In March 2016, DeepMind's AlphaGo program defeated Lee Sedol – one of the world's strongest Go players – with a score of 4:1. This wasn't just a tournament result. It was the moment conversations about artificial intelligence stopped being about the future.
For a long time, Go was considered a game beyond the reach of machines. Not because the rules are complex – they're actually quite simple. But because the number of possible moves on the board is so vast that no computer could possibly brute-force them all. It was believed that intuition was required. As it turned out, it wasn't. More accurately, it turned out that intuition could be replicated by learning from a sufficiently large number of examples and playing against oneself millions of times.
Move 37: The Move No One Saw Coming
In the second game, AlphaGo made a move that professional commentators considered a mistake. Lee Sedol left the table – he needed time to process what he had just seen. It was later revealed that this move was brilliant. The probability of a human playing the same move was estimated at one in ten thousand.
This moment became symbolic of something more than just a victory in a game. The machine didn't just replicate the human style of play – it found an approach that humans hadn't conceived of in the millennia that Go has existed. This was a signal: AI is capable not only of learning from humans but also of going beyond what humans already know.
From the Game Board to the Real World
After AlphaGo, the DeepMind team began applying the same approach to problems that seemed far more complex than any board game.
The first high-profile result was AlphaFold. Go has clear rules and a clear criterion for victory. Biology works differently. But the principle turned out to be similar: give the system enough data and the ability to figure out solutions on its own, and it will tackle a problem that scientists have been struggling with for decades.
AlphaFold solved the problem of protein structure prediction. Simply put: every protein in our body is a long chain of molecules that folds in a specific way. How the protein functions depends on exactly how it folds. Figuring this out manually took years of research for a single protein. AlphaFold learned to do it in minutes. Today, its database covers over 200 million protein structures – essentially the entire protein universe known to science.
This fundamentally changed biology. Researchers worldwide gained a tool that has accelerated work in drug development, disease research, and understanding living organisms at the molecular level.
It Didn't Stop with Biology
AlphaFold is the most famous example, but not the only one. In the years since, the approach pioneered by AlphaGo has found applications in a wide variety of fields.
- Mathematics. AlphaProof and AlphaGeometry have shown that AI can not just calculate, but prove theorems – and do so at a level comparable to contestants in the International Mathematical Olympiad.
- Nuclear Fusion Plasma Control. AlphaGo gave rise to reinforcement learning approaches that are applied to control tokamaks – devices for plasma confinement in nuclear fusion. The challenge there is similar: a vast space of possible states, no ready-made solution, and the need to find a strategy through trial and error.
- Computational Optimization. AlphaDev discovered new sorting algorithms that are more efficient than those humans had used for decades. These algorithms are already integrated into widely used software libraries.
- Weather and Climate. GraphCast, developed using similar ideas, makes weather forecasts more accurately and faster than traditional methods.
A common thread runs through all these cases: instead of hard-coding knowledge, you give the system the ability to learn – from data, from its own experience, and from feedback on its results.
Why AlphaGo Is More Than Just a Success Story
It's easy to tell this story as a string of triumphs. But it's more honest to say that AlphaGo also raised some uncomfortable questions.
One of them is about the limits of understanding. AlphaGo makes winning moves. But the system can't explain why a particular move is the right one. This works in Go. In medicine or law, it's not so easy to accept.
Another question is where the tool ends and something more begins. After AlphaGo, DeepMind clearly started thinking not just about specific tasks, but about a more general AI – systems that can transfer experience from one domain to another. This field is now known as AGI, or Artificial General Intelligence. AlphaGo was not AGI. But it showed that learning without rigidly defined rules could work – and work better than expected.
What Has Changed in Ten Years
In 2016, many experts predicted it would take AI another ten years to beat the top Go players. AlphaGo did it sooner. That in itself was a lesson: our intuitive estimates of progress in AI are systematically underestimated.
Ten years have passed since then, and AI has entered fields that in 2016 were spoken of only as distant prospects. Language models that can hold meaningful conversations. Systems that generate images and video. Tools that assist scientists in real-world research. All of this is part of a wave that AlphaGo helped set in motion.
Interestingly, after the match series against AlphaGo, Lee Sedol continued to play Go – and, in his own words, the game became different for him. He began to look at the board differently, noticing moves he had never considered before. The machine changed how a human thinks about the game to which he had dedicated his life.
In a way, this is the best description of what is happening to science as a whole. AI is not replacing researchers – it's changing how they look at problems. It opens up perspectives that humans wouldn't have thought of on their own, or would have arrived at much later.
What's Next: An Honest Answer
No one knows exactly where this is heading. DeepMind openly states that its ultimate goal is to create systems capable of solving a wide range of problems as flexibly as a human. But between today's systems and that goal lies a vast gap, the nature of which is not yet fully understood.
What is clear is this: AlphaGo was not an endpoint, but a starting point. It showed that AI can find solutions where human thinking hits the ceiling of its own experience. And that is, perhaps, more important than any score on a game board.
