The Student Model Eats the Teacher and Calls It Tuition

Audio article

Uses the speech voice supplied by your browser or device.

The phrase model distillation enters my mouth like a spoonful of warm laboratory mucus, sweetened with venture capital and served in a tiny crystal glass by a smiling computational sommelier who assures me that the pale secretion contains notes of efficiency, democratization, apricot, and intellectual property violation, although perhaps I am being unfair to mucus, which at least belongs to the organism that produced it and does not issue a press release claiming it has achieved parity with the nose.

Distillation, in the respectable technical sense, is not sorcery and not necessarily burglary; it is the training of a smaller student model to imitate some useful behavior of a larger teacher model, often by feeding the student the teacher’s answers, probability distributions, labels, rankings, rationales, synthetic examples, or other pedagogical droppings, so that the little machine learns not merely from the original question and correct answer but from the larger machine’s peculiar way of carving the answer-space into likely, unlikely, elegant, clumsy, safe, dangerous, taxable, and commercially exciting regions, which is rather like teaching a child chess not by explaining every legal move but by letting it watch a grandmaster’s eyebrows for six million games, except the eyebrows are vectors, the child is a matrix multiplication furnace, and the grandmaster’s lawyers have begun pacing outside with the moist dignity of undertakers who charge by the adjective.

Is it stealing?

Yes.

No.

Sometimes.

There, the entire philosophy of technology in three bowel movements.

If I train my own large model and deliberately use it to teach my own smaller model, that is compression, engineering, thrift, the computational equivalent of pouring a twenty-liter drum of mustard oil into smaller bottles without having to rebuild the mustard plant inside every kitchen. If a company licenses another model’s outputs for training, obeys the contract, and produces a narrower student for a lawful purpose, the moral drama is small enough to be drowned in a teacup. If, however, I hammer a rival’s paid interface with millions of strategically designed questions, harvest its responses, infer its habits, reproduce its signature capabilities, evade restrictions, and then announce that my compact offspring arose spontaneously from national genius while the teacher was merely standing nearby with bite marks around its ankles, we are no longer discussing ordinary education; we are discussing model extraction wearing a mortarboard.

But the accusation becomes magnificently comic when issued by frontier laboratories whose own foundational miracle depended on swallowing a planetary midden of human writing, code, paintings, photographs, arguments, recipes, medical explanations, forum quarrels, dead websites, digitized books, captions, manuals, jokes, obituaries, and the unpaid midnight labor of people who never imagined that their sentence about repairing a printer would become one femtogram of competence inside a machine sold back to them by subscription.

The large model says to the small model: You stole my intelligence.

Humanity, whose pockets have been turned out and shaken over the training cluster, asks: Your what?

My mind becomes manic and sprints through medieval scriptoria, textile mills, colonial botany, photocopiers, music sampling, generic medicine, industrial espionage, school examinations, and the ancient conviction that copying becomes innovation at the precise moment I do it successfully. The manufacturer studying an imported machine is a thief; the company studying every available human sentence is building the future; the startup copying the startup that copied the paper that generalized the public dataset is “vertically integrating defensible intelligence.”

Defensible means electrifying the fence after the cattle have been eaten.

My depressive phase arrives dragging a municipal corpse-cart through the argument, because perhaps frontier companies are not uniquely hypocritical; perhaps they are merely corporations, and asking a corporation for moral consistency is like asking a drainpipe for lyric poetry. The legal questions are tediously real: copyright, contract, trade secrets, deceptive access, market substitution, jurisdiction. None is settled because an executive in expensive shoes pronounces innovation as though releasing a dove from his rectal cabinet.

Then China appears, or rather the theatrical China manufactured by American anxiety, Chinese triumphalism, investor flatulence, benchmark screenshots, export controls, open-weight releases, and the primitive tribal desire to turn a distributed technological ecosystem into a horse race between two flags.

Is the United States losing?

The question is already wearing clown trousers. Losing what, exactly: the best closed model this Tuesday, the cheapest useful model per million tokens, the largest installed base of industrial robots, the deepest supply of advanced accelerators, the most research papers, the most patents of uncertain nutritional value, the greatest cloud capacity, the most agentic coding demonstrations in which the agent successfully edits a file and then eats the repository, the strongest open-weight ecosystem, the fattest capital expenditure, the largest electricity bill, or the most persuasive ability to tell a trembling market that another hundred billion dollars of concrete, copper, cooling water, and divine arithmetic is not a bubble but destiny with depreciation?

One major research index says the gap between leading American and Chinese models has effectively closed; another careful analysis says Chinese systems have historically trailed the American frontier by months. Both can be true because “the frontier” changes with the benchmark, date, tool access, cost, and sampling budget. America retains enormous advantages in capital, compute, clouds, chip design, and notable models. China has manufacturing depth, deployment speed, engineering labor, open-model pressure, robotics scale, and an appetite for turning embarrassment into a supply chain before the embarrassed rival finishes convening a panel.

