Alan Turing saw it coming. 鈥淢achines take me by surprise with great frequency,鈥� he wrote in 鈥� the same 1950 paper that begins 鈥淚 propose to ask the question 鈥楥an machines think?鈥欌�� and goes on to set down the rules of an Imitation Game, or what we call the Turing Test, designed to get the measure of an artificial intelligence.
Turing was pushing back against the idea that computers could do nothing more than what they were instructed to do. Despite his arguments, it is a reputation they were stuck with for decades. Machines are robotic. They cannot come up with anything new, they cannot be creative.
Advertisement
But fast-forward 77 years and our machines are full of surprises. What鈥檚 more, the ability to lead us to unexpected places 鈥� from new kinds of art to new discoveries in science and medicine 鈥� is arguably the biggest breakthrough artificial intelligence could make.
2017 has been a good year for computers with a creative bent. September鈥檚 , an annual online event in which coders, artists and game developers get together to make things that make things, had more entries than ever. Participants use an AI technique called procedural generation to produce new games and animations automatically. The results are often delightful and unexpected.
Surprise is important, says Michael Cook of the at Falmouth University, UK, who is a co-organiser of ProcJam. 鈥淏eing surprising means you鈥檝e shattered people鈥檚 expectations and for software that is usually a huge achievement.鈥�
Two of the most striking entries were , a series of eerie alien worlds built from 3D fractals that you can wander around, and , which generates an endless run of weirdly lifelike animated figures using .
Surprisingly unhuman
Working with creative software can inject uniqueness into whatever you are making, says writer Jupiter Hadley, another ProcJam organiser. 鈥淧rocedural generation adds an unpredictable element, a surprising thing that humans wouldn鈥檛 do.鈥�
That鈥檚 also true of some of the entries to November鈥檚 National Novel Generation Month 鈥� known as 鈥� which celebrates computer-generated fiction. For example, Janelle Shane 鈥� an optics researcher at Boulder Nonlinear Systems in California, who 鈥� trained a machine learning system to generate the first line of a story by feeding it 11,135 opening lines from existing novels.
Many of the results are nonsensical or just plain odd. But a few hit you like good poetry, fresh and evocative: 鈥淭he silence was unlike a place鈥� or 鈥淭he sky was at the door鈥�.
Computers have even had a stab at creating entirely new painting styles. This was done using generative adversarial networks 鈥� or GANs 鈥� which were everywhere this year. GANs work by playing two neural networks off one another, batting works-in-progress back and forth until one judges the others鈥� efforts a success.
But the importance of surprise goes far beyond quirky works of art or catchy first lines. When DeepMind鈥檚 AlphaGo software beat South Korean Go master Lee Sedol last year, Sedol was shocked by the AI鈥檚 style of play. Trained on many thousands of past Go games, AlphaGo went on to outstrip its mentors and made winning moves that would have been unthinkable to a human player.
In October, DeepMind revealed AlphaGo鈥檚 successor 鈥� AlphaGo Zero 鈥� a system that ignored several millennia of human experience and learned how to play Go by itself. In three days, not only had it rediscovered the best strategies humans had devised, but it had also created new ones. 鈥淭he union of human and computer players will usher in a new era,鈥� said Chinese Go champion Ke Jie, who lost to AlphaGo earlier this year. 鈥淭ogether mankind and AI can finally find the truth of Go.鈥�
AI leg up
AlphaGo鈥檚 creativity is already making human players better: European champion and early sparring partner Fan Hui jumped up the world rankings after playing the AI. Deep Blue didn鈥檛 do the same for chess players. Today鈥檚 grand masters are about as good as Kasparov was in 1997. This is because chess programs beat humans simply by being far quicker at calculating. AlphaGo actually has new ideas about the game, which humans can learn.
DeepMind co-founder Demis Hassabis wants software like AlphaGo to help us invent new drugs or open doors to unknown corners of science. He spent much of the year talking about creativity 鈥� including at in September. For Hassabis, creativity is about conjuring counterfactual situations and using them to think through solutions that might never be needed. It means being one step ahead rather than waiting for the world to hit you on the head.
We have long thought of computers as reasoning machines, solving problems with methodical steps. But problems are actually solved with brilliant flashes of inspiration 鈥� ideas that pop up from nowhere. Mathematicians and artists alike know this feeling. And it is where our best AI is headed.
鈥淢aking a computer make something is still one of the most magical things I do,鈥� says Cook. 鈥淚 always cherish the first time software I鈥檝e made surprises me.鈥�
