smeeth

The illegality usually comes from widescale fraud which includes accessing services you are banned from accessing.

Agree, and this is exactly what Anthropic is alleging.

data you get from distillation is first and foremost not owned by anthropic nor is it copyrightable. If a user willingly gives up their anthropic reasoning data/traces that is 100% legal no matter what the "terms of service" say as it's not enforceable and would fall apart in court.

It's important to note this is NOT what happened. Anthropic was able to trace data directly back to employees at the company: "We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff."

none of the quoted laws apply here since no trade information is stolen

There is a lot of work showing Kimi models produce similar outputs to Anthropic models, which constitutes trade information. This is not dissimilar to past and ongoing IP suits against Anthropic and OpenAI by showing the models would recreate images of Mickey Mouse/NYT articles etc.

For the record, I'm a researcher myself and I'm well aware how competent the researchers are at the open-source labs/how much they've contributed. But that's not at issue here, my disagreement with you is specific to your arguments about legality; you're conflating what you think should be legal with what actually is legal.

"I think we shouldn't have IP protection at all" is a totally valid position to hold, but that's not the law is. OP said it didn't violate the law, and it does.

This is illegal in China too, there's just an enforcement asymmetry. I understand what you're saying is de facto true, I'm just taking issue with people saying either

1) its not illegal (it is)

2) it shouldn't be illegal because Anthropic stole training data (thats not how the law works)

It's a little different in this case, since 1) not all the data Ant used was stolen and 2) they did contribute significantly to the value of the stolen good.

An analogy might be a baker stole 20% of the flour used to bake their special bread, which was then stolen. Both thefts are obviously wrong and bad.

Do you care about the law or not? I think theft is bad everywhere, not just when Anthropic does it.

To be clear, I think theft is also bad when Anthropic does it.

US and CA law really don't care that Anthropic violated IP law elsewhere.

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...

You're conflating two concepts: native bit width and post-training quantization.

Consider two models: one is 16B and trained natively in 2 bits; one is 8B and trained natively in FP4. These models have the same total number of weight bits, but one has twice as many parameters. There is no real reason the FP4 one should be better just because it's FP4. It might be, but that is an empirical question, not a general rule.

Post-training quantization is another thing entirely. Taking a model trained at higher precision and forcing it down to 2 bits is going to hurt performance, often very badly. But this was never my point.

1. yes. https://www.alphaxiv.org/abs/2607.bonsai-27b table 14 shows that bonsai retains roughly 95% of the fp16 27b model's average performance and outperforms post-training quantization at a similar bit width. it doesn't directly compare against every top 10-18b model, but it is clearly still performing like a large model.

2. quantization != native low precision training. a model trained in native ternary should generally outperform a full-precision model quantized after the fact.

even if a ternary model only retains 90-95% of the performance of its fp16 equivalent, who cares? if a 200b ternary model retains most of the capability of the 200b fp16 model while using a fraction of the memory and bandwidth, it can be substantially less efficient per parameter and still dominate a smaller fp16 model under the same hardware budget.

Probably won't have to wait that long. Prism released Bonsai 27B (https://huggingface.co/prism-ml/Ternary-Bonsai-27B-mlx-2bit) as a ternary model a few days ago, its just ~7GB and runs at 44+ t/sec on an m4 max laptop. That's already in the ballpark of active parameter count of most 200B+ models, so we will get a model like this whenever Prism feels like releasing one.

It is debatable if we will actually need that many parameters though, since recursive nets like HRM (https://huggingface.co/sapientinc/HRM-Text-1B) don't need to parametrize as heavily.

The sooner the USG figures out a standard process for approving releases the better. There are many differing opinions on how much to regulate AI, but I think we can all agree ad-hoc policy sucks.

Tim was a great CEO.

I'm just pointing out product velocity slowed. I'm far from the first person to say it, it's just a fact. In the five years before Cook we got first generation Apple TV, iPhone, iPad, and MacBook Air. Your list spans 14 years.

I'm quite curious what Tim Cook's legacy will end up being.

There is no question many of Apple's business experienced significant, impressive growth during his tenure. Amazing capital efficiency.

There is also no question Apple lost product velocity. Few new products were launched, and those that were had mixed success.

Tim was, at the end of the day, an elite financial operator. Apple shareholders were lucky to have him. Customers like myself probably have mixed opinions, and it remains to be seen how he set the company up for the future.

The "rent" in "rent-seeking" does not refer to "rent" it refers to "economic rent."

Totally different concept. But don't take my word for it:

"Rent-seeking" is an attempt to obtain economic rent (i.e., the portion of income paid to a factor of production in excess of what is needed to keep it employed in its current use) by manipulating the social or political environment in which economic activities occur, rather than by creating new wealth.[0]

In economics, economic rent is any payment to the owner of a factor of production in excess of the costs needed to bring that factor into production. [1]

[0] https://en.wikipedia.org/wiki/Rent-seeking

[1] https://en.wikipedia.org/wiki/Economic_rent

Methinks this post conflates “rent seeking” with “return on investment” just a tad.

