The Overmind vs the Hivemind

I was introduced to the Overmind last week, which is a Starcraft (Brood War) AI developed at Berkeley. Starcraft is a "real-time strategy" game, abbreviated as "RTS", where two or more players face-off, competing for resources and building troops to fight and ultimately destroy each other. Last one standing is the victor.

It's an interesting game to develop AI for because it requires a great deal of intricate management, both at the macro level (managing your economy; that is, resource gathering and allocation) and at the micro level (positioning and controlling your individual troops in battle). There's enough complexity that a game can go in many unexpected ways; it is an ideal environment for humans to creatively respond and adapt to - something which computers traditionally have a great deal of trouble with.

When I think of "AI vs human", I tend to think of it in a sort of Deep Blue vs Garry Kasparov sense. A solitary expert against an expertly programmed machine. The machine proves its superiority by consistently beating the human. The Overmind fits that narrative (though it still had trouble with human professional Starcraft players).

It's interesting to think of the Overmind in contrast to Twitch Plays Pokemon, where masses of participants essentially button-mashed their way through Pokemon Red (and is now making its way through the other titles). Anyone can join the game's chat room and submit a button to press. The system had two modes: "anarchy" and "democracy", the former meaning that every command someone submitted was executed, the latter meaning that there was consensus on what the next command would be. It took 16 days non-stop to complete the game which typically takes a bit more than a day. TPP is the complete antithesis of unitary control by a program.

Sci-fi AI narratives (and the AI field's ultimate aspirations) are far more ambitious than these video game bots. In Iain M. Banks' The Culture, general artificial intelligences are organizing galactic societies, thus liberating organic beings from the complexities of deciding policy and the stress of providing for and governing themselves.

This is a huge leap from video game bots-impressive as they are-but it was explored long before the conception of these bots, as early as the mid-20th century, in Soviet Russia. Francis Spufford's Red Plenty details an attempt at integrating computers into the central planning model:

Soviet planners, economists, physicists and mathematicians...persuaded the Soviet leadership that, using cybernetic principles and the newly developed computers, the centralised, planned Soviet economy could at last be made efficient.

These ambitions for AI are far removed from one-on-one video game contests. Rather than substituting for an individual, these programs are meant to replace the collective decision-making capacity of entire societies of people. Naturally, at this scale the evaluation of success is much murkier than a game with simple victory conditions. And when we begin discussing decisions which impact peoples' lives, the particular strategies the AI uses become matters of ethics than merely mechanics.

It feels as if we are still quite a long way off from a managerial AI becoming a reality. But with the frenzy around big data, where companies like Palantir aggregate massive amounts of disparate data to draw and act upon high-level conclusions, the technical feasibility of a computational caretaker (/overlord) will only grow. When will we see organizations managed by "Computer Executive Officers" instead of by people? In The Culture, such artificial intelligences are the foundation for utopia, but they are equally the fodder of nightmarish futures - Skynet, ARIIA, etc... - will we be willing to use them?