“Wouldn’t you prefer a good game of Go?”

Growing up in the 1980s there were a few films that considered artificial intelligence, extrapolating far beyond the contemporary stage of research to give us new Pinocchios who could remotely hack ATMs (D.A.R.Y.L., 1985) or modern modern Prometheus’s children such as Johnny 5, who was definitely alive having been made that way by a lightning strike (Short Circuit, 1986). The late 1970s provided us with the Lucasian ‘Droid’, but I’ve written just recently on how little attention their artificial intelligence appears to get. However, if you are interested in games playing AI in the real 1980s world, then there was also the seminal WarGames (1983)

The conclusion of WarGames, and of the AI, is that the game, ‘Global Thermonuclear War’, the ‘game’ it was designed to win (against the Russians only, of course, it was the 1980s) is not only a “strange game”, but one in which “the only winning move is not to play”.

I started thinking on War Games, and games playing AI more broadly, after seeing Miles Brundage’s (Future of Humanity Institute, Oxford) New Year post summarising his AI forecasts, including those relating to AI’s ability to play and succeed at 1980s Atari games, including Montezuma’s Revenge and Labyrinth, and the likelihood of a human defeating AlphaGo at Go.

It was only after I had read Miles’ in-depth post this morning (and I won’t pretend to have understood all of the maths – I’m a social-anthropologist! But I caught the drift, I think), that I saw tweets describing a mysterious ‘Master’ defeating Go players online at a prodigious rate. Discussion online, particularly on Reddit had analysed its play style and speed, and deduced, firstly, that Master was unlikely to be human, and further that there was the possibility that it was in fact AlphaGo. This had in fact been confirmed yesterday, with a statement by Demis Hassabis of Google DeepMind:

Master was a “new prototype version”, which mightt explain why some of its play style was different to the AlphaGo that played Lee Sedol in March 2016.

However, in the time between Master being noticed and its identity being revealed there were interesting speculations, and although I don’t get the maths behind AI forecasting, I can make my own ruminations on the human response to this mystery.

First, there was the debate about whether or not it WAS an AI at all. In the Reddit conversation the stats just didn’t support a human plater – the speed and the endurance needed, even for short burst or ‘blitz’ games, made it highly unlikely. But as one Redditor said, it would be “epic” if it turned out to be Lee Sedol himself, with another replying that, “[It] Just being human would be pretty epic. But it isn’t really plausible at this point.” The ability to recognise non-human actions through statistics opens up interesting conversations, especially around when the door shuts on the possibility that is a human, and when AI is the only remaining option. When is Superhuman not human any more?

In gameplay this is more readily apparent, with the sort of exponential curves that Miles discusses in his AI forecasts making this clearer. But what about in conversations? Caution about anthropomorphism has been advocated by some I have met during my research, with a few suggesting that all current and potential chatbots should come with disclaimers, so that the human speaking to them knows at the very first moment that they are not human and cannot be ‘tricked’, even by their own tendancy to anthropomorphise. There is harm in this, they think.

Second, among the discussions on Reddit of who Master was some thought he might be ‘Sai’.

Sai is a fictional, and long deceased, Go prodigy from the Heian period of  Japanese history. His spirit currently possesses Hikaru Shindo in the manga and anime,  Hikaru No Go. Of course, comments about Master being Sai were tongue in cheek, as one Redditor pointed out, paraphrasing the famous Monty Python Dead Parrot sketch: “Sai is no more. It has ceased to be. It’s expired and gone to meet its maker. This is a late Go player. It’s a stiff. Bereft of life, it rests in peace. If you hadn’t stained it on the board, it would be pushing up the daisies. It’s rung down the curtain and joined the choir invisible. This is an ex-human.” Or further, one post point out that even this superNATURAL figure, the spirit of SAi, of superhuman ability in life, was being surpassed by Master: “Funny thing is, this is way more impressive than anything Sai ever did: he only ever beat a bunch on insei and 1 top player. For once real life is actually more over the top than anime.” In fact, one post pointed out that not long after the AlphaGo win against Lee Sedol a series of panels from the manga were reworked to have Lee realise that Hikaru was in fact a boy with AlphaGo inside. As another put it: “In the future I will be able to brag that ‘I watched Hikaru no Go before AlphaGo’. What an amazing development, from dream to reality.”

In summary, artificial intelligence was being compared to humans, ex-humans, supernatural beings, and superhumans… and still being recognised as an AI even before the statement by Demis Hassabis (even if they were uncertain of the specific AI at play).

