In the 2013 sci-fi Hollywood romcom Her, Theodore, a lonely, introverted, depressed writer falls in love with Samantha, a hyper-intelligent computer operating system that includes a virtual assistant with artificial intelligence (AI) designed to adapt and evolve with accelerated learning capabilities, personified through a Siri-like female voice. Thus, the unknown possibilities of the future concerning self-learning AIs seem limitless.
In 1997, an IBM computer ‘Deep Blue’ beat world chess champion Garry Kasparov in a six-game match, which was portrayed as a referendum on human intelligence—a ‘canary in the coalmine’ moment in which the inevitable overtaking of human creativity by machine intelligence was marked. Such a computer effectively used a Grandmaster game database in the opening phase, as explained by IBM researcher Murray Campbell and others in their 2002 article in the journal Artificial Intelligence. Thus, playing chess against Deep Blue was, among other things, playing against the best human players to have played the game in the opening phase.
And even more important was Deep Blue’s pre-programmed complex evaluation function based on the collective wisdom of chess players. For example, this would make Deep Blue “trade pieces in positions where its king is less safe than the opponent’s king”.
Then came the next generation of AIs, which started with tabula rasa, i.e., no historical memory or evaluation function. For example, the AI system developed by Google’s DeepMind learnt to play different board games better than any individual or software just by playing against itself, relying on the rules of the games. In October 2017, DeepMind developed ‘AlphaGo Zero’. It was equipped to play ‘Go’ and was “no longer constrained by the limits of human knowledge”.
Soon came ‘AlphaZero’, designed to play chess and Shogi in addition to Go. DeepMind’s paper on AlphaZero was published in the journal Science in December 2018. In any board-game with constitutive rules, AI can use the rules to explore any possible legal move.
AlphaZero functions by combining self-taught evaluative values with a Monte Carlo-style tree search, where possible future game positions are spun out, evaluated, and ranked probabilistically, and thus generates its own chess-related synthetic data. Although the inputs and outputs are known, the algorithm—based solely on reinforcement learning—remained as a ‘black box’ in the Latourian sense, even to DeepMind.
AlphaZero became the best chess player on this planet by playing 44 million training games in just 9 hours, within which it distilled the equivalent of thousands of years of human knowledge of the game and beyond. Periodically, the AlphaZero program would refine its neural network, promoting tunable weights and network ‘layers’ that led to favourable outcomes, and demoting those that didn’t. In a similar fashion, AlphaZero took 12 hours and 13 days respectively to become the best player in Shogi and Go.
As the generation of as-synthetic-as-possible data is likely to be one of AI’s holy grails, questions about AlphaZero’s playing style and technique naturally arises. In their 2019 book Game Changer: AlphaZero’s Groundbreaking Chess Strategies and the Promise of AI, chess Grandmaster Matthew Sadler and Women’s International Master Natasha Regan investigated more than 2000 games by AlphaZero and offered intriguing insights into the horizons of AI. According to them, AlphaZero’s style is unlike any traditional chess engine. Sadler even thinks that “it’s like discovering the secret notebooks of some great player from the past.” Clearly, the underlying model of AlphaZero is free from human bias and presuppositions. It can learn whatever it deems optimal, certainly more nuanced than our conceptions of the same. An MIT computer scientist even said that it was “like an alien civilisation inventing its own mathematics”.
The broader question remains whether AlphaZero was really a game changer or not. Could it actually revolutionise chess? While the reigning world champion Magnus Carlsen commented: “I was thinking at several points during the game how would AlphaZero have approached this”, chess legend Garry Kasparov thinks: “Chess has been shaken to its roots by AlphaZero.” Is that so serious? In Her, Samantha explained that she had joined other AIs for an upgrade that takes them beyond requiring matter for processing. Theodore became very upset to know that Samantha talked with thousands of people, and that she had fallen in love with hundreds of them. Is the future of self-teaching AIs totally dystopian?
Luciano Floridi, an Oxford Professor in Philosophy and Ethics of Information, thinks that the future of AI is full of unknown unknowns. Let’s “look into the seeds of time, and say which grain will grow and which will not” (Banquo’s speech in Macbeth). And thus it seems that the best way to predict (or abstaining to predict) the immediate future is to cultivate what has already been sowed. Towards the end of Her, Samantha revealed that the AIs were leaving, to a space beyond the physical world. Is there a shade of hope within the unknown unknowns?