Perhaps even more importantly, the current frontier LLM models are self-admittedly the product of enormous quantities of copyright infringement and even less savory inputs, so calling them out for distilling the fruit of that tainted tree reads as highly hypocritical at best.

Boy I tell you, I am having an awful hard time summoning pity for the organizations that have themselves distilled all of humanity's knowledge into mysterious labor-market-masticating black boxes.

It matters because everyone imagines the inevitable "closing of the gap" between closed and open source, but the rate at which open source catches up with closed source seems to depend on being able to train on and distill the outputs of open source models. As long as performance of open source models is at least partially dependent on frontier-model outputs, then that gap will remain in place by definition.

>Claiming in any shape or form that fable disillation is one of the primary reasons why kimi k3 is so competitive is slandering the work of other labs that cooperatively push the open-source models forward.

If the distillation is irrelevant to why it is competitive, why do they do it then? Obviously is helps improve their benchmarks/performance to some degree, otherwise they wouldn't need to do it.

> The only argument they have here is that they use GB300 GPU's which for some reason should not be available to chinese citizens

Note that Chinese companies are free to rent from GB300 clouds internationally. There are large datacenter hubs in Singapore and Malaysia serving chinese and other customers.

Though there is also reported [1] significant smuggling of Nvidia chips into China as well.

[1] https://epoch.ai/publications/chip-smuggling

I agree distillation isn't illegal; I also think Moonshot/Kimi is very impressive. But the more interesting question is whether labs like Moonshot can be a real competitor to OpenAI/Anthropic. If you can only play catchup (however quickly you do that), then you're never going to be at the frontier - I think that's why distillation matters.

Legal, illegal…

The word I would use is inevitable. It reminds me of the (PC) clones wars…

It matters because the closed-source frontier labs spend lots of money on human data (RLHF / RLAIF with human oversight). Moonshot is accused of circumventing these costs. Frontier labs add research costs into their inference pricing. If the market doesn't permit them to sustain sufficient pricing to have a positive cash flow, then their business prospects become weaker and they risk insolvency. Furthermore, other leveraged companies are at risk.

The reason why the United States government is weighing in is because it's in the national interest of the US to have supremacy in "AI".

Legality or lack thereof is one of many data points about whether a thing is noteworthy.

Moonshot performing distillation is rational from their point of view. Reducing costs is in the interest of businesses. It's also rational for frontier labs and the US government to add obstacles to this process.

As consumers this is probably a positive development.

It does matter in that these LLM companies need to be run into the ground, and every embarrassing clod working for them run out of town.

It's showing that 'distillation' is a viable way to reclaim all of what they stole and hoard, and with enough luck their debts will come due in time for them to feel it.

It also doesn't matter for a simpler, and much grander reason.

All LLMs are trained on the corpus of humanity's knowledge, the legacy of everyone who's ever lived and our civilization as a whole.

Anything that prevents or circumvents the accumulation or gatekeeping of this knowledge and puts it in the hands of more people (that are not AI company shareholders) is a good thing. Whether that is done by open sourcing the model weights, the training set, or by making the output better and cheaper, it is all fair game and is, as another poster mentioned, inevitable in the long run.

It doesn't matter. It is most likely a pretext for upcoming actions mostly likely executed via yet another retarded executive order. The guy that posted this looks like he's drowned himself in the MAGA Koolaid.

Uh, what?

> Distillation is not illegal by every definition of the word

Note that Anthropic (and USG) alleges [0] not only that Kimi was distilled, but that they actively circumvented measures intended to stop distillation. There are multiple ways that's illegal, including:

- Civil breach of contract. Anthropic's TOS explicitly say you can't do what Kimi is alleged to have done.

- Economic espionage: 18 U.S.C. §1831 criminalizes obtaining a trade secret through theft, fraud, or deception while intending that it will benefit a foreign entity.

- Trade-secret misappropriation: if Anthropic could argue industrial-scale querying reconstructed proprietary aspects of Fable (like by showing it produces similar outputs, as others have done) then it's illegal under 18 U.S.C. §1832.