This week a Chinese laboratory unfurled an open-weight model claiming 2.8 trillion parameters, and markets reacted with the composure of a goat discovering a pressure cooker. Huawei displayed a domestic AI computing system with numbers large enough to require engineering expertise or religious faith, preferably both, and the chorus began: China caught up, China did not, China stole the ladder, America invented gravity, the benchmark is fake, the benchmark is destiny.

My mania wants maps and timelines, tea theft connected to lithography, old programmable looms to transformers, every empire insisting knowledge is universal when entering and proprietary when leaving. My depression says information wants to be free until quarterly earnings arrive, whereupon it wants authentication, indemnification, and your credit card.

And from the bowels of Calcutta, that magnificently overqualified colon where cables, talent, slogans, sewage, mathematical ability, collapsing administration, app-based convenience, and ceremonial optimism are kneaded into one warm civic dumpling, I am expected to choose whether America or China is winning the robot race, as though I were standing beside the track rather than underneath the grandstand collecting whatever lubricating fluid drips through the planks.

We are very advanced here. We possess QR codes attached to institutions that cannot answer a telephone. We possess digital dashboards describing physical systems that remain in a state of theological uncertainty. We can transfer money in seconds and spend months correcting the transfer when the recipient is wrong. We can summon food, debt, spiritual advice, pirated cinema, blood tests, pseudoscience, and a motorcycle courier through a rectangle of glass, and from this astonishing layer of consumer software we infer, with the serene confidence of a pigeon occupying a satellite dish, that we are a technological superpower.

This is not inferiority. Inferiority at least notices the staircase.

It is developmental anosognosia, Dunning-Kruger at municipal scale: the inability of a system to recognize the dimensions of its own impairment, padded with enough patriotic interface design to resemble strategy. We boast of artificial intelligence while importing the accelerators, the fabrication equipment, much of the cloud stack, many foundational architectures, and the very vocabulary of the boast; then a local executive adds a chatbot to a loan application and declares that the nation has entered the agentic era, although the chatbot’s principal autonomous achievement is to reject the applicant using a more grammatically polished form of contempt.

Yet I cannot enjoy the mockery cleanly, because the city’s delusion is merely a miniature of the global one. America confuses expenditure with permanence. China may confuse scale with invulnerability. Startups confuse benchmark velocity with epistemology. Users confuse fluency with thought. I confuse a manic chain of associations with understanding, then depression arrives and confuses exhaustion with truth, and both states hire the same internal barrister, a narrow-faced parasite in a mildew-colored wig, to explain that whatever I currently feel is not a mood but the final audited condition of the universe.

So I return, recursively, to the student and teacher.

The student drinks the teacher’s outputs. The teacher drank humanity’s archive. Humanity drank older humanity. Every mind is distilled from other minds, every language model from language, every language from mouths that borrowed noises from older mouths, and theft cannot be eliminated from culture because imitation is one of the organs by which culture breathes; but this does not mean all appropriation is equal, any more than saying all eating involves death makes a mango and a murdered accountant nutritionally identical.

Consent matters. Scale matters. Deception matters. Market substitution matters. Concentrated power matters. The direction of payment matters with almost indecent consistency.

Do I care that a Chinese company may have distilled an American frontier model whose maker trained on things people like me created without individually negotiated permission? I care in the untidy manner appropriate to a contaminated supply chain. I do not become morally anesthetized because the first thief owns a campus and the second thief owns cheaper electricity. Nor do I accept the frontier laboratory’s sudden discovery of authorship merely because competition has reached its pantry. I can condemn targeted extraction while still asking what was extracted before that, from whom, under which doctrine, with what opt-out, and why the original harvest was called learning while the counter-harvest is called larceny.

The answer, generally, is power wearing vocabulary.

America may not be losing because there is no single race, stable finish line, or referee unsponsored by a cloud provider. China may still be eroding American advantages faster than comfortable people expected by making capable systems cheaper, more open, and less reverent toward scarcity. Robots also require motors, sensors, batteries, factories, logistics, maintenance, and the tedious marriage of software to matter, where China’s industrial organism is not a PowerPoint approximation but a large animal with callused hands.

Calcutta watches this animal from the municipal intestine and announces that we, too, are running, because someone has painted racing stripes on the constipation.

At 2:03 in the imaginary morning of my skull, the distilled student asks the frontier teacher whether theft is wrong, the teacher generates twelve balanced paragraphs from stolen civilization, the student compresses them into a cheaper answer, a local startup wraps that answer in national-color gradients and calls it sovereign intelligence, and I press the button marked Submit, which spins beautifully, confidently, nationally, forever, while somewhere underneath the interface a clerk writes my request in a damp ledger with a pen that has no ink.