The Creativity Code

BGJ 191 Spring 2020,
Reviewed by Colin Maclennan

How AI is Learning to Write, Paint and Think
by Marcus du Sautoy

Marcus du Sautoy’s latest book The Creativity Code will be of interest to many Go players. The book is an exploration of the developing world of artificial intelligence (AI). Du Sautoy probes the question raised by Ada Lovelace, the 19th century mathematician who worked with Charles Babbage on his ‘Analytical Engine’. She was the first to recognise that the machine had applications beyond mere number crunching and proposed what is thought to be the first computer program. However she did not think it could ever be possible for a computer to go beyond what had been entered by the coder. Any creativity must lie with the coder, she thought. It is this question that du Sautoy probes in this book, and he suggests a Lovelace Test.

After a discussion of creativity, du Sautoy starts with a blow-by-blow account of Demis Hassibis and the development of DeepMind’s AlphaGo to the point where it beat Lee Sedol, South Korea’s eighteentime world champion. I found du Sautoy’s explanation illuminating that in hundreds of years of playing Go, human players have built up conventions that lead to what mathematicians would call a ‘local maximum’. What AlphaGo has done by starting from scratch and learning to improve its play is to find another maximum which is higher.

The book moves on from Go to describe the development of the algorithms and their use in selecting recommendations for music, films, books etc. on the basis of users’ previous choices, and explains how they modify their results as more data accumulates. Later chapters bring us to the application of these techniques to creating new works in the style of famous composers and artists.

Du Sautoy is an amateur musician and he focuses in some detail on development of AI algorithms to compose music. He refers to DeepBach, which generates Bach chorales that music students were unable to tell apart from the real thing.

Du Sautoy does not reach a firm conclusion on Lovelace’s question, though he thinks the Lovelace Test has been passed in the case of music, but he certainly does enough to suggest to me that the question is very much an open one. Meanwhile AlphaGo has now been retired from competitive play and DeepMind are now looking to other goals: healthcare, climate change, energy efficiency, face and speech recognition. It’s all getting serious!

So where does AlphaGo leave Go played by humans? Will it will be possible to build upon the lessons learned from AlphaGo to develop new conventions and styles of play, such that a human player will eventually be able to beat AlphaGo and claim the higher peak that AlphaGo has discovered?