(image via NYPL)
Guess what? I’m challenging you to a game of tennis in three days. Here’s an issue though: I don’t know anything about tennis and have never played it, and the same goes for you.
In order to prepare for the game, we are each going to do something very different. I’m going to practice playing with someone else who isn’t very good. You, meanwhile, are going to train with an expert. But you are only going to train by talking about tennis with the expert, and never actually play. The expert will tell you everything you need to know in order to win at tennis, but you won’t actually get any practice.
Chances are I’m going to win the game. Why? Because the task of playing tennis isn’t just reducible to learning a set of things to do in a certain order. There’s a level of knowledge and skills that become unconsciously incorporated into the body. As David Foster Wallace wrote about tennis, “The sort of thinking involved is the sort that can be done only by a living and highly conscious entity, and then it can really be done only unconsciously, i.e., by fusing talent with repetition to such an extent that the variables are combined and controlled without conscious thought.” Practicing doesn’t mean learning rules faster; it means your body knows instinctively where to put the tennis racket.
The same can be said of most skills, like learning how to play an instrument. Expert musicians instinctively know how the instrument works. And the same goes for driving. Drivers obviously learn certain rules (“stop at the stop sign”) and heuristics (“slow down during rain”), but much of driving is done unconsciously and reflexively. Indeed a driver who needs to think through procedurally how to deal with, say, a snowy off ramp will be more at risk of an accident than someone who instinctively knows what to do. A proficient driver is one who can spend their mental energy making more subtle and refined decisions based on determining what is salient about a specific situation, as past experiences unconsciously influence current experiences. Our bodies and minds aren’t just a series of logic statements but also a series of lived-through meanings.
This is my intro-level remembrance of Hubert Dreyfus’ argument against artificial intelligence via Merleau-Ponty’s phenomenology (more via Wikipedia). It’s been a long time since I followed any of this, and I’m not able to keep up with the current debates. As I understand it Dreyfus’ arguments were hated by computers scientists in the 1970s, then appreciated in the 1990s, and now computer scientists assume cheap computing power can use brute force and some probability theory to work around it.
But my vague memory of these debates is why I imagine driverless cars are going to hit a much bigger obstacle than most. I was reminded of all this via a recent article on Slate about Google’s driverless cars from Lee Gomes:
[T]he Google car was able to do so much more than its predecessors in large part because the company had the resources to do something no other robotic car research project ever could: develop an ingenious but extremely expensive mapping system. These maps contain the exact three-dimensional location of streetlights, stop signs, crosswalks, lane markings, and every other crucial aspect of a roadway […] But the maps have problems, starting with the fact that the car can’t travel a single inch without one. […]
Because it can’t tell the difference between a big rock and a crumbled-up piece of newspaper, it will try to drive around both if it encounters either sitting in the middle of the road. […] Computer scientists have various names for the ability to synthesize and respond to this barrage of unpredictable information: “generalized intelligence,” “situational awareness,” “everyday common sense.” It’s been the dream of artificial intelligence researchers since the advent of computers. And it remains just that.
Focus your attention on the issue that the car can’t tell the difference between a dangerous rock to avoid and a newspaper to drive through. As John Dewey found when he demolished the notion of a reflex arc, reflexes become instinctual so attention is paid only when something new breaks the habitual response. Or, experienced human drivers don’t see the rock, and then decide to move. They just as much decide to move because that forces them to see the rock. The functionalist breakdown, necessary to the propositional logic of computer programming, is just an ex post justification for a whole, organic action. This is the “everyday common sense” alluded to in the piece.
Or let’s put it a different way. Imagine learning tennis by setting up one of those machines that shoots tennis balls at you, the same repetitive way. There would be a strict limit in how much you could learn, or how much that one motion would translate into you being able to play an entire game. But by teaching cars to drive by essentially having them follow a map means that they are playing tennis by just repeating the same ball toss, over and over again.
Again, I’m willing to sustain the argument that the pure, brute force of computing power will be enough – stack enough processors on top of each other and they’ll eventually bang out an answer on what to do. But if the current action requires telling cars absolutely everything that will be around them, instead of some sort of computational ability react to the road itself, including via experience, this will be a much harder issue. I hope it works, but maybe we can slow down the victory laps that are already calling massive overhauls to our understanding of public policy (like the idea that public buses are obsolete) until these cars encounter a situation they don’t know in advance.