became some all-knowing scientist. Something simpler happened: we built machines that can test enormous numbers of ideas without caring whether those ideas look normal to us.

That may turn out to matter far more than beating humans at chess, because somewhere there may be a better battery, a better material, a better algorithm or a better medical treatment that looks completely wrong according to what humans currently believe.

For most of history, strange ideas often died when the expert in the room said, “That’s not how it’s done.”

AI doesn’t know that’s not how it’s done.
Sometimes that’s exactly why it finds the move.

After only 4 hours of playing chess AI basically said “Hold my beer” then pulled out a 37.