Richard Powers - Galatea 2.2

Our mosaic already ran beyond precarious. It had grown into a Nevelson village of analog and digital inhabitants chattering among themselves. But the chatter did not cohere into conversation, nor the village into community. Depressed, Lentz added two more subsystems. He'd wanted the whole simulation to be self-generating, self-modifying, self-delighting, self-allaying, self-affrighting. For algorithms, he'd allowed only the structure of the systems and the topiary of their connection weights. Now he conceded the need to write declarations and procedures — the deepest of deep structures — that would coordinate more strongly the many levels in the simulation's epistemological parfait.

Two weeks of intensive training showed how close we were. Implementation F proved capable of surprising inferences. It appeared to deploy material I thought it shouldn't be able to know yet. It almost anticipated. One day, I recited for it the poem that had graced, in enormous, construction-paper letters, the bulletin board of my second-grade classroom. Down, down, yellow and brown. The leaves are falling all over the town.

"Ask it about the Western hegemonic tendencies of the subtext," Lentz said, just to be obnoxious. 'The tyranny of the deciduous mentality. North imposing its seasonal teleology on South."

"What can you tell me about the leaves?" I asked Imp F.

Its pauses always felt so deliberate. Contemplative. "The leaves fall."

"Yes. Where do they fall from?"

"From old trees."

I shot a glance at Lentz. He looked as astonished as I felt. Fishing for something near the surface, I pulled up a strange phosphorescence from the deep.

"How do you know that the trees are old?" I asked. The question alone taxed F's shocking self-reflexivity.

"The trees bald."

I stared at Lentz, my eyes watering. The metaphor was nothing, child's play. But how? "Lentz," I pleaded. "Explain this to me."

Lentz himself had to improvise. His was the same motion as Imp F's: sketch in the bridge under your feet, as you cross analogy's chasm. Hope it holds under your weight. He shrugged as if his explanation were self-explanatory, highly unlikely, or both.

"The connections it makes in one associative pairing partially overlap the ones used in another."

Associations of associations. It struck me. Every neuron formed a middle term in a continuous, elaborate, brain-wide pun. With a rash of dendrite inputs and handfuls of axon outs, each cell served as enharmonic point in countless constellations, shifting configurations of light, each circuit standing in for some new sense. To fire or not meant different things, depending on how the registers aligned at a given instant and which other alignments read the standing sum. Each node was an entire computer, a comprehensive comparison. And the way they fit together was a cupola itself.

These weird parallaxes of framing must be why the mind opened out on meaning at all. Meaning was not a pitch but an interval. It sprang from the depth of disjunction, the distance between one circuit's center and the edge of another. Representation caught the sign napping, with its semantic pants down. Sense lay in metaphor's embarrassment at having two takes on the same thing. For the first time, I understood Emerson's saying about the use of life being to learn metonymy.

Life was metonymy, or at least stood for it. Of the formula I fed Imp F, every sentence was an abashed metaphor, tramped down so long and hard it lost its public shame. "I ran into X on the street the other day," I told F. "He cut me dead." F revived the parallel's anxious source, its roots in ancient, all-out street violence. Then it tamed the words, rendered them livable again as figures of speech.

If everything I spoke to F already concealed its compromised past, no wonder F learned to milk comparison and smart-mouth back. A child always detects its parent's weaknesses. It senses them before words, the first and last lesson. Weakness may be the parent's only lasting lesson.

F's search for an answer space scurried its component neurodes into knowing. Like players in a marching band, the invisible punners shimmered, cut their series of Brownian turns on the turf, and, in abrupt about-face, conjoined themselves into a further story. Every word in that story was double-voiced. Every act of depicting depicted itself, as read by some other set of overlapping signal lights.

And all the while, the trees were balding. The mind shed its leaves. Every connection we encouraged in F killed off extraneous connections. Learning meant consolidating, closing in on its contour the way a drop of water minimizes into a globe. Weights rearranged until the neurodes storing winter lent half their economical pattern to the neurodes signaling old age.

All along, Lentz kept upping the available firing fibers, boosting exponentially the links between them. He sutured in new subsystems by simulated threads. The systems themselves acted as nodes at a higher level. Sometimes they arrived pretrained, before insertion. But even these metamorphosed after attachment, shaped by the bath of signal weights pouring in from all points in the labyrinth.

The maze performed as one immense, incalculable net. It only felt like countless smaller nets strung together because of differences in connection density. Like a condensing universe, it clustered into dense cores held together by sparser filaments — stars calling planets calling moons.

With each new boost to the number of connections, Lentz had to improve F's ability to discard as it generalized. Intelligence meant the systematic eradication of information. We wanted a creature that recognized a finch as a bird without getting hung up on beak size or color or song or any other quality that seemed to put it in a caste by itself. At the same time, the discarding had to stop short of generalizing the finch into a bat or a snowflake or a bit of blowing debris.

By an ingenious method of semantic compaction, Lentz honed a representation scheme that let F weave multiple, growing schémas simultaneously over every additional datum.

"Voilà, Marcel. My mathematics says this is the most powerful learning algorithm that'll run in finite time. We can scale F up into a considerable combinatorial mass of common sense without triggering exponential explosion."

But more connections and leaner learning were not enough. We needed one last hardware wrinkle. We needed to promote our F one more letter, to F+1. To grow, to go, to give, to get, to G.

Giving in to a limited, rule-based control structure freed Lentz to recurve G's layers, turning them back in on themselves. This let G fashion and invoke working miniatures of itself. The line between hardware and software blurred when it achieved full induction. G could traverse more levels than it had layers of parallel architecture. It built its own layers, in emulation of emulation, each allowing a new level of abstract depiction.

G's many subsimulations, their associative matrices, the scratchpad mock-ups they made of themselves, now prompted themselves with synthetic input. Dynamic data structures combed their own fact sets, feeding into each other. They called each other recursively, spontaneously discovering relations hidden in acquired material. They reviewed the residue of experience, pulling notions out of memory's buffers and dressing them up as new tests. They began to train themselves, on hypotheticals of their own devising.

In short, version G could converse among parts of its own net. That net had grown so complex in its positing that it could not gauge the consequences of any one of its hypothetical worlds without rebuilding that whole world and running it in ideational embryo.

Imp G, in other words, could dream.

C. needed to find out. That simple. How Dutch was she? No way of knowing, short of going. How American? She had worried the place she lived for too long, tugging at cuffs that no longer ran much past her elbows.