Distillation Is a Mirror; the US Labs Just Hate What They See
🍕 I’m JA Westenberg. I write essays about technology, philosophy, productivity, and what it means to be human. If you like ideas that take their time, you’re in the right place.
While you’re here, a few things I’ve built: Kerouac, a writing app | Post Through It, a solo microblogging tool | Distributism.ai, my MCP platform | Laterware, all the AI news you need | Studio Self, a thought leadership agency for founders x VCs.
You can buy my book Permissionless here.
The US AI companies are screaming about distillation.
(Again.) To OpenAI and Anthropic, the idea of an open-source or Chinese AI company building their own AI by querying Claude, ChatGPT, etc is abhorrent - it’s a form of theft, taking that which doth not belong to them.
But there’s a deep hypocrisy at work here, and I don’t think it takes a highly developed moral compass to find it.
The US frontier companies built their own models on the entire corpus of human knowledge; they trained on books, images, music, videos, software and so on - entirely without the consent of the creators.
It’s the original sin of artificial intelligence. I won’t go so far as to say they stole these works - although in the case of torrented and pirated books, I think that case makes itself - but they certainly didn’t get permission from any of the folks whose work now constitutes Claude et al.
We weren’t allowed to opt out.
And in most cases, we certainly weren’t given compensation for our work.
In the vast scheme of things, this doesn’t actually keep me up at night. I’ve been writing on the internet long enough to know that my own work has been chewed up and digested by every model out there. I’ve understood that since the first models were able to do a passable job at generating a JA Westenberg-style blog post from first principles.
I remain largely unbothered by it.
Most of what I produce - including this blog - I consider to be open source anyway.
Reproduce it however you like; send me a link for my archives.
My problem is the moral inconsistency.
Distillation is when a smaller AI model learns from a bigger model. It’s almost copying homework; it extracts the patterns created by the frontier labs without having to copy the weights. The method enables AI models like DeepSeek to be built - and therefore, provided - at a fraction of the cost of US models.
According to the luminaries running Anthropic and OpenAI, this constitutes theft, inasmuch as it’s the unauthorized use of intellectual property.
But if Anthropic has extracted value from someone else’s work without permission, and DeepSeek has extracted value from Anthropic’s work without permission, I fail to see which sinner has acted more egregiously.
Both have used the output without direct permission; both claim:” Transformative use” as a defense. The logic is identical. If, as a writer, Anthropic has trained on my work without my consent, I don’t see that I should be at all invested in protecting Anthropic’s rights if and when Moonshot (the makers of Kimi) train on Anthropic’s work.
It seems a fairly clear-cut case of what goes around comes around.
Anthropic, of course, has teams of lawyers, lobbyists, and flaks. They have an army of folks whom they can call on to protect their IP. The individual artists, creators, etc whose work wound up in their training data have no such advantages.
They have every right to attempt to use that army to fight for their commercial moat, but they’re playing a very different game than the rest of us. They’re not the victims here, and pretending to be so when their own logic is turned on them seems a bit rich.
A bully who goes crying when someone else punches back is hardly likely to garner sympathy from anyone but the bully’s own overindulgent mother.
If training data is fair use, it must logically follow that distillation is fair use; and if distillation is theft, it must follow that training is also theft. You can’t have your cake and eat it too. You have to pick one. And the truth is, they already picked theft, by their own definition of the word.
Nobody in AI has clean hands; everyone is tainted to some degree by the original sin of training data. The entire stack is built on borrowed ground, and arguing about distillation won’t change that fact. I don’t think you need to abandon AI entirely because of it, but you do have to take the indignation of the injured AI labs with more than a grain of salt.
Anthropic and OpenAI forfeited the moral argument at the inception of their commercial models. They may be able to make a compelling legal case to the US Government, but they have no ethical case they can make to the rest of us that won’t show them up to be sore losers who can’t handle their own medicine.
Distillation is a mirror.
Just because the US labs hate what they see in it doesn’t mean the rest of us should weep tears of sympathy.