A historic change, only noticed by a small minority of enthusiasts
A recent scientific breakthrough made possible by AI has been well publicised by the media, but the main learning opportunity it presents has remained unheralded.
I am referring to AlphaZero’s victory in a chess game: this AI from Google was initially programmed to learn the game of Go, then extended to chess and Shogi.
It has received accolades from some for its learning speed: 4 hours was enough for the machine to become a leading chess master. Impressive of course, but that only reflects its extraordinary processing power, not a scientific breakthrough.
Others focussed on the fact that AlphaZero improved its game by playing against itself, starting from zero knowledge on the game: no library of opening moves, or strategic principles. This modern form of solipsism is indeed noteworthy. But it only proves that chess, like all strategy games, is a closed semantic universe, however complex it may be.
However, we now come to the main achievement: the nature of AlphaZero’s opponent: Stockfish, the most powerful chess program based on combinatorial computation. The final score out of 100 games played: 28 wins, 72 draws and no defeats. AI proved victorious over brute force, where Stockfish already far exceeds the best human players.
More important than the victory itself is the way it was achieved. It was not lost on the chess-playing microcosm that they were experiencing a historical moment, not only for the game of chess, but for the very nature of cognition.
Google’s AI playing chess: revenge of the human
Paradoxically, AlphaZero’s resounding victory is not a sign of the machines increasing their crushing domination over man, but on the contrary the restoration of the human spirit.
Even the best human chess players fall very short of the combinatorial power of any machine. We compensate for this weakness in calculation by geometrical reasoning, by using pieces against their own side, by broad reasoning making it possible to reduce a very large number of combinations to a few main lines of possibility.
When Kasparov faced Deep Blue, IBM’s machine based on brute force calculation – even when he narrowly lost the second match – he showed that geometrical reasoning makes it possible to level the playing field with a machine even when calculating billions of times fewer combinations.
It is this human capacity which allows us to create mathematics. It gives human consideration a certain beauty, a geometrical purity, bringing together countless scenarios in a few lines.
Until recently, chess machines were based on combinatory brute force. Neither strategy nor holistic concept, the game seemed to be reduced to just a series of calculations. Under the pressure of these machines, the game became “ugly” and meaningless, a chaotic sequence of combinations, not a paradise teeming with brilliant designs. And if artists of the game did exist, their destiny would be to be infallibly outclassed by the brute force of raw calculation.
But AlphaZero’s game is quite different. AlphaZero reinvents the concepts that chess masters took centuries to forge: domination on certain squares, blocking of opponent pieces at the cost of material sacrifices to save time, advantage of a pair of bishops sweeping the board on both colours.
The miracle is that this idea prevails over brute force. Mind and time prevail over matter. The human nature of the game is restored and justified by the beauty of its reasoning processes: they are indeed the most effective, as well as being the most elegant. The game of chess is good at hiding the inner beauty that its devotees find in it.
The triumph of the mind
In that case, why have humans not maintained their dominion over the game playing machines? Because you still have to combine geometric meaning with unparalleled combinatory power to give beautiful designs the last word. The best human players manage to find the same geometric lines as AlphaZero, but are flawed in their execution. At the point where calculation supports a strategic vision, AlphaZero can assert its superiority over brute force.
This reversal of fortunes was eminently foreseeable. The best grand masters can beat the best computers if they themselves are assisted by a computer to which they outsource the analysis of options. A man assisted by a machine beats a machine.
Analysing AlphaZero’s games gives you something to marvel at. In one game it allows itself to sacrifice an entire piece and a pawn, solely to make things awkward for the opponent pieces between them, a move that no human would have the audacity to try faced with an analytical monster like Stockfish.
Faced with this new kind of adversary, Stockfish finds itself in the same position as the first-generation computers of the past, when grand masters humiliated them: its game seems “unbalanced”, its analytical power only makes it myopic, enduring its opponent’s initiative without understanding his intention. AlphaZero is a kind of compilation of everything that human ingenuity could find in chess, extended over a deeper concept.
The essence of all games
AlphaZero’s learning principle is as minimalist as it is essential. What has fascinated chess or Go enthusiasts is that they got the impression that the very essence of their game was being extracted as they watched. That AlphaZero is able to determine the intrinsic complexity and logic of any rule-based game.
You can imagine the range of possible applications of a discovery like this: many real-world phenomena, in hard sciences and in social sciences, can be modelled by games. Michel Crozier’s sociology of organisations, for example, shows the laws of human behaviour inside companies demonstrated with the certainty of a theorem. We always want to know what game we are playing, what train of thought we have boarded.
The dialogical man drives out the ghosts of transhumanism
Some will worry about this new demonstration of the power of AI: alarmist articles about transhumanism or the rise of the machines to gradually eliminate us will raise their heads again.
While the revolution heralded by AlphaZero is impressive, we must not allow it to mislead us about the nature of the dangers of AI.
The fundamental difference between man and machine is that we know how to manage in semantically open universes. We change the game when necessary, even creating the one we think best suits the situation, and we know when to do it.
That makes the complementary cooperation between man and machine a natural outcome. It’s up to us to navigate between the different games and to decide when it’s time to change them. It’s up to the machine to give us the essential keys for each game, better than we could do ourselves.
“Decision support” is not blind confidence in the machine’s diagnosis, but a regular consultation of its evaluation of a situation. Major decisions in economics, geopolitics or organisation could be taken with support from this intelligence.
This is a sort of “augmented man”, but it is the result of a choice, of a constant dialogue between man and the game-playing machines he has created for himself, not a prosthesis violating his biological integrity. In the same way, Renaissance man cultivated arts that made him more confident.
Don’t be afraid of your own intelligence
Curiously, it is always an excess of intelligence that we are afraid of, and which feeds the fantasies about AI eliminating mankind. While in fact the progress made by AI is only an exploration of our own intelligence.
Don’t be afraid to make AIs more and more intelligent, without limitation. The real dangers associated with AI come from our greed for power that makes us use everyone’s personal data in a totalitarian way. The danger lies in Big Data, if AI is pressed into its service blindly.
We must not allow ourselves to believe that such temptations will only occur in non-democratic countries. The current leaders of democracies are showing worrying signs of refusal to listen, and personal power grabbing. It doesn’t take AI for us to slide towards totalitarianism, our greed and our natural stupidity are more than enough for that.
Unlimited exploration of cognition is not a danger. On the contrary, it is the antidote to deviant use of technology. It is an extra intelligence that makes us develop encryption methods that preserve our individual freedom, for example homomorphic encryption that allows us to carry out “cloud computing” without revealing our personal data.
For those who are afraid of ghosts, there are two famous phrases to remember. “If you think intelligence is dangerous, try ignorance.” And the second, which we owe to Douglas Hofstadter: “Man will never be able to do without being intelligent, whatever intelligence he may deploy to achieve it”.
This post is also available in: FR (FR)DE (DE)