07 January 2021
Protein-Folding and Other Games
Demis Hassabis, homo ludens.
By Neil Tidmarsh
Last month, science put on its Santa-suit and gave humanity a whole sledgeful of early Christmas presents: the Pfizer/BioNTech vaccine, the Oxford/Astrazeneca vaccine, the Moderna vaccine – and the secrets of protein-folding.
Understanding the mysteries of protein-folding – the three-dimensional shapes that proteins form – was considered for a long time to be one of the most difficult problems in science. Now, thanks to DeepMind’s breakthrough, the road appears to be open to the development of treatments for serious conditions such as Alzheimer’s and Parkinson’s. Scientists have been working on this problem for decades and they expected that they’d be working on it for a good few years to come before they finally cracked it. The reaction to DeepMind’s revelation was one of amazement and celebration; it has come much sooner than anyone expected and it’ll accelerate the race to find cures for certain illnesses to a previously unimagined degree.
So who is the genius behind this breakthrough? What is his or her background? A wise old white-bearded professor, perhaps, in some eminent university’s biology department? A white-coated scientist in some multi-national pharmaceutical giant’s gleaming laboratory? A machine-brained mathematician feeding data into his algorithms in some academic ivory-tower?
No. DeepMind, dedicated to machine-learning artificial intelligence, is a classic information technology start-up (although now owned by Google) based in Kings Cross, only a stone’s throw from Shoreditch and silicon valley – hipster-central with its bearded and tattooed denizens of small, funky, non-hierarchical and entrepreneurial outfits. The man behind the DeepMind organisation is Demis Hassabis, and his background is in gaming and games playing. He was a schoolboy chess champion, becoming a Master at 13 with an ELO rating of 2300 (the second highest in the world for his age); he was the World Team Champion of the strategic board game Diplomacy in 2004; he cashed at the World Series of Poker six times including in the Main Event; he was the World Decamentathlon (part of the mind sports Olympiad) champion twice; and he has been the World Pentamind board games champion a record five times.
As a games developer rather than a games player, he’s had an award-winning computer games career at companies such as Bullfrog, Lionhead Studios and his own Elixir Studios. He co-designed and programmed Theme Park (a seminal and best-selling management sim game which won a Golden Joystick Award) as a teenager working at Bullfrog while filling in time waiting to go up to Cambridge University (he left school at 16, having taken his A levels a couple of years early).
All this is rather disingenuous, of course. He does have an outstanding academic CV as well. He gained a double first in computer sciences at Cambridge; he has a PhD in cognitive neuroscience from UCL; he’s further researched neuroscience and artificial intelligence as a visiting scientist at Massachusetts Institute of Technology and at Harvard; and he’s published numerous influential papers and won many academic awards.
But the focus on his gaming background is more than valid; it emphasises the vital and often overlooked importance of games and play as the generator of innovation, learning and discovery. A surprising number of the great developments in the science of mathematics has been prompted by games players and gamblers; the theory of probability, for instance, was developed by the philosopher Blaise Pascal for a gambling mate of his. The algorithms developed by and for hedge-fund managers and automated trading systems and other punters in the stock exchange casinos of Wall Street and the City of London and other rivals to Las Vegas have fast-forwarded the development of the science this century. The only professional gambler I’ve ever met looked nothing like your typical Mississippi riverboat swashbuckler – no Stetson, no swaggering moustache, no boot-lace tie, no gold-and-diamond flashy Clark Gable grin; he looked and sounded just like a prep-school maths teacher, quietly feeding lots of juicy data into his algorithms – which he’d developed in the City – and placing bets as instructed by them after they’d fully digested the meal (although he did admit that he played poker and backgammon for money from time to time).
Demis Hassabis is a classic exemplar of homo ludens, ‘playing man’ or ‘man the games player’. This is central to his approach and to that of his organisation. He set up DeepMind in 2010 with the aim of developing and mastering artificial intelligence and then using that to crack the problems which humanity and the wider world are facing; its mission was and is to “solve intelligence” and then use intelligence to “solve everything else”. And he saw that teaching machines to play human mind games was fundamental to “solving intelligence”, ie to modelling the human brain and creating even more powerful versions of it.
Humans have been teaching machines to play games for some decades, of course. A machine beat a human being at a board game for the first time in 1997, when the backgammon-playing program BKG 9.8 defeated world champion Luigi Villa (funnily enough, BKG played badly, nowhere near as skilfully as Villa, but the machine was luckier with the dice!). Later that year, IBM’s Deep Blue computer beat world chess champion Gary Kasparov.
But DeepMind has taken it all to a new level. First of all, it has produced machines which are uniquely powerful and, yes, ‘game-changing’. Its first breakthrough came in 2013 with the algorithm Deep Q-Network (DQN) which played Atari games “at a superhuman level”. Three years later its program AlphaGo beat world champion Lee Sidal at the game of Go. Until then, no program had come anywhere near to mastering this complex game let alone beating a world-class human player; many doubted that it was even possible.
The traditional method of teaching a machine to play games was to load its memory banks with hundreds and thousands of past games which humans had already played; the machine would then play by precedent, recognising each position it found itself confronted with and then playing the best move recorded for it. In other words, it didn’t play better than human beings, it just played like the best human players of all time all playing on the same side. Hassabis, however, thanks to his researches into cognitive neuroscience and his theory of ‘scene construction’ (concerning the way memory and imagination work together), has produced algorithms which enable machines to play each other and learn from their moves as they do so, rather than simply draw on massive amounts of data from matches which human beings have already played.
This means that machines learn much, much quicker and develop strengths far beyond that of the human brain. These days computers are so advanced that they’re shedding light on areas of chess and backgammon which the human mind has hitherto been unaware of. (And as well as playing games, they can analyse games so quickly and accurately that the most skilful play in a backgammon game can be instantly identified, divorced from the element of luck; in recent months, tournaments have begun to award points to the most skilful players even if they lose.)
Second, and most importantly, the creation of a machine capable of out-playing and out-thinking human beings isn’t an end in itself as far as Hassabis and DeepMind are concerned. It’s simply the means towards the higher goal of creating artificial intelligence which can solve mankind’s problems and improve his life and world. The modelling of proteins by DeepMind’s tool AlphaFold in last year’s CASP (Critical Assessments of Techniques for Protein Structure Prediction) is its first big success. But it’s only the beginning. Almost certainly there’ll be more to follow.
DeepMind’s work is a reassuring counter-balance to the current fears and warnings about a dystopian future dominated by an all-powerful but hostile (or even merely neutral) artificial intelligence. Mr Hassabis predicts that AI will be “one of the most beneficial technologies of mankind ever”.
A footnote for the current lockdown; all this suggests that, if you have young children stranded at home, with schools closed, perhaps the best thing you can do for them is to forget trying to follow a syllabus on-line and instead simply teach them how to play chess, or backgammon, or bridge, or draughts, or Go, or ludo, or any board game. Playing board games with them must be easier and more enjoyable than trying to home-tutor them in academic subjects. And with a recent government report concluding that computer games don’t harm children’s cognitive abilities but indeed enhance them, who could blame you for leaving them to follow in Demis Hassabis’s footsteps while you get on with your own work or read the paper or help yourself to a well-earned drink?