Hassabis on chess and AI

There are two things to worry about. One is bad actors, whether that's individuals or all the way up to nation states, repurposing these technologies that we're trying to build for good, like curing diseases and advancing material science and energy for harmful ends?

Fritz 21

YOUR PERSONAL CHESS COACH - Whether you’re taking your first steps into the world of club chess, or already playing at a tournament level: with FRITZ, you can train more efficiently, intelligently and with a more personalised approach than ever before.
FRITZ is more than just a chess engine – it’s a training revolution! Whether you’re taking your first steps into the world of club chess, or already playing at a tournament level: with FRITZ, you can train more efficiently, intelligently and with a more personalised approach than ever before.

The above 15-minute video tells us how one of the central scientists working on AI evaluates the current situation. It is definitely worth listening to. But it is the vivid description of how AI handled chess will be of particular interest to our news page readers.

Hassabis describes how early artificial intelligence programs – specifically "expert systems" like IBM's Deep Blue managed to defeat world champions at complex games like chess. It was done using human-coded knowledge. Instead of the AI learning on its own, a team of smart programmers collaborated with domain experts (chess grandmasters), extracting the experts' strategies and manually coding them into the system as a strict set of rules and heuristics. The resulting systems did not actually "think" or understand the game. They used massive computing power to mechanically apply the human-coded rules to calculate which move had the best mathematical outcome.

Deep Blue was not proper AI, Hassabis says. It achieved world champion in chess, but it couldn't do anything else. It couldn't even play simpler game like tic-tac-toe. Can you imagine a human grandmaster not being able to learn how to play tic-tac-toe?

So where did the intelligence of Deep Blue reside? It wasn't in the system, it was in the minds of the grandmasters and the programmers who solved the problem and then implemented the solution. The program just dumbly executed the solution.

Hassabis and his team took a different path: a learning system, a neural network, Alpha Zero, that starts off with only the rules of a game. It played randomly, 100,000 games against itself. They let the program create its own data set – it could see which moves were slightly better than others. The 100,000 games were used to train a new version that became slightly better than version one. Similarly version two was used to train version three, which trained version four, and so on. It turned out that around 16 or 17 generations were enough to go from random chess to better than world champion.

Demis was a world-class child chess prodigy who reached an Elo level of 2300 as a 13-year-old, before stepping back from competitive play to pursue computing, video game design, and artificial intelligence. He was fascinated to watch the neural network learn chess live:

"It starts in the morning, then by lunchtime I could still just about compete with it. By tea time it's better than all grandmasters, and then by dinner time it's better than the world champion. And it's playing interesting new chess, new types of moves. I think we need these types of ideas with our foundation models, the new Gemini and these kinds of things. You can think of them as generalized models of everything, language, the world around us, not just a game.

If you are interested in the subject, and concerned about what lies ahead for humanity, you would do well to watch the rest of the interview. It is with a consummate expert on the subject, one who is very clear in conveying his knowledge and thinking.

Demis Hassabis

Sir Demis Hassabis (born 27 July 1976) is a British computer scientist, artificial intelligence researcher, neuroscientist, and entrepreneur who co-founded DeepMind in 2010. He has served as CEO of Google DeepMind following its acquisition by Google in 2014. Today he is the Chief Scientist of Alphabet Inc. (Google's parent company), and also continues as the CEO of Isomorphic Labs.

Hassabis's work on AlphaFold, an AI model that predicts protein structures with unprecedented accuracy, addressed a longstanding challenge in biology known as the protein folding problem. For this contribution, he shared the 2024 Nobel Prize in Chemistry with John Jumper and David Baker.

Also read

The adventure of chess programming (3) – originally published in Der Spiegel.

ChessBase '26 & Fritz 21

EXPAND YOUR CHESS HORIZONS
ChessBase 26 + Fritz 21: The Summer Special Power Bundle
Study like a champion. Play like a legend.
Real progress happens in two places: at the desk, where you prepare and analyze, and at the board, where you test yourself under fire. For one week only, we are bringing both ChessBase flagships together in a single bundle so you can close the gap between the two. Prepare with ChessBase 26. Play against Fritz 21. Improve faster than either program can take you on its own.