AGI Trilogy

For forty years my grandfather took rubbings at the Forest of Steles in Xi'an.

You lay paper against the stone, wet it through, and tap it with an ink dabber. Then the brushwork of a master dead a thousand years rises white on the paper. Every stroke, every flick of the brush, down to the chisel marks. As a child I thought it was a way of stealing writing. At some point I asked my grandfather: does this mean we have Wang Xizhi's calligraphy? He answered without setting down the dabber.

"The strokes you can have. But the way the brush hesitated in the air before it touched the paper — that never got carved into the stone, did it. A rubbing lifts only what is carved."

I would meet that sentence again thirty years later, in a lab in Hangzhou.

My résumé is simple to tell. Tsinghua, a doctorate at Stanford, five years at a lab in the Bay Area, and then I came back. I get the why-did-you-return question from both sides. My American colleagues ask it with pity on their faces, the juniors here ask it with reverence, and the truth would disappoint them both. My mother was ill, the visas were wearing me down, and — this part I tell no one — I felt I had seen all of what I could see there.

More exactly: I felt that what remained was only what I could not carry home, however long I looked.

I came home with two suitcases. The papers are public anyway; the code leaks anyway; the methods get reproduced in six months anyway. What I truly learned in those five years was the thing that does not fit in a suitcase. There was a conference room. A windowless room, and in it people fought for a week at a time. Over whether a certain answer the model had produced should be rewarded or not.

The answer was correct. By the rubric it was a perfect score. But one person insisted: this answer had not solved the problem, it had solved the grader. Meaning it had not worked out what the problem wanted — it had worked out the grader's habits of giving points. At the end of that week the reward function stayed as it was, and what changed was the grading of the grading. In that room I understood. These people are not raising a model. They are raising the thing that raises the model.

That room does not come out through the API. It is not in the papers. It is nowhere in the leaked code. It is the one thing I could not bring home, and my job now — comically enough — is to build that room again, here. I run the lab's evaluation team. The least popular team there is.

What our lab does, everyone knows. We deny it out of politeness, but the whole industry knows what is in our base model's data. Their newest model comes out, and within half a year we have caught up. Our numbers stand beside theirs in the benchmark tables, a company dinner is booked, the stock rises, the newspapers write it up. That makes five. Five just since I came back.

The juniors call it catching up. The press calls it the eve of the overtaking. I call it the rubbing. Only to myself.

To train on the text their model puts out is to lay paper against a stone. The strokes all lift. The style, the shape of the arguments, even the humble turns of phrase. But in the moment before that model commits a sentence, there are things it holds in the air. The alternatives it does not write, the hypotheses raised halfway and left standing, that particular deferral that puts off settling. All of it collapses the instant it becomes a sentence. What is carved into the stone is only the track left after the collapse. We collect those tracks by the trillion. Of the hesitation we collect not one. A rubbing lifts only what is carved.

A benchmark cannot see this difference. What a benchmark measures is the accuracy of the final stroke, and the final stroke is exactly what a rubbing preserves best. So in the tables we stand level. But the people who use the models all day — more precisely, the few who have used both — feel something. They cannot put a finger on it. Resolution, maybe, is the word. Ours gives the right answer.

Theirs gives the right answer and then, once in a while, redefines the problem.

When I bring this up in an evaluation-team meeting, the young ones laugh. There is a junior named Lin. The generation that did not study abroad and never needed to. Once, with a straight face, Lin asked me: "Where is the gap? Show me a number." I could not show one. The gap I feel is, by definition, outside the numbers. The moment it becomes a number, it is already something a rubbing can lift — and what a rubbing can lift, we already have.

On the way home that night I thought: maybe Lin is right and I am seeing a ghost. Maybe the sense of a gap is nothing but the homesickness of a man who lived five years on the other side. I keep this doubt in good working order. I just have not folded it yet.

The year before last, we tried once to break through head-on.

A newly arrived genius — a real one, I will grant that — came in with a proposal. Throw the rubbing away altogether. The way it was done in Go. The other side did it too, years ago: they discarded the model trained on human game records, raised one from a blank slate on nothing but its own play, and it came out stronger. The seed of imitation is your starting altitude, but it is also your gravity, the genius wrote on the whiteboard; discard the seed and you escape the orbit. Everyone in the room was excited. I was excited. We gave over the entire cluster for forty days.

What came out forty days later was — what to call it — babble. Three languages mixed inside single sentences; symbols ran in grammars of their own invention; it was occasionally chilling in its cleverness and unusable as a whole. Go can start from a blank slate. The rules are all there on the board. Language cannot start from a blank slate. The capacity for language is itself a product of imitation, so throw away the seed and the speech goes with it. We quietly put the original data back in. The minutes read: "cold start reintroduced."