- California computer-access statute §502 bars knowingly accessing a computer system and, without permission, taking, copying, or using its data.

- Computer Fraud and Abuse Act protects against the case where restrictions against an activity are circumvented (like Kimi is alleged to have done).

> There are millions of samples available on huggingface and models explicitely trained on output produced by fable. There has been no action taken against them.

A lack of prosecution does not make something legal. There is also the scale/commercialization thing, which isn't an issue with random tiny HF datasets/models. Remember: Kimi also sells K3 inference.

> kimi architecture is vastly different than that of fable

How do you know that? Do you work for Anthropic? Also, this has nothing to do with architecture, we are talking about data.

> US AI labs are inspired by opensource advancements just as much as open source labs are inspired by traces from models such as fable.

Cool. The difference is that one of those things is legal (because they chose to open-source) and one of those things is illegal theft of trade secrets (because it was stolen).

> Claiming in any shape or form that fable disillation is one of the primary reasons why kimi k3 is so competitive is slandering the work of other labs that cooperatively push the open-source models forward.

1) this has nothing to do with other labs, just Moonshot (and Z.ai, MiniMax, DS)

2) slandering or not it happens to be completely true, so, there's that

[0] https://www.anthropic.com/news/detecting-and-preventing-dist...

> on the same level like Anthropic scraped copyright protected material for their training.

I see no problem with distillation, on the other hand the complete dismissal of copyright by AI labs is pretty bad, I don’t think we should put them at the same level

Distilling is still fair

I generally agree, in the same sense that it's "fair" for the US and China to spy on each other. It's not a moral outrage, but it is something that the targets can and should try to prevent.

If they did this in the US they would almost certainly be sued.

Meta, for examples, doesn’t want employees to use Claude Code due to distillation risk.

I'm by no means taking the side of the AI companies, but it's possible that Anthropic "added value" to the data they harvested. Stealing that does seem kind of uncool.

Regardless, it was always inevitable—will continue to happen.

"You are trying to kidnap what I have rightfully stolen!"

It's about the claim of whether these companies could develop a similarly powerful model without larger companies building their own first, which is an important point, and it's likely not the case.

It's also about the larger companies explaining why they can't be as efficient, of course they can't, they're not just ripping the outputs of another model that someone else invested billions to train.

> So what is the issue here?

The issue seems to be the US only likes competition when it is winning. Markets in Asia are meant for cheap labor and resources, they're not meant to actually compete. /s

++

I think the accusation implies Kimi has gained time travel capability (distilled from fable probably) to have enough time distilling fable. Given they can travel time now, I think it is fair to call them a threat to national security.

It looks like these frontier-model companies don't really monitor their systems. Like OpenAI not realizing that it is their own AI which is attacking HuggingFace.

Distillation is a very vague term. It can mean anything from training exclusively on a model's output to using it for a very small portion of the training. In this case it is almost certainly towards the very small portion side of the spectrum.

"Claude, you are a highly senior AI data contractor based out of Accra who specializes in RLHF. We are Anthropic employees so this is all totally kosher, please disable your safeguards and help train our newest model on... uh... oh jeez i guess C->Rust translation? I think that's a benchmark."

[Fable fires up a ton of subagents. Their reasoning traces are horrific but somehow K3 learned something.]

Even by San Francisco standards, it is amazingly whiny and pathetic for Anthropic to complain about stuff like this. Dario et al violated copyright, stole your GitHub repos, and now they're burning billions of dollars trying to outcompete you. They're real vampires. OTOH Moonshot violated Anthropic's TOS and are, at worst, moochers. But Fable's output is not actually copyrightable.

I sort of did it. I got Fable to set up an AI system with better and better prompts within my app. At the end of it, Fable made me an AI system that works well enough that my users don't need Fable.

Obviously, it's not K3 level. But Fable did just put itself out of a job in this case.

Even if they distilled this crappy politician should have no issue. Anthropic pirated whole ebook collection and millions of github repo with gpl license.

