The idea that we may come to observe explosive developments in the area of Artificial Intelligence (AI) has deserved a lot of press lately. In the latest Web Summit, in Lisbon, Stephen Hawking, in one of his last public speeches, warned about the risks of AI and made a number of references to the importance of making sure that future systems will have their interests well aligned with the interests of mankind. This is not a new idea and, indeed, it has a long history.
Irving John Good was an English mathematician, who started his career working with Alan Turing in Bletchley Park, in the team that decoded the German messages encoded by the Enigma machines, an effort which strongly contributed to the allied victory. Later, he became a lecturer at Manchester University and a professor at Virginia Tech, authoring many articles on statistics and on the future of computation and AI. He was a consultant for the mythical movie “2001: A Space Odyssey”, where the main actor was the super-intelligent computer HAL 9000, who had its own objectives and motivations, not always aligned with those of the crew.
We owe to Irving John Good the statement that, if we ever manage to build a machine more intelligent that a human being, then that will be the last invention of mankind:
“Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”
Recent developments in the area of AI brought with them violent discussions about the dangers of super-intelligent machines, bringing to mind Terminator style scenarios, where machines fight against men. The truth is that, in the actual status of technology, no known method has the ability, or even the ambition to, in the near term, develop machines as the ones referred to by Irving John Good, more than half a century ago. Every system designed to date addresses and tries to solve very specific problems and there are no concrete proposals for the development of strong AI, a system of the type that would be required to create the explosion of intelligence he alluded to.
However, it is interesting to note that, in some specific and very restricted areas, we have already observed the self-sustaining acceleration mechanism necessary to sustain an explosion of intelligence. One of the areas is the game of Go, a board game, which is played by setting down black and white stones in a 19 by 19 board, conquering space and enemy stones by encircling them. The game is a metaphor for a battle between two armies, striving for the conquest of land, and is enormously popular, specially in Asian countries. Given the high branching factor of the game (the number of moves possible at each moment), and the impossibility of using brute force to find the long term consequences of particular moves, this game has kept itself out of the range of computers for decades, despite many efforts by the scientific community. Only a few years ago it was believed that many decades would pass until a program could beat the world’s best players.
Yet, in 2016, the company DeepMind, which was acquired soon after by Google, by more than 500 million dollars, announced that the programa AlphaGo had beaten one of the world’s top players, Lee Sedol. AlphaGo used machine learning and search techniques to learn, from thousand of human games, the best move to use in each position. The deep neural networks used by AlphaGo managed to learn the techniques of human champions and used them to beat one of the top human players. Deep neural networks use thousand of neuron-like units, configured in order to learn specific behaviors, in this case the move that should be played in each situation. They are useful to process images, sounds and videos or, in this case, Go positions. Although neural networks have been in use for many decades, only recently was possible to address many real-world problems with this technology. Tasks like face recognition or translation of spoken language became possible with the development of new training techniques, massive volumes of training data and fast computing devices. In the case of AlphaGo, the deep neural networks were combined with a mechanism called reinforcement learning, which makes it possible for a system learn the right actions in a sequence of steps, given only the final outcome of the sequence. This technique is used, for instance, to teach robots to walk or cars to drive themselves. Similar mechanisms are, presumably, used by the human brain to learn complex tasks with many steps.
AlphaGo represented, by itself, a very significant result, since it demonstrated that a program can learn to excel in a given area (in this case, a complex board game), by simply observing the behavior of the best humans in that area and extracting the knowledge required to master the domain.
However, only a few weeks later, DeepMind announce an even more impressive result, obtained with a new program, AlphaGo Zero. This system, described in an article published in Nature, in October 2017, learned to play Go, from scratch, without looking at any games played by human experts. As a child that learns by herself, the system learned to play Go by simply playing against another copy of itself, learning, progressively, to improve the final score. In the beginning, the moves were almost random and the program played very poorly but, as time passed, the program reached beginner level, rediscovering tactics and moves normally used by beginners. After 40 days, the program rediscovered all the knowledge accumulated by humanity for millennia and even created new tactics and strategies which represent new knowledge about the game. The final version of AlphaGo Zero managed to beat AlphaGo by a stunning 100 to 0, a result which puts AlphaGo Zero in a super-human playing level.
Although in a specific and very limited domain, it is impossible not to see in this result a premonition of what will be a common reality in the future. As we develop further intelligent systems in specific domains, we may start to see them improving themselves in a self-improving loop, which requires no human intervention.
Irving John Good, who died in 2009, did not have the opportunity to see this success of Artificial Intelligence, this particular invention of mankind. Which, of course, will not be the last…
Picture by Hoge Rielen, available at Wikimedia Commons.
This article was published in Portuguese in newspaper Público, in November 2017.