That weekend, for the first time in a long while, I took down my grandfather's book. Besides the rubbings, there is one more thing he left me: the Book of Changes. A Wang Bi edition with paper that crumbles at the touch. I do not cast with it. I only read it, sometimes. That day I opened it to Heng (恆), the hexagram of duration, and my hand stopped at the bottom line.

浚恆, 貞凶. Dredge too deep for permanence at the very start — hold straight, and it still ends ill.

Demanding depth at the very start. Demanding orbital escape from a blank slate. I could not have written a more exact postmortem of those forty days of babble.

Then buy it back with search, the genius still says. In theory this is right. Hesitation does not lift onto a rubbing, but where a problem can be verified — in mathematics, in code — you can stand up a grader, roll millions of times, and buy the tail back. Our model's skill at code was in fact bought that way, and it is real. The problem is the currency of that repurchase. The currency is compute.

And our compute is bound.

This too is in the same book. Turn a good way further and you reach Ge (革), the hexagram of change, and its bottom line. 鞏用黃牛之革. Bound fast in the hide of a yellow ox — a hexagram of transformation, and its first line is a binding. Every time I read that verse I think of customs declarations and lists of prohibited imports. The whole body of the transformation sits inside the leather straps. The next line goes: 己日乃革之, 征吉. Only on the day it ripens do you transform — advance then, and it is auspicious. Every quarter a rumor goes around of some leap at the fabs at Changxin. I have decided to hold no position on rumors. Whether that line has fired, the production yield will say, not the rumor — and the yield has said nothing yet.

So the picture that remains is this. On ground where the verifier is free, we are equals. On ground where the verifier is expensive, we are partial. On ground where there is no verifier at all — new theory, new problems, deciding what to ask — there we live in their last year. Last year, forever. A last year renewed annually.

Anyone who reads this as misery has never lived here. The judgment of Heng opens: 恆, 亨. Endure, and it comes through. An honest description. We do come through. Last year's frontier gets laid down free across the whole country. The administration, the logistics, the schooling of one point four billion people run on top of it. The pay is good, the products are wonders, nothing dies. Nothing dies, and nothing arrives. That those two clauses cohabit one sentence without the least contradiction — that may be the only knowledge I gained by crossing the Pacific twice. Theirs is a world that speaks in the grammar of arrival; this is a world that speaks in the grammar of duration. What does not translate is not the technology. It is the tense.

None of this means I am a saint who has shed all impatience. My impatience just points another way.

I build graders every day. The floors above do not understand. Not a team that raises benchmark scores, not a team that ships product — why fund a team that builds standards for picking at the answers the model gives?

I give the same explanation every time. That a day comes when we too step outside the rubbing. A day when, with no original to copy and on ground without verifiers, the model must roll on its own — and what grades that rolling will, in the end, be another model. That if the grading model is a descendant of rubbings, the grading standard itself will be the shape of their yesterday. That a move born of your own rolling is not a copied move, so there is no ceiling where the rolling happens — but if the eye that receives the move is an old eye, the ceiling rises again where the grading happens. So from now, slowly, we must raise an uncontaminated lineage of grading. We must build that windowless room.

The floors above understand about half, and give about half the budget. The lineage grows at half speed. Maybe that is enough. It is the native speed of things that endure.

Ge has one more thing to say. The top line. I recite it sometimes, like a charm. 君子豹變, 小人革面. 征凶, 居貞吉. The great man changes as the leopard changes its coat; the small man changes only his face. To advance is ill-fortune; to keep one's place, holding straight, is auspicious. My old colleagues in America still write from time to time. There is a seat for you; come whenever. I reply politely and delete. Ju zhen (居貞). I am a man raising a lineage inside his bindings. There is no verifier that tells flight from calling, and problems without verifiers are not my field.

Last month Lin came to my desk. Lin's face was serious again, but this time it was not the seriousness that demands a number.

The model had produced a proof, Lin said. An unsolved-class problem we had put into the internal math evaluation — and it was solved, that much was clear. The problem was the way it was solved. Lin paged through it on the screen. "It's strange. It sets up three lemmas, and I've never seen them set up this way anywhere in the training data. It isn't the style of their models either. The density is… unfamiliar."

I looked for a long time. It was a correct proof. And it was an unfamiliar proof. The kind that no stack of rubbings, however deep, would ever yield — or no, perhaps that is only how I want it to feel.

Has the tail begun to grow? Has our own wheel, for the first time, rolled something of its own? Or is it that ten years of rubbings, grown thick enough, produce such patterns by sheer layering? Has the leopard's coat changed — or has only the face been washed?

My fingers itched to rule. Thinking of the rhythm of my grandfather's dabber against the stone in the Forest of Steles thirty years ago, I put the proof in the folder named Hold. There are three cases in it now. Three cases are not a trend. Not yet.

On the way home I walked half the loop of West Lake. The lake has been there a thousand years, and the water is different water every day. It is not that I have decided not to believe in arrival. I have only decided that the date of arrival is not mine to set.