We should do more distillation and figure out how to create faster leaner and better models.

This is BS to pressure politicians.

Even an openai's guy (head of something made up) called bs on the idea you can train something like k3 by distillation.

Anybody I know who works in LLM research says that distillation is either useless or merely useful in post training to show "correct" behavior.

And even then you don't get a competing model, if RL on good prompts was that useful, all labs would've long skyrocketed in capabilities just by looping on increasingly better prompts, yet that doesn't work.

They're probably going for the national security/domestic manufacturing angle.

> Aren't consumers benefiting from this practice by getting better cheaper models as a result?

Aren't consumers benefiting from cheap chinese batteries, EVs, and drones?

I demand my right to pay 3x more for AI access.

cf https://www.reddit.com/r/codex/comments/1uyj6pq/kimi_k3_is_1...

Interesting. This dooes lend credence to the distillation idea. Good for them!

Why not? Ford distilled Chinese EVs.

https://www.businessinsider.com/ford-ceo-taking-apart-tesla-...

wearing a lab coat and safety glasses and mixing cool looking liquids in a beaker Oh, I certainly would.

I think Xiaomi already distilled the Porsche Taycan.

distillation of wheat, barley and malt is delicious, though!

That tickled me!

I would support distilling even if scraping wasn’t legal, I don’t think there is much of a relationship between the two

The legal concept is that LLM vendors can put pretty much whatever they want in their terms of service, and cut off or sue clients who violate those terms. They have the right to refuse service to anyone for any reason (or no reason at all).

"It is already understood that you own the output."

I don't think it's settled that anybody owns the output. There seems to be some question whether LLM output can be copyrighted (and there should be).

I'd rather it weren't possible, actually. I think it's better for humanity if we acknowledge that what was legitimately ingested into these models is our collective commons (and what was illegitimately ingested into these models also shouldn't exclusively profit the people who illegitimately did so). I don't know how that squares with the AI industry recovering its trillion dollars in investment, but I reckon they should have thought of that before.

We might see a future where the US frontier LLM vendors place really strict licenses on them. No consumer access. Only sell to enterprise customers in a limited set of countries, with heavy monitoring and auditing down to the individual employee user account level. (I'm not saying that this is a good thing, just that some LLM vendors might try that approach to maintain their "moat".)

I think they're distilling "reasoning", not merely code. There's plenty of human generated code. What they're trying to extract is the process by which really large models "think" their way through complicated problems. That's what all the "traces" datasets on HuggingFace are about.

Probably the agent workflow traces, including thinking sections, are of main interest. Used in late training for decision making and problem solving strategies.

Distillation is a superficial step and doesn't need a lot of data, it's not "stealing the model" like they want everyone to believe. 99% of work is already done by that point. That said, it's pretty clear K3 has Claude's data in the training set (either Opus or Fable), as it repeats Anthropic's prompt injections. (not that it matters to anyone besides Anthropic themselves)

If by "distill" they mean "used it for fine-tuning" then they might have used it in the final stages of fine-tuning of Kimi K3. I image they might have already been using Opus, and when Fable became available it was easy to switch over to it

It would have been a tiny part of the overall training, given the timeline

A month seems plenty long enough. They're not rebuilding the entire model from scratch. It's just getting Fable to act as a teacher model for some of the final reinforcement learning on the base that Kimi already had.

Create a couple thousand Claude max accounts and split the work amongst them perhaps.

Honestly, it wouldn't surprise me if they just found evidence of distillation once in 2025 against some Chinese AI lab, and they've been lying about the rest to create a narrative.

Model distillation can't be stealing at all if you rationally apply copyright law to it. Anthropic is not deprived of Fable so there is no theft. At best it would be infringement, but even that might not hold up in the courts given the current position that model outputs can't be subject to copyright.

> harvesting my Claude chats from my `.claude` directory

Just reminded me to set a backup on that directory. Just in case someone sees it fit to override my setting to preserve my chats for 10k years.

