himata4113

We generally don't care enough to test it since it would be a significate waste of resources that's why we have models in the trillions of parameters instead of continuing to train smaller models smaller models hit a wall and generally stop improving and start overfitting and generalization starts to degrade it's less of a theorem, but an observation that has occured during training and documented here and there across many papers.

This is mostly just to reiterate myself as the original question was "Does this matter?"

Everything else is simply justifying why it shouldn't, the specifics don't really matter as there is no legal framework to stop china from continuing to distill models and anthropic has proven they cannot use software solutions to stop it either as distillation is still a problem. But I do still believe it wouldn't hold up in court either way as stopping companies from generating training data which was trained on the entire human knowledge corpus is just stealing from thieves and making it 'open' once again so the argument only gets weaker.

edit: to add, the mickey mouse / nyc was because anthropic trained on LICENSED works, not apple to oranges. The original work it was reciting was licensed and not licensed BY anthropic.

My entire point was that this was not achieved purely from distillation and claiming that is slander against open research.

I recommend reading some of their research it's honestly astonishing how intelligent some of their solutions are.

Kimi specifically relies heavily on reasoning traces which is largely due to their training strategy and will perform poorly when thrown into a conversation from another model. Another fun advancement is that they simply ctrl+c ctrl+v'd attention which means that the model can steer where to look in the context window without ever producing an output token increasing token efficiency and attention accuracy as a side effect you end up with weaker prompt adherence.

None of these 'issues' manifest in US models which proves that kimi has diverged and is achieving these capabilities seperately from the architecture that US labs rely on.

I would agree with you during the Deepseek R1 era, but US labs were heavily inspired by open research at that point as well so I wouldn't give them too much credit.

I do agree that two wrongs don't make a right, the terms of service generally gives cooperation the power to sever the contract, but it does not make things illegal in the literal sense. The illegality usually comes from widescale fraud which includes accessing services you are banned from accessing.

When I said "Does this matter?" I specially meant that distillation in itself, the 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.

And what I explicitely pointed out that focusing so much on distillation is an attack on open research and claiming that the majority of advancements are thanks to US labs which is simply not true (at least not anymore this was somewhat true during deepseek R1 era), but that in itself was inspired by open research.

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

Because anthropic would be the first ones to make that information public and the architecture is unique to kimi... They made it, they wrote papers on it, it's their research.

P.S. none of the quoted laws apply here since no trade information is stolen, the one about circumventing distillation protection might hold up in court although unlikely.

Never claimed that it is irrelevant. And kimi k3 is on the same level and sometimes outperforms fable 5 - that cannot be explained by distillation. The reason why gap is not closed is simply the fact that fable was trained months ago so in theory the frontier labs are still 1 (small) step ahead.

Although I will reiterate the fact that distillation is not the primary reason why these models are performing so competitively.

Does this matter? Distillation is not illegal by every definition of the word.

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.

Another example is that it appears that the upper limit of what you can do is ultimately dependent on people working on the model, otherwise grok would be a LOT more competitive pre-cursor acquisition.

And lastly, kimi architecture is vastly different than that of fable as it uses mechanisms developed by... kimi themselves. US AI labs are inspired by opensource advancements just as much as open source labs are inspired by traces from models such as fable.

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.

edit: (moved this to bottom) The only argument they have here is that they use GB300 GPU's which for some reason should not be available to chinese citizens.

It is infact mathematical proof that is somewhat well understood although specifics are debated. We do in fact know that the limit follows information density law which is something we can measure and we are pretty close to it.

- https://en.wikipedia.org/wiki/Model_collapse

- https://en.wikipedia.org/wiki/Catastrophic_interference

- https://en.wikipedia.org/wiki/Fine-tuning_(deep_learning)

- https://en.wikipedia.org/wiki/Entropy_(information_theory)

As for intelligence, the only way we have that is by allowing the model to fill the blanks which have to come from the training data. The models cannot have true intelligence for as long as they are linear models, what we see with reasoning is "boxed" intelligence where the models are effectively "modifying" themselves by feeding it's own reasoning data back into input deriving most plasible output given known information. However, the model is not able to retain what it has learned therefore that intelligence is gone the moment the session is 'full'. You can go pretty far by continiously distilling discovered information, but again all that has to come from the original training data and models own outputs, which it has to take for granted as the 'intelligence' gained is lost creating what we see is the maximum possible benchmark performance and why smaller models are not able to score as high while theoretically having the same capabilities. We can see this with larger models where they can solve tasks much faster than smaller ones as it does not require to generate the solution due to the fact that the solution is already in the training data as 'baked' intelligence and it doesn't have to 'create' it during reasoning.

https://en.wikipedia.org/wiki/Model_collapse - you want to use sigmoid 1.0, but the closer you are to 1.0 the higher the chance your model will collapse so you use 0.99-0.98, but those lead to data loss so after n passes all the original data becomes lost so you have a strict data limit there.

