Richard Powers - Galatea 2.2

'There is that. But I was thinking how you want to put everything into words. Think about this. You want to make a data structure that will say everything there is to say about 'ball.' You have to have facts for roundness, cohesion, size, weight. You have to have all sorts of probabilistic rules. It's more likely to be ten inches in diameter than ten feet."

"Unless it's a wrecking ball."

"It's more likely to be made of rubber or plastic or wood than of, say, plant fiber."

"Unless it's a softball, or a cotton ball. ."

"Or a fuzzball." Diana laughed.

"Or a spitball. Or a ball of yarn."

"Enough already. You're making me crazy. The list of predicates is — forever. And the exception list is even longer. You can talk through a whole encyclopedia, but you won't yet have said 'ball' itself. Let alone what you do with one. What one means. How to throw one. How a kid feels when he gets one for a birthday present."

"But that's where associative learning comes in, isn't it? We don't have to enumerate all those qualities. If the machine keeps coming across them, and sometimes people are gazing into crystals, and sometimes they're pitching a curve, and sometimes they are having—"

"That's just the problem. Any baby can hold a ball in its hands. Your machine can't. How many words is it going to take to say what that globe feels like? The heft of the thing. The possibility."

"Red rubber sphere that you can chuck. ."

"Richie. Richie."

I didn't have the heart to tell her that no one called me that. Besides, it sounded good in her mouth.

"Are you going to do 'red' as fitting between x and y angstroms?"

"Not at all. That's the classical AI people. Chen and Keluga. We're going to do red sky, red-faced, red flag, in the red, red dress. ."

"Sexist. Okay. Go on to 'sphere.' How is the poor thing supposed to get 'all the points equidistant from a center' when it has never seen distance? When it can't possibly measure?"

"Good point. You mean we have a problem. No symbolic grounding."

"Symbolic grounding." Diana grinned, and wiped the tartar sauce from her lower lip. "That's the phrase I was looking for."

"You're saying a reading machine is hopeless. We should give it up."

"I'm saying, if you're going to make such a thing, you have to give it eyes, hands, ears. A real interface onto the outside."

"The literary theorists think a human's real-world interface is problematic at best. And greatly overrated. They say even sense data must be put into symbols."

"The literary theorists have to get tenure. And they have no hard facts to get tenure with. They have to fight for a slice of a pie that's getting smaller every day."

"Getting smaller because of people like you."

I was teasing. She took me at face value. "Yes, maybe. But tell me, then. Why do your people need to either emulate mine, attack us as nature molesters, or dismiss us as irrelevant self-deluders?"

"My people are scared shitless of your people, that's why. They're terrified that Dad and Mom really do love you best."

Diana smiled like a teenager. "Are you kidding? We're the problem child. Don't you remember? You heard the parents tell us never to play with matches."

We walked outside into a day remade. The trees had stripped bare, but for a moment, I couldn't tell if they were coming or had just gone. We two were the only ones on the whole street wearing coats. A pack of earphoned roller bladers with wraparound shades so silvered they couldn't see almost mowed us down.

My disorientation was total. "Wasn't it almost winter when we went inside?"

"It's getting so that these once-in-a-century freak weather spells occur every other week."

"That's the fault of you scientists."

"Doubtless."

"It's still nippy. Have I lost my mind, or are those women wearing shorts?"

"Youth is never cold," Diana lamented.

Ten years ago, fifteen, I'd pitied those who'd let their skin go papery and their blood thin. Back then, this had been my haunt. Now I'd lost the lease, could not even name the street.

"I want it to freeze. Freeze hard."

"It's going to freeze," Diana guaranteed. "You've forgotten. It's rough out in these parts. Four months of interstellar blackness. Deep space. Zero. The air loads up with sleet BBs, with nothing between here and Pennsylvania to slow the wind."

"How long have you been here?"

"Five years. You know the funny thing? I never thought of myself as a Californian when I lived there."

We walked toward campus. "Does the brain have a 'decaying leaf recognition' subsystem? A 'sheepskin coat just back from the cleaners' subsystem?"

"I know what you mean," Diana said. "Smell is everything, isn't it? When I smell the air right now, I think, I know what's coming. Time to batten down."

"I think, how can I leave all this?"

She said nothing for the longest time. Then Diana said, "Thrill."

"Yes. Thrill."

It took me until the Quad, until we walked over the spot, to make the connection. To realize what day it was. You see, it didn't feel like November. Same day, different year: I no longer knew what people meant by that. Four days later, it snowed.

"Symbolic grounding," I told Lentz.

D, our first implementation to run on the sprawling community of connection machine, was my slowest charge so far. But parallel architecture and several software twists made it the one with the most astonishing reach. D came into this world recursive. It took forever to grasp that two was the integer just past one. But the instant it got that, it had infinity in the same breath. It could watch itself learn. When it got "Dogs bark," it also got "Baby says, 'Dogs bark.' " The breakout, when it came, was going to be unbelievable.

The shape of its brain was generative. It matched the arc of an uncoiling sentence. But deep syntax wasn't enough. Words weren't enough. "The policeman gave the motorist his badge," I told it. "Who does 'his' stand for? The policeman gave the motorist his license." Now who?

How much would D have to know before it could get that one? Tracing my own inference process dizzied me. The linked list of properties hanging off the world's every object required a combinatorial explosion of sense. But to see even part of what the host symbols stood for required that D first see.

"We have to give it eyes," I decided. "Why train it with all the properties of a ball when we can show it what a ball is?"

"Nice leap, Marcel," Lentz said. "I've been waiting for you to make it."

"And here I was afraid you were going to kill me for bringing it up."

"On the contrary. In fact, I still have a passive retinal matrix lying around intact from work I did last year. We can paste it in."

"Now, where the hell did I put that eyeball? It was around here somewhere."

Lentz laughed. "All right. So the work habits could be more systematic."

Sight was not the sudden beamburst I hoped for. Revision E could convert objects into retinoptic neurode maps. Training got it to associate words with each visual clump. But navigating by this crayon cartography resembled sewing silk with a Lincoln log needle.

I gave it several common objects, in stiff cross section. E made only static traces: photos, not footage. I sorely doubted that the speckles of light and dark did anything to round out the ball in E's associative memory. We boosted the resolution. Added bits for color, sixteen million shades. Whether E traveled those surfaces or twirled them in its mental space I couldn't say.

The creature in Frankenstein learned to speak by eavesdropping on an exiled family, the most astonishing act of language acquisition until Taylor's beloved Tarzan, the books on which the best reader I ever met grew up. Frankenstein's creature had his chattering family and a knapsack of classics: Paradise Last, Plutarch's Lives, Goethe's Werther. E, like Tarzan, learned to talk more or less on print alone.