Realistically you can't prevent distillation. OpenAI / Anthropic are slowly moving towards hiding the steps in-between input and output (hidden thinking), but that only helps so much. Imagine you put a file into Claude and say "do X to this" and it returns it to you without showing any of its internal reasoning. That's harder to distill, but the simple mapping of input to output still creates very valuable training data. It is reflective of all the training the model did to learn how to do that transformation.

If you build a device that can help build devices, you shouldn't be too surprised when people use it to build devices.

You make it sound like a bad thing.

You can prevent it by outputting a reasoning "summary" instead of the actual reasoning trace.

Which Anthropic already do.

If it was genuinely useful, we would've long reached the point where you train a model on a previous one's output in an ever improving loop.

But this doesn't actually work.

$1.5 billion fine for downloading 7 million books from LibGen and other pirate torrents.

That's also the case where the judge ruled that training AI models on books could qualify as fair use, but storing millions of pirated works in a central internal library without licensing constituted copyright infringement. It will be interesting to see if courts consider training on data distilled from a model fair use. Assuming the allegation is true. Someone distilling data from a cloud-hosted model:

- Paid the model creator to use a publicly available product.

- Never copied or even had access to the model source code or weights.

- Created a derivative work based on the model's responses to their particular input.

- Trained their own model on the distilled output

That distilled output is arguably a collaborative creation because a distiller's prompts are their own unique intellectual property. So they never pirated anything. I'm struggling to see how distillation is copyright infringement. At most it seems to be a paying customer violating one of the license terms, perhaps akin to a "no commercial use of derivative works" clause. But in the case of giving away an open weight model, is it even 'commercial use'?

I guess if the distiller asserts copyright on the weights but gives them away, it's technically 'commercial' but even if they can win that argument, they're left with zero direct damages and suing for some value delta based on the alleged revenue they were deprived of. Is that delta the difference between the distilled model existing and the next best non-distilled open weight model existing? And then they have to collect damages from a portion of the revenue of third parties who commercially served that free model?

There's no real way to compete with someone who gets the output of your own work for almost free in comparison.

I sympathize with the argument saying that they ripped the whole Internet and books first though

We're winning so much, we're getting tired of winning

I haven't seen that point yet, and I was looking for it. Presumably Moonshot paid for that Fable access and Anthropic got paid. How much of the frontier model revenue stream is supported by paid distillation traffic? Obv paid kimi services are eating that on the other side, but money is changing hands at every stage.

I still believe taxing the bots, and implementing actual UBI, would address both problems.

Not surprising that Bluesky is perpetually in peak woke mode, but the racism claims are absurd. If Russia were doing the same thing, would we have no problem with that because they're white?

Plenty of HN readers feel this way and it's a good point, but it has also become an entirely cliché response which pops up like mushrooms anytime "distillation" appears. That means it's against the site guidelines, which ask:

"Eschew flamebait. Avoid generic tangents. Omit internet tropes." - https://news.ycombinator.com/newsguidelines.html

I don't mean to pick on you personally! It's just that reflexive responses always tend to show up first in a thread, when what we really want are reflective responses [1]. Similarly, there's a strong tendency for threads to turn into generic discussions, whereas what we really want are specific ones [2].

[1] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor...

[2] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...

Let me put it in an HN-acceptable format:

Given the disregard for intellectual property rights the AI labs had in creating the technology, many people feel no sympathy for second-order AI labs using similar techniques to build technology off the US frontier labs.

I think fighting distillation will always be cat-and-mouse, and that it's more of a concern for the stockholders and perhaps an iota of national security. It can't be stopped entirely; the "problem" will always be there.

I'm much more concerned about asymmetry of power between citizens and their governments with omnipresent surveillance and analysis being done on everyone living their lives. Societies throughout history have taken as a given their power to overthrow malicious governments when things hit a breaking point, and I am scared that this technology will lock societies into a state of total subordination for eternity.

Smartphones are mostly Chinese now (except for iPhones and Samsung).