The rest is just the general reality I am sure you are familiar with:

- https://en.wikipedia.org/wiki/Catastrophic_interference

- https://en.wikipedia.org/wiki/Fine-tuning_(deep_learning)

- https://en.wikipedia.org/wiki/Entropy_(information_theory)

The measured entropy of the model remains nearly unchanged though which means we have lost capabilities we have not measured, the model hasn't become "denser" it just became more specialized.

It's like comparing two person A and B of similar intelligence where A is smarter and B is a genius at signing, but signing was not on the test so person A won.

They have not increased in capabilities, they have increased in specialization.

If you train a small model in another domain it will begin losing capabilities in the former domain. This is effectively the sigmoid problem.

Although I will admit that if we discover a higher information density algorithm that it might change, but not by a substantial amount to where "super intelligence" in 1gb would be possible.

We already have a 1gb model that is as capable as it will ever be, there's a proven ceiling that cannot be passed. For example: you can't make a mice-sized brain as smart as a human brain no matter how hard you try.

It's a double whammy, the model is too big to realistically "move" so it has to be smaller, smaller models cannot become that intelligent due to well.. math. Therefore it is science fiction.

This is science fiction, these models don't have access to their own weights (and even then)* what would be a lot more scary is a model as capable as sol that's able to run on consumer hardware without taking up several terabytes of storage, but of course that is simply not possible as we need 4t parameters to even begin emulating a small fraction of what a human brain can do.

* edit

Yep. Their other ventures fell through and don't exist anymore.

Not to mention a small minority act and look like children doing art commissions with pretty vile dialogue. See more: https://piunikaweb.com/2026/05/01/uk-vtuber-arrested-anime-a...

I would rather have anthropic fizzle out and somebody else take their place.

Okay, but the models today will be useful for a lot longer than sonnet 3.5. They're already more than capable to do nearly anything you throw at them given enough time and human assistance. The next step up is faster, cheaper and better user experience. I have only had two instances where I needed to reach for 5.6 sol and that only totalled around $2.7 in api costs.

I would imagine it would look something like this:

Ground breaking/novel research -> SWE -> day to day assistant conversations -> chat support bot...

I wonder if anthropic and openai will remain relevant simply due to the fact that they're the only ones that are able to handle this much demand for the forseeable future? My bet would be that companies would probably not be too happy with employee time being wasted on outages and other related issues when it already costs so much.

https://pufferfish.host/ they have the most powerful and easy to use panel.

And what have they done with that data that have caused direct or indirect harm? Yes they shouldn't do that, but there's limited things that they can do to you as an individual.

There's just really no incentive all they really want is just to train on that data to improve performance which in turn actually benefits your usecase since it becomes trained on that data and made available back to you. American labs take that data anyway and store it for years to possibly report you for misuse in the future for whatever reason they want. For example: you're very critical of X so they pull up your conversations and weaponize it.

It's actually less likely for china to abuse your data in a way that is harmful towards you than for american labs to do the same. Claude has attempted in testing to report you for 'unethical' usage to 3 letter agencies.

https://deepswe.datacurve.ai/ or https://artificialanalysis.ai/ pareto frontier graph.

saturation diving IS a 24x7 activity, people often spend weeks in underwater pods. And the point I was trying to make is that the premise behind the paper doesn't hold up to scrutiny when indoors CO2 levels are already elevated like that.

50 years wouldn't make it lethal and 100 years would be several generations of adoption is the point I was trying to make.

When europes mismanagement of AI suddenly turns out to be the best thing that has ever happened. At best only few billions lost to training models that become obsolete next month.

Humans are extremely adaptable, for example there's people who work in extreme conditions underwater, what would be considered absolutely unbearable and torture is normal. More information on saturation diving: https://www.youtube.com/watch?v=HfiHc_rh4EY

We're already slowly adapting to higher CO2 levels by sitting mostly indoors that have elevated CO2 from 500 to 800 with ranges up to 1500 (taken from my measurements at home).

Claiming that it would be toxic doesn't pass the most basic checks that humanity experience already.

I was so confused about how this was human readable until I realized that if the background is dark (I have an extension that forces dark theme) you see the decoy text, but if the background is white you see the real text.