In the mornings, I'd try to write. I now had perhaps three pages. In those pages I did nothing but unwind my first line. I made my relative clauses jump through all sorts of novel hoops. The performance could have gone on for several hundred pages on style alone. I could get my train to leave most eloquently. But I could not get it to go anywhere but back.
Sustaining so much as a paragraph became impossible. I'd juggle the first verb in my head, overwhelmed by my own weighted backlist. I thought of the four books I'd written between leaving U. and returning. The round trip seemed too immense. I no longer had the heart to extend it.
I'd missed my connection. Stranded at the terminal. I didn't want to write anymore. I was sick of speculation and empathy and revision. All I wanted was to read word frequency lists to Implementation A.
Lentz thought it time to see if Implementation A made sense. We drilled it for days on two-word sentences. Then we prompted A with nouns, to see if it could supply a valid predication. We would see if A acquired, by example, the pattern of generating grammar.
It recognized the simple S-V phrases we fed it. Dogs bark. Birds soar. Night falls. You vanish. Father hugs. Baby cries. It could distinguish these ideas from the formulations that we earmarked as inappropriate for well-brought-up machines to use in public.
But when it came time to perform, A clutched. It seemed to stab at pattern-completion. It formed sentences by chance guess.
It could learn words. It could identify well-formed sentences. But Implementation A refused to elaborate its own. No matter how Lentz tweaked, Imp A seemed baffled by the task. It would not take the first generalizing step to learn the language.
Lentz tacked into the failure. He tried everything he could think of to coax the machine up a notch. Nothing worked. Asked to complete a thought, the network improvised wildly. Fish…? Fish sky. Shines? Hopes shines. Forests floor. Laugh efforts. Combs loneliness. Even when, by chance, it got the syntax right, the sense was not even random.
Imp A spoke the way a toddler gave directions. Its eager-to-please finger serpentined over all possible routes. This way? That's right! Or is it this way? Sure, if you like. Where's the horse? Come on, point to the horse. That's not a horse. That's not a horse either.
Yet it was, at least, pointing. Maybe our putting the request as pointedly as we did flustered it.
"Couldn't we hard-code a properties list? I mean, someone here must have compiled a semantic catalog. ."
"That's your coeval, Keluga. He's trying to write the entire Roget's as a series of nested, rule-based schematics. Containment, relation, exclusion. ."
"Coeval? I'm thirty-five, Lentz. That guy's a kid."
"You're both post-Tupperware. Chen is Keluga's supervisor. Outrageous. The Bergen and McCarthy of AI. Chen knows every algorithm in existence, but can't speak a single natural language, including his own. Keluga grew up on Hollywood robots and the microprocessor revolution. One of those brats who got his first PC free in a cereal box. He gets nervous when a sentence doesn't start with 'While' and end with 'Wend.' He's going to reform English by taking out all the inefficient synonyms."
"But is his list any good? Couldn't we just use the semantic data structures as a reservoir of associations that the net could dip into?"
"We could. But it wouldn't be satisfying. That's not the way the cells do things."
"I thought the point wasn't to duplicate mind. I thought—"
"The point is to get this boutique of ICs to comment intelligently on William Bloody Wordsworth. Are you going to give me rules for doing that?"
I said nothing. Lentz realized he'd gone overboard. He regrouped and spoke again.
"We could try to feed it algorithms for everything. There are only slightly more of them than there are particles in the universe. It would be like building a heart molecule by molecule. And we'd still have a hell of an indexing and retrieval problem at the end. Even then, talking to such a decision tree would be like talking to a shopping list. It'd never get any smarter than a low-ranking government bureaucrat."
Lentz returned to the net design. I could do little to assist hut run upstairs to the laser and retrieve the printouts. Lentz liked to crack morose asides while he worked. "You better start thinking of how you're going to fill the rest of your year."
Sometimes he would set me tasks, puzzles that weren't clearly puzzles until after I solved them. "Fill in the blank, making a simple, two-word sentence. 'Silence. .'"
". . fears," I suggested. "Make that 'Silence fears.' "
"Oh, very good, Marcel. I knew there must have been a reason I asked you to be my apprentice."
But the more he joked, the worse Lentz's asthmatic anxiety grew. We washed up at the starting block. My future again began to seem as unbearably long as my past.
Angry, Lentz lobotomized the thing. "Just punishment," he declared. "This box didn't deserve to think." Hurting the circuit seemed no more sadistic than forcing it to learn in the first place.
He blurred the device's retention. He reduced the scope and breadth of connections. Then he tested it a last time in its weakened state, empiricism now taunting all three of us. To my astonishment, the learning algorithm rose up from its reduced straits and began to get the picture.
Miraculously, Implementation A began to pick up our patterns, arranging our words into meaningful relationships. "I might have known." Lentz sounded disgusted with himself. "Too much retention. It was learning. But learning got swamped in its own strength. The creature was driving itself batty, holding on too tenaciously to everything it had ever seen. Dying of its own nostalgia. Mired in the overacquired."
"Autistic," I remember saying. Particulars overwhelmed it. Its world consisted of this plus this plus this. Order would not striate out. Implementation A had sat paralyzed, a hoary, infantile widow in a house packed with undiscardable mementos, no more room to turn around. Overassociating, overextending, creating infinitesimal, worthless categories in which everything belonged always arid only to itself.
A few deft neurological nicks and a new run-time behavior did the job. Implementation B lived inside the same hardware, indistinguishable from its short-lived predecessor. This time, fuzzy logic and feedback-braking kept the box as discriminately forgetful as any blossomer.
From the get-go, B was a different animal.
Diana Hartrick came by. She surprised me one Saturday afternoon. The air had just started to turn chilly. I remember that she had a denim windbreaker on and was shivering on the stoop when I let her in.
"Hi. I know you're not supposed to drop in unannounced in this country. ."
"Not a problem. I lived half my adult life in a country where dropping in unannounced was de rigueur.'''
"Good. I was on my way to the lab. I thought I'd bring you this." She handed me a Ball jar full of soup. "I figured a single guy could use some. I'm afraid it's very hearty. You may want to thin it out some when you heat it up." She stopped to look around my place for the first time. "Hey! I thought you were staying until summer."
"I am."
"What's the matter? Aren't you comfortable here? It looks like you're trying to weasel past worldliness without actually having to touch it. By the way, you'll want to plug your fridge in for the soup. Unless you'd like to share it with the microorganisms."
"Thanks. You wouldn't happen to have a bowl on you, would you? Just kidding."