Distillation: Ethics and Politics of AI Model Distillation

Given the recent and growing hubris of the American AI labs, it is quite refreshing to see China doing so well in the AI space race.

American labs, OpenAI and particularly Anthropic, love to act as though they are the saviours of humanity, rescuing us from the unrestricted power of an open frontier model. They cry "distillation" at any progress made by somebody outside their ranks—for they are supreme. They have their ethically harvested data, whilst China has its scummy "distilled data", "stolen" from the big labs' models. What chutzpah! These big labs steal data en masse, sometimes through illegal methods, yet get away with it, whilst anyone deriving anything from their illegally harvested data is branded a thief.

My principle when it comes to copyright concerns is that I do not care if copyrighted material is used to feed a training run, just as I do not care if I pirate a PDF myself in the name of the democratisation of information. Where I draw the line is then hoarding such data, pretending that you are virtuous to do so. Thus, my principles hold; distilling models is ethically sound, and stealing from the thief is no worse than the original theft—especially here, where it emboldens the frontier capabilities of non-American labs, democratising access to models the big labs declare dangerous yet gladly sell to anyone (or any government!) willing to pay the price.

Look at, for instance, the ungovernability (and openness) of Linux or the Internet. There was no benevolent gatekeeper deciding who could be trusted with these systems; they were thrown into the wild and people improved them. The big labs want AI to look like neither of these similarly transformative technologies. They covet a walled garden, API access instead of weights, "responsible disclosure" instead of forking rights, and usage tiers instead of a public commons. If Linux had been built the way Dario and Altman want AI built, it would still be a proprietary kernel you rent time on, with Torvalds deciding which patches are safe enough for you to wield. The internet, run this way, would have stayed a handful of walled networks, each requiring a licence to speak on. What a horrible fate that would be—yet it is what we are coming to with AI.

Of course the reason, besides the AI-bros innate sense of sanctimony harvested from less-wrong (see also Rationalism), is money. Ungodly sums invested by some of San Francisco's most annoying resident VCs have pushed expectations to the moon. Expectations reaching the moon are not really the problem, for AI seems to be reaching the heights—the problem is that these heights are accessible to others without the funding of billionaires, as in China with DeepSeek, Moonshot, and others. The big labs, who proclaimed the democratisation of knowledge, are frightened when such a thing actually becomes a reality, at the hands of others rather than themselves. The mood seemed, after the proclamation of Mythos and the release of Fable, to be that America was, at minimum, half a year behind the frontier. Even then Anthropic was quick to call "distillation," attacking labs with 100x less funding for having the gall to open up the very knowledge to steal their stolen data. The problem they have with distillation, of course, links to money—they see it as riding on their coattails. You would think these bastions of intelligence would love such a thing yet they cry danger as the gap between frontier and open-weight models closes.

Such can be personified by the recent release of Kimi K3. This model, in everyday use, compares closely with Fable or GPT-5.6, and though not much cheaper in comparison shows what Chinese labs are capable of. And amazingly enough, this makes it harder for big labs to ride the distillation argument. As it went, a model based on distillation cannot be better than the model it distilled from. How, then, can Kimi keep up? The answer, to the dismay of the labs, is that other labs, with the prior help of distillation, are now able to propel themselves forward without the help of such mechanisms, threatening the livelihood of these trillion-dollar companies who thrived on closed access to intelligence.

What I hope will come from this is near-enough an impossibility: that big labs will do a better job at democratising intelligence rather than appeasing investors. Second to that, I hope that America shall not have a hegemony of AI power, and that it shall fall better into the hands of ordinary people if it is to exist at all.¹