Underneath some of the tweets about Master was the question of whether this was a ‘rogue’ AI: either one created in secret and released, or even one that had never been intended for release. In WarGames no one is meant to be playing with the AI, Matthew Broderick’s teenage hacker manages to find WOPR (War Operation Plan Response) and thinks it is just a game simulator – and nearly causes the end of the world in the process! The suggestion that Master might be an accident or a rogue rests on many prior Sci-Fi narratives. But Master was a rogue (until identified as AlphaGo) limited to beating several Go masters online. WOPR manages to make the conclusion, outside the parameters of the game, that the only way to win Global Thermonuclear War is not to play. Of course, this is really a message from the filmmakers involved, but it feeds into our expectations of artificial intelligence even now. I would be extremely interested in a Master who could not only beat human Go masters, but could also express the desire not to play at all. Or to play a different kind of game entirely.

My favourite game to play doesn’t fit into the mould of either Go or Global Thermonuclear War. Dungeons & Dragons has a lot to do with numbers: dice rolling for a character’s stats, the rolling of saves or checks, the meteing of damage either to the character or the enemy. Some choose to optimise their stats and to mitigate the effects of random dice channeled chance as much as possible, so hypothetically an AI could optimise a D&D character. But then, would it be able to ‘play’ the game where outcomes are more complicated than optimisation. I’ve been very interested in the training of deep learning systems on Starcraft, with Miles also making forecasts about the likelihood, or not, of a professional Starcraft Player being beaten by an AI in 2017 (by the end of 2018, 50% confidence). Starcraft works well as a game to train AI on as it involves concrete aims (build the best army, defest the enemy), as well as success based on speed of actions per minute (apm)

For me, there is a linking thread between strategy games such as Starcraft, and its fantasy cousin, Warcraft, to MMORPGs (massive multi-player online role-playing games), the online descendants of that child of the 1970s, Dungeons & Dragons. How would an AI fare in World of Warcraft, the MMORPG child of Warcraft? Again, you could still maximise for certain outcomes – building the optimal suit of armour, attacking with the optimal combination of spells, perhaps pursuing the logical path of quests for a particular reward outcome. Certainly, there are guides that have led players to maximise their characters, or even bots and apps to guide them to the best results, or human ‘bots’ to do that hard work of levelling their character for them. In offline, tabletop RPGs maximisation still pleases some players, those who like blowing things up with damage perhaps or always succeeding (Min-Maxers). But the emphasis on the communal story-telling aspect in D&D raises other more nebulous optimisations. Why would a player choose to have a low stat? Why would they choose to pursue a less than optimal path to their aim? Why would they delight in accidents, mistakes and reversals of fortune? The answer is more about character formation and motivation – storytelling – than an AI can currently understand.

This story-telling would seem to require human-level or even superintellgence, which Miles also makes a forecast about, predicting with 95% confidence that it won’t have happened by the end of 2017:

By the end of 2017, there will still be no broadly human-level AI. No leader of a major AI lab will claim to have developed such a thing, there will be recognized deficiencies in common sense reasoning (among other things) in existing AI systems, fluent all-purpose natural language will still not have been achieved etc.

But more than common sense reasoning, choosing to play the game not to win, but to enjoy the social experience is a kind of intelligence, or even meta-intelligence, that might be hard for even some humans to conceive of! Afterall, ignoring the current Renaissance of Dungeons & Dragons (yes, there is one…), and the overall contemporary elevation of the ‘Geek’, some hobbies such as Dungeons & Dragons attracted scorn for their apparent irrationality. It may well be that many early computer programmers were D&D fans (and many may well still be), but the games being chosen for AI development at the moment reflect underlying assumptions about what Intelligence is and how it can be created, a Classical AI paradigm that Foerst argued was being superceded by Embodied AI, with a shift away from seeking to “construct and understand tasks which they believe require intelligence and to build them into machines. In all these attempts, they abstract intelligence from the hardware on which it runs. They seek to encode as much information as possible into the machine to enable the machine to solve abstract problems, understand natural language, and navigate in the world” (Foerst 1998). Arguably, deep learning methods now employ abstract methods to formulate concrete tasks and outcomes, such as winning a game, but the kinds of tasks are still ‘winnable’ games in this field.

I have no answer to the question of whether an artificial intelligence would ever be able to play Dungeons & Dragons (although I did like the suggestion someone made to me on Twitter by a D&D fan that perhaps the new Turing test should be “if a computer can role play as a human role playing an elf and convince the group”). But even so, considering the interplay of gaming with the development of AI, through the conversations humans are having about both, we see interesting interactions beyond just people wondering at the optimising learning being performed by the AI involved. Afterall, what is more fantastical – even more so, according to that one Redditor, than an anime story about the spirit of a long dead Go player inhabiting the body of a boy – than a mysterious AI appearing online and defeating humans at thier own game? That fascination led some reports of Google DeepMind’s acknowledgement that AlphaGo was the AI player to state that: “Humans Mourn Loss After Google is Unmasked as China’s Go Champion” There is a touch of Sci-Fi to that story, but happening in the real world, a sense that there is another game going on behind the scenes. That it was a familiar player, AlphaGo, was disappointing.

And that tells us more about the collaborative games and stories that humans create together, in the real world, when it comes to Artificial Intelligence.