A large model has learned something you cannot easily read off its weights: not only which answer is right, but how nearly right all the wrong ones were. Ask it to name an animal and it does not simply say dog. It says dog, overwhelmingly, with a faint residue of wolf, a fainter one of fox, and essentially nothing for teapot. That residue is the model’s accumulated sense of how the world resembles itself. Distillation is the practice of training a smaller model on those full distributions rather than on the bare correct answers — letting the student inherit the teacher’s judgement instead of merely its conclusions.
The obvious objection is that a copy should be worse than the original. It usually is — but by far less than anyone expected, and at a small fraction of the cost. This is where the trouble starts. A model that took hundreds of millions of dollars and a data centre to train can be approximated, well enough for most purposes, by a competitor who simply asks it a great many questions and trains on the answers. The technique does not care whether the teacher consented. It is also lossy in ways that are hard to see: the student inherits the teacher’s blind spots along with its judgement, and inherits them stripped of the training data that would let anyone work out where they came from.
So the practice sits in two places at once. Inside a laboratory it is ordinary and useful: you train one very large model carefully, then distil it into something small enough to run on a phone, cheap enough to serve at scale, fast enough to answer while the reader is still paying attention. Almost every model you actually interact with is smaller than the best model its maker owns, and distillation is much of why that is affordable. Outside the laboratory it is a governance problem nobody has solved — terms of service forbid it, detecting it is difficult, and proving it in court is harder still.
Strip out the machinery and this is apprenticeship, with the same old asymmetry. The master’s skill took a career to acquire; the apprentice acquires a working version of it in a few years, never having made the mistakes that produced it. Something real transfers. Something also fails to. What the apprentice ends up with is not the master’s knowledge but a compression of it — reliable in the ordinary cases, quietly unreliable at exactly the edges the master would have recognised on sight. We have argued for centuries about whether that counts as theft or simply as how knowledge has always moved. The machines have only made the argument urgent.