Economic rent is the extra money you can charge for owning a scarce resource. ML models are not waterfront real estate, they are IP. Other people can make more models if they can/want to.

Now, whether IP should be legally protected is a totally separate question, and while we in the West tend to assume the answer is obvious geohot would certainly not be the first person to suggest broadly applying private property rights to information makes questionable sense.

I always understood this to be why Tesla started working on humanoid robots

Model A consistently outperforms Model B under identical conditions, that tells you something meaningful about the model.

Not really! Sorry to harp on this, but there are two ways one model could outperform another:

1) It adheres to your strategy better

2) It improvises

If the prompt was "maximize money, here's inspiration" improvising is fine. If the prompt was "implement the strategy," improvising is failure.

Right now you have a leaderboard; you don’t yet have a benchmark, because you can’t tell whether high P&L reflects correctness.

Just reading your description, it sounds like there are two variables:

1. Prompt adherence: how well the models follow your stated strategy

2. Decision quality: how well models do on judgment calls that aren’t explicitly in the strategy

Candidly, since you haven’t shared the strategy, there’s no way for me to evaluate either (1) or (2). A model’s performance could be coming from the quality of your strategy, the model itself, or an interaction between the two, and I can’t disentangle that from what you’ve provided.

So as presented, the benchmark is basically useless to me for evaluating models (not because it’s pointless overall, but because I can’t tell what it’s actually measuring without seeing the strategy).

I suspect this sort of thing starts to happen when UX decision-making gets decentralized. No single god-king would allow six or more different new icons; the lack of uniformity is obviously nonsensical to anybody, but not necessarily to a disorganized collective of anybodies.

If your "impossible" designs are manufactured by non-exclusive suppliers it isn't much of a moat.

…so?

A realistic stay-at-home subsidy would max out around $30k. Your proposal only meaningfully shifts incentives for the bottom income quintile. For everyone else:

- Upper-income families can already afford to choose whatever setup they want.

- Middle-income families couldn’t take it because it’d mean too steep a drop in income.

So the alternative you proposed economically benefits the bottom quintile while leaving their kids worse off. For everyone else, it probably either doesn't matter or gives them cash they don't need as much.

Do you have any evidence for that?

From what I’ve seen, the research leans the other way. For example:

Children from more advantaged families were actually more likely to view unfair distribution as unfair, while poorer children were more likely to accept it. [0]

Mother’s work hours show no link to childhood behavioral problems, it’s schedule flexibility that matters. [1]

For working-class families, more father work hours correlated with fewer behavioral problems.[2]

The idea that “well-off kids” end up with morality deficits because their parents work a lot doesn’t seem to hold up.

[0] https://onlinelibrary.wiley.com/doi/10.1111/desc.13230

[1] https://pmc.ncbi.nlm.nih.gov/articles/PMC9119633/

[2] https://pmc.ncbi.nlm.nih.gov/articles/PMC7021583/

The most obvious example is the children of addicts. It’s hard to imagine a kid is better off stuck at home with druggie parents than spending the day in daycare.

Hate to break it to you, but many kids actually do better away from their parents than with them.

It's extremely sad, but a consistent finding in early childhood education is that the children who thrive most in daycares tend to come from the least advantaged backgrounds.

So a policy of paying parents to stay home would mostly benefit kids who are already well off.

Another day, another person not getting discounted cash flow.

Models trained in 2025 don’t ship until 2026/7. That means the $3bn in 2025 training costs show up as expense now, while the revenue comes later. Treating that as a straight loss is just confused.

OAI’s projected $5bn 2025 loss is mostly training spend. If you don’t separate that out with future revenues, you’re misreading the business.

And yes, inference gross margins are positive. No idea why the author pretends they aren’t.

As far as analogies go I prefer approximate database

I've been annoyed for a while people don't use a common parameter weight/compute budget for benchmarking papers.

That said, it does make it easier to claim progress...

Their FOSS local inference service didn't go anywhere.

This isn't Anaconda, they didn't do a bait and switch to screw their core users. It isn't sinful for devs to try and earn a living.

You don't need the analogy. If you have a tool that does a job for you your capacity to do the job degrades alongside other associated skills.

Tools that do many things and tools that do a small number of things are still tools.

"do we really even need a body at all anymore?"

It's a legitimate question. What's so special about the body and why do we need to have one? Would life be better or worse without bodies?

Deep down I think everyone's answer has more to do with spirituality than anything else. There isn't a single objectively correct response.

It's worth noting this is the exact argument people used against adopting electric calculators.