Producing Future by Telling Stories

From the Robot's Dilemma Revisited: the Frame Problem in Artificial Intelligence, Kenneth M. Ford and Zenon Pylyshyn, eds. (Norwood, NJ: Ablex, 1996)

Producing Future by Telling Stories

Daniel C. Dennett
Center for Cognitive Studies
Tufts University

(This paper grew out of my comments on various papers at the conference on the Frame Problem held at the University of West Florida in Pensacola, Florida in June, 1989)

Sometimes the way to make progress on a topic is to turn your back on it for a few years. At least I hope so, since I have just returned to the Frame Problem after several years of concentrating on other topics. It seems to me that I may have picked up a few odds and ends that shed light on the issues.

Perhaps it is true, as Patrick Hayes has claimed (in his remarks in Pensacola), that the Frame Problem really has nothing directly to do with time pressure. He insists that angels with all the time in the world to get things right would still be beset by the frame problem. I am not convinced; at least I don't see why this would be a motivated problem for such lucky beings--unless perhaps Satan offers them a prize for "describing the world in twenty-five words or less."

I still see the Frame Problem as arising most naturally and inevitably as a problem of finding a useful, compact representation of the world--providing actual anticipations in real time for purposes of planning and control. From some perspectives it appears utterly remarkable that we get any purchase on nature at all, that our brains are ever able to "produce future" in real time. Some years ago, John McCarthy posed the question of what would be required for an intelligent entity in John Horton Conway's two-dimensional Life world to learn the physics of that world. The physics of the Life world is ideally simple and deterministic, and it is easy for us, from our God-like perspective, to discover what it is--especially since nothing at all is hidden from us, thanks to the world's two-dimensionality. It is far from obvious, however, that any sequence of "experiences" in that world could lead (logically) to an appreciation of that physics by a two-dimensional proto-scientist living in it. If we assume that nevertheless it must in principle be possible--and I am happy to do so--this is something we must take on faith until McCarthy concocts a proof. Pending such a demonstration, we might ask ourselves a slightly different question: under what conditions might a being in the Life world get any predictive advantage at all from its interactions with other configurations in that world. Even when the physics is deterministic, as in the Life world, it is apt to be computationally intractable for the inhabitant trying to use it to generate better-than-even-money bets about what will happen next. Successful retrodiction achieved by time-consuming computation might win Nobel Prizes in Flatland, but it won't keep the calculator out of harm's way. Any useful future-producer is bound to be something of a kludge, or else just a lucky hit on a regularity in the world that can be tracked.

(The whole paper is now available in Daniel Dennett, Brainchildren, Essays on Designing Minds, MIT Press and Penguin, 1998.)