AGI Trilogy

There is no proof that a thing is your own. There is only evidence that it is someone else's.

My profession stands on that asymmetry. I examine the artificial-intelligence models submitted to the state, to see whether they are truly their makers' own, from the ground up. My official title is long and dull, so I will skip it. Around the office they just say Verification. Send it over to Verification. Verification flagged it. I am the suspicion the state hired.

No one taught me this work. Three years ago the job did not exist in the world. There was no predecessor and no handover file. On my first morning, the desk held a computer and one empty approval binder, in a format I could not trace to any ministry. I began the job by filling in those blanks. Half of the verification manual my juniors use today is what I made up at night back then. I wrote it in formal prose, so the procedures would look like they had a lineage. Perhaps that is how lineages begin.

Before this, I was on the building side myself. Seven years on a mid-sized team, doing about what everyone else did. I could give several reasons for quitting, but only one of them is honest. One day I found I could no longer say where, in what I made, my own part began and ended. When I quit building, the question was all that remained, and just then the country began paying a salary for it.

The official name of our program contains the words national team.

The first time I heard it I laughed, and three years on I no longer laugh, because the name is exact. This country decided to hold the arrival of artificial intelligence as a tryout. There are preliminaries, there are finals, there are elimination rounds, there is an announcement date. There is screening, there is pleading, there are objections. Other countries waited for the arrival, or built it. We drew up a bracket for it. A colleague I met at a conference abroad found this fascinating. The state raises models by tournament? I have not yet ruled on whether that is an embarrassment or a boast. We are the best country in the world at turning anything into a selection process. College entrance, the civil-service exams, corporate recruiting, television auditions. There was no reason the arrival should be the exception.

My post is the threshold of that bracket. Every model that reaches the finals crosses my desk.

When a model comes in, the first thing I do is chew it.

Not a metaphor — the name of a procedure. I wrote it into the manual myself: the mastication test. You apply compression, shake it minutely, resect it layer by layer and watch the response. A model that truly grew from the bottom has the same texture wherever you bite. That is what training is. The whole of it passes through the same stretch of time together, and its parts leave tooth-marks on one another. A transplant is different. At the boundary, the teeth slip. However much new flesh is grafted on top, a layer that did not hurt with the rest — you know it the moment you chew.

In October, on that team's submission, my teeth slipped.

A big team. The kind everyone knows when you say the name. They had hidden nothing, either. The submission papers stated that the encoder of an open model from across the sea had been referenced. In the industry, that sentence is a form of greeting. Everyone writes it, everyone reads it and moves on. What was not written were two things. How far the use went. And whether those layers had afterward hurt together with the rest.

I hung on for three days, teeth in. One entire vision encoder was somebody else's, and it was frozen. Detached, carried over, planted still frozen. A frozen layer does not chew. It carries no tooth-marks of training. Where the papers said meat, I bit, and metal came out. What I confirmed was not the use. The use was on paper. What I confirmed was the boundary. Boundaries are not written into papers. You only learn them by biting.

The report was four pages. Those four pages eliminated two years of several hundred people. The day it went up for approval I left work early and walked a long time. The sense of having caught something and the sense of having bitten something are different. What you catch stays in the hand; what you bite stays in the mouth. I am still holding that team's arrowhead in my teeth.

There is a place called the pleading room. It has an official name of its own, but that is what we call it among ourselves. The room where an eliminated team comes and speaks.

From that team, three came, and the talking was done by the youngest researcher. Most junior on the org chart — and probably the hands that had planted the encoder. The kid did not make excuses. The kid did a calculation. Numbers marched down the whiteboard. With the compute we were allocated, from a blank slate, by that deadline — the score does not physically exist. The people in this room know that better than anyone. We didn't steal time. We borrowed it. We meant to pay it back.

I know. Before the kid came, I had run the same calculation. The numbers do not come out. What I had drawn was, before it was the boundary of misconduct, a waterline. The height of this country's water. The bracket and the announcement date were set from above; an honest route that fit the schedule never existed in the first place; and the kid had simply chosen a route that existed over one that did not. The rules fitted the team with stocks. A first offender's stocks, the kind that corrects the walk. A re-entry clause was attached, so it was not a death. On the way out, the kid gave me a slight bow. A greeting without resentment, one profession to another. What I should have answered that greeting with, I do not know to this day.

It was around then that I went to the dentist. My temples ached in the mornings and one molar had gone sensitive.

Do you grind your teeth? the dentist asked. No. That was my answer.

You grind. Have for a long time. The dentist showed me the wear facets on my molars in the mirror, and at the word wear I had already conceded. Wear is my specialty. The dentist had me bite down on a strip of blue paper. When I bit and let go, every spot the teeth had touched stayed blue. Articulating paper, it was called. This is how we see where you bite hardest. What touches, remains. That thin strip of paper was more honest than any log I had seen that week.

The dentist said we should take an impression. I held a spoonful of cold silicone in my mouth and waited three minutes, and three minutes without speech is longer than you would think, so thoughts came. I am taking an impression, here, now. A man who catches other people's copying by day is having a rubbing taken of his own teeth for the night. It was the only copy I approved that week. Copying what is yours is legal under any law. The problem is an impression taken from what is someone else's. Then how far down is the thing we call from-the-ground-up actually one's own — before I got there, the dentist took the tray out of my mouth.

A week later the object came back. Transparent, hard, the exact negative of my lower teeth. It was called a night guard.

There is something people get wrong about night guards. It is not a device that stops the grinding.

It cannot stop it. The biting force belongs not to the night but to the day, and the day is outside this object's jurisdiction. The dentist said as much. We can't remove the cause — stress is not my department. What this object does is one thing. It slips itself between tooth and tooth, so that what gets ground down is not the teeth. Which is why it is a consumable. Use it well and it wears well, and as much as it wears, that much tooth remains.

That summer, the cafeteria television was always showing the same building. Not ours — the courthouse. The sound was off; only the captions ran. Appearance, summons, sentencing, dismissal, appeal. At lunch one day, mid-meal, no guard in my mouth, I bit down on the thought.

The courts are the night guard this country bites down on every night.

A trial cannot stop the grinding. The biting force comes from outside the courtroom. What a trial does is slip a material between tooth and tooth — the material called procedure. Filing, pleading, ruling, objection — so that what gets ground down is not live human flesh. So that the material wears instead. So if that building is on the evening news every day, it means this country bites hard every night. That is bad news. It is better news than biting with bare teeth. After that day the captions read a little differently. What was being ground down, who was doing the biting — I still did not know. But why that object got bitten night after night — that, I thought I understood.

Three months later, at the checkup, the dentist held my guard up to the fluorescent light. The spot over the right molars had been ground cloudy. You're grinding a lot — has it been a hard stretch? The dentist was reading my nights. About the days, I said nothing. Only what is ground remains. Thinking I had heard that grammar somewhere before, I ordered a new guard.

At one objection hearing, I was for once the one being bitten.

The eliminated team's counsel came prepared. He did not contest the misconduct. He contested my scoresheet. Commissioner, these benchmarks you used in the verification. Which country made them? Who verified the independence of these metrics? A national team selected by measuring ground-up models with someone else's ruler — which nation's team is that?

Procedurally, the question was dismissed as irrelevant to the merits of the screening. I could not dismiss it. He is right. Our graders are all imports. I was measuring other people's traces with other people's rulers, to establish something of our own. Nowhere in the program plan is there a line for making rulers. A ruler is a thing that simply exists — that is what we were all taught. On the subway home that evening I noticed the force in my molars and let it go. The days of daytime biting were beginning. The guard is for nights.

The week the benchmark cluster was moving into the basement of the evaluation building, I ran into an old man in coveralls in the freight elevator, two days running.

He was a cooling-pipe man. Busy season, I suppose? I asked, meaning nothing by it, and the old man thought a moment and answered. Down below, yes. Banks, hospitals, law firms. These days everything moves in through the basement. Regardless of the picking and choosing up top. The elevator opened, and the old man shouldered his tool bag, added one more sentence, and disappeared into the dark on the pipe side. Heat doesn't lie, you see.

I stood there a long while. Upstairs, the state was selecting the arrival; down in the basement of the same building, the arrival was already coming in through the pipes. The announcement date belongs to the upper floors. The dates belong to the basement. Though the two of them share an elevator. That an arrival is not a thing you select — I learned that in a freight elevator.

In the second round of evaluation, one model was strange.

Not because something foreign turned up. The opposite. Wherever I bit, the teeth went in. The texture was uniform. But the arrangement was unfamiliar. If it was an impression, there has to be an original, and no original I know carries this pattern. If it began from a blank slate, this level of finish does not come out of this much compute. I chewed for three days and then admitted it. There are objects my teeth cannot rule on.

Some evaluator across the sea would have made a folder. A folder named Hold, with three or so cases inside. I cannot make a folder. The verdict field in my system has two boxes. Fit. Unfit. Hold is not on the form. I built this form too, three years ago. Back then I did not know anything strange enough to need holding would ever come, and above all, this country's examination halls had taught me that if a hold box exists, everyone flees into it — so I dug only the two boxes, on purpose. The trap I dug three years ago is the trap I am standing in now.

The deadline is two days off. For two days now, the cursor has stood between the two boxes.

Some nights, near dawn, I get the order of things confused.

Is the state chewing up the fakes, spitting them out, establishing the real thing? Or is this tryout itself a foreign body lodged in the country's jaw, and some larger jaw slowly chewing on that? By day I work on the first sentence; near dawn the second one comes. Either way the motion is the same, and by sound alone there is no telling them apart. Maybe, from inside a chewing country, you can never know, right to the end, whether you are the teeth or the meat.

My wife said, one night near dawn: You've started grinding your teeth in your sleep. It sounds like you're trying to bite through something.

Tonight, again, I set the transparent thing between my teeth and lie down. The third-round notice will post in the morning. The cursor stands between two boxes, below ground the water is circulating, and the courthouse steps will be crowded again tomorrow. The sound of chewing will not stop. I cannot write that it is all right. But this object has taught me that unscathed is not the same as whole. What is ground down and still remains. When I wake in the morning my jaw aches, and my teeth are all in their places.