Regardless of whether they achieved parity via distillation, or whether they got here via independently constructing a model from scratch, it was always going to end this way for the frontier American labs. Distillation “attacks” are not attacks. The frontier labs “distilled” all existing human written knowledge into their models, there was always going to be a second class lab that would distill that model into a cheaper version of it. There was never any plausible explanation for why this wouldn’t happen. There was never any practical mechanism to prevent someone from saving a conversation and using it to train their own model.
Even if it didn’t happen here, it was still the case that it was going to happen going forward. It was always going to end like this. Invest in the hardware companies, not the model companies.
Say it louder for the people in the back. All these complaints about "distillation" from frontier labs are bordering on felony contempt of business model at this point. It's great for us.
Almost all markets depend on some form of regulation whether its as simple as "leave everyone alone but no stealing" or "every participant has to source every object through mountains of red tape."
Thus far the US has not really chosen to go the Chinese rare-earth method yet. The problem with distillation attacks is the end result is everyone who is not doing them is going to deal with some kind of regulation whether it's complete loss of access, or the amount of control you'll have to give up to access them will be ridiculous.
Sort of like the "stealing music is fine" but "lets freak out now that it's producing visual art", in the end the entire thing is a social construct. Whether this is treated as theft or "business as usual" is entirely societal.
Eventually the gap will close, unless there's a major breakthrough that hasn't been made yet.
Given these models could not have been trained in the first place if they had to license every line of random fan fiction on the internet, I think distillation also being fair game is a tradeoff everyone should be willing to take (unless they want to decelerate, but that's a different conversation).
We're still in the early days of the AI industry timeline(relative to traditional industries). Not everything has yet been litigated.
Taxes on AI subscriptions or AI capable hardware, to financially compensate IP holders for (potential) IP theft, could very well arrive in the near future, once the industry is mature.
If this shocks you and sounds preposterous, I'll remind you that in several EU countries, we still pay extra taxes on any and all storage mediums and on devices with built-in storage (tapes, CDs, DVDs, HDDs, SSDs, tablets, phones, etc) simply because they can be used to store pirated content, decisions based on laws from 50-100 years ago, and the money goes to the national unions and associations of music and arts IP holders. It's basically a lobby pushed and government legalized extortion racket that no voter agrees with or can change but has no choice but to conform either way.
So I guarantee you in the future, it will be the same for AI subscriptions and hardware capable of running LLMs locally. Every time you purchase a Claude or ChatGPT subscription, an Nvidia GPU, Intel/AMD SoC PC or an Apple/Qualcomm powered smartphone, you'll pay a government enforced tax to the likes of Sony, Axel Springer, etc. for licensing their IP, whether you want to or not. In the EU at least. US maybe not.
That is incorrect. Anthropic paid $1.5 billion in compensation to copyright holders for use of their content in training data. OpenAI pays hundreds of millions per year across 150+ licensing deals for access to copyrighted data. Meta and Alphabet have similar arrangements.
Under the settlement, Anthropic was forced to delete the pirated data they were training on.
Chinese labs can still train on pirated data. I doubt the Chinese models operate under similar licensing agreements.
they didn't pay yet, because court challenged settlement as inadequate.
> I doubt the Chinese models operate under similar licensing agreements.
US corps likely pay licenses when afraid to be sued, or have troubles getting that data, otherwise they just take data, which was demonstrated many times. The same apply to Chinese corps, alibaba totally can be sued in US.
China is infamous for weakly enforcing copyright law. Even when it is completely obvious that Chinese labs are training models on pirated data, US copyright holders face a virtually impossible task of proving it in court. Those lawsuits won't go anywhere.
That's like saying someone is a big proponent of community law and order, and they donated $1000 to the county sheriff when actually they got caught drunk speeding in a school zone.
A false equivalence. A more correct example is: Anthropic was speeding, got caught by the county sheriff, and paid the fine. Anthropic stopped speeding.
Meanwhile, Chinese labs are speeding in a different county. Everyone knows they are speeding, yet the sheriff won't pull them over, so they just keep doing it.
This lax enforcement gives Chinese labs a structural advantage over American ones.
After the fact. They did the same thing Youtube, Uber and Airbnb did: Break the law, eventually get caught, cut some deal where they pay a pittance and keep doing the same thing but now with leverage on their side.
How is distillation an "attack" but gigascraping the Internet to the point of crashing servers and everyone needs Cloudflare and Anubis now not an "attack"?
I'm not aiming for a what about kickflip here: I'm saying we need to either agree on some rules or stop crying foul. Maybe the coherent legal theory is that neural networks and intellectual property don't interact. That would be weird but it would be consistent, a market could price it, I could do coding stuff and know if I was illegaling.
But this weird gerrymander that no judge will really rule on in an emphatic way is like, bad for the planet, bad for markets, bad business.
There are a lot of reasons to look forward to DeepSeek Huggingface drop kicking the unambiguous frontier weights in like, November, but I think my favorite one will be "who's distilling now bitch?"
I think you've basically got the legal theory. Training a neural network isn't prohibited by copyright law so if you can legally get your hands on something (e.g. by sending a GET request to someone with rights to serve the contents of their web page, or by buying a book) without signing a contract to not train on it, you can train on it.
But the American AI companies only let you query their models if you first sign a contract to not train on the output.
It's hypocrisy and unfair, but I think there's a strong legal argument for it.
Of course China can simply decline to assist in enforcing that contract... But I would expect US courts to do their best to.
Now THAT'S doing some heavy lifting lmao. The vast, vast, VAST majority of the original datasets were from pirated books and the like. Also, arguably a robots.txt is the exact mechanism to follow to do the mass GET-ing, yet the AI cos choose time and time and time again to simply ignore it and be as abusive as they possibly fucking can
Yes, anthropic and openai have really been brought to their knees and ipos cancelled because of the legal consequences of obtaining their training data.
This would have been a problem but it turns out that Anthropic is actually valued multiple orders of magnitude more than a copy of all the books in the world. So they survived the significant legal consequences.
> but I think there's a strong legal argument for it.
Maybe today. I doubt it tomorrow. Legal and not legal, largely, has to answer to the population sooner or later. Ultimately, humanity decides legality. And I don't think the frontier labs will get a pass from humanity in the midterm, let alone the long term. I think you'll see the rules change towards something more "intent" driven. And there's absolutely no difference in intent between Frontier labs and everyone chasing them.
Frontier labs just want the door closed behind them, as do their investors, because they know the money will never be recouped if others can do the same magic tricks.
I mean the latter, but more narrowly: China would never allow the United States to have a monopoly on machine intelligence if the only thing standing in the way of a