AlphaGo Moment

An AlphaGo moment is a turning point: a machine suddenly solves a task better than the strongest human. The name comes from a computer program called AlphaGo, which in 2016 played against the South Korean Lee Sedol. The game was Go, a very old Asian board game with simple rules and an unfathomable number of possible moves. Experts at the time had estimated that machines would need at least another ten years for this. AlphaGo won nonetheless, 4:1. Since then, the term has been used for any case in which a technology matures faster than expected.

Why Lee Sedol’s defeat was a turning point

Go had long been considered the last bastion of human strategic thinking. Computers had already cracked chess back in 1997, when IBM's Deep Blue system defeated world champion Garry Kasparov. In chess, raw computing power helped: the computer checked millions of moves per second. That doesn’t work with Go, because there are more possible game sequences than atoms in the visible universe. It is therefore impossible to calculate everything. Good Go players instead speak of feeling and shapes on the board.

This is exactly why the victory was a shock. It showed that machines can also succeed where humans saw their strength as lying in intuition. Move 37 in the second game became famous. AlphaGo placed a stone in a spot that all commentators considered a mistake. Only many moves later did it become clear that the move had decided the game. The machine had not merely calculated faster – it had discovered something new.

For business and politics, the impact was even greater than for the world of gaming. China soon after launched a national program for artificial intelligence. Investors poured billions into companies working with similar methods. A single board game thus changed how seriously the world took the topic.

The pattern behind the leap

At its core, AlphaGo consisted of two parts. One estimated which move was likely good. The other estimated how favorable the overall position was. Both parts were neural networks, i.e. programs that learn patterns from many examples instead of following fixed rules. In addition, the program mentally tried out possible continuations, but only the most promising ones.

At first, AlphaGo learned from games played by human professionals. After that, it played against itself millions of times and improved through its own mistakes. This principle is called reinforcement learning: the program is not given a ready-made model to imitate, only feedback on whether the outcome was good. The successor version, AlphaGo Zero, ultimately needed no human games at all. It learned solely from the rules and clearly beat the older version.

This gives rise to the typical pattern of an AlphaGo moment. A capability grows slowly for a long time, seems harmless, and then in a short span becomes better than any human. It’s important to draw a distinction here: not every record is such a moment. It requires a task that was considered typically human, and a lead that can no longer be caught up.

Where the comparison is drawn today

In business news, the term often comes up when a company presents a surprisingly strong product. The launch of ChatGPT at the end of 2022 was frequently described this way, because text programs suddenly seemed ready for everyday use. The program AlphaFold is also considered an example: it predicts the three-dimensional shape of proteins, a problem that biology had worked on for decades.

The comparison is especially useful as a warning against false confidence. Those who claim that a task will remain too difficult for machines for a long time to come have often been wrong in the past. This is discussed, for example, with regard to autonomous driving, medical diagnostics, or programming.

At the same time, the term is used excessively. Companies like to use it in marketing because it sounds like a historic breakthrough. A sober look is therefore worthwhile: was a leaderboard only narrowly improved, or was an entire task actually solved? Only the second case truly deserves the name.