Google has accomplished what many artificial intelligence researchers had thought was still years or even decades away, as AlphaGo, an AI developed by Google’s DeepMind project, has successfully defeated legendary Go player Lee Sedol.

“I’m kind of speechless,” DeepMind founder Demis Hassabis said in a post-game press conference. “It’s kind of the most mind-blowing kind of game experience we’ve had so far.”

AlphaGo won four of its five games against Sedol, showing that while the AI may be considered one of the best Go players in the world, it is not unbeatable. Not yet anyway. Because AlphaGo uses machine learning to improve its skills as it gains more data, the AI will only grow stronger over time.

When Google published AlphaGo’s first success against European Go champion Fan Hui, top-ranked player Myungwan Kim analyzed the AI’s playstyle and concluded that at its skill level at the time, it would not be capable of defeating Sedol.

“No offense [to Google], but it’s not going to happen,” Kim said at the time.

However, thanks to its rapid learning ability, AlphaGo was able to strengthen its playstyle in the short time between its competition with Fan Hui and its competition with Sedol. One can only imagine how much stronger it will become over time, and future Go players will likely continue trying to defeat the AI.

What this means for AI

Defeating a top-ranked Go player has been on the AI researcher bucket list for some time now, but Google’s accomplishment represents more than simply building a smart board game-playing program.

The real breakthrough was the methodology used to create and train AlphaGo, which was never actually told how to play Go. Instead, AlphaGo was fed raw data from real Go matches and had to use machine learning to determine the rules of the game and how to be successful at it.

AlphaGo played the game against itself over and over to determine optimal moves in millions of different situations until it eventually rivalled the best players in the world.

Rather than programming an AI and telling it exactly what to do and how to respond to certain stimuli, the methodology behind AlphaGo effectively let the AI program itself, which is a small step on the long road to the technological singularity, where AIs become capable of building better version of themselves without requiring human intervention.

Screenshot via DeepMind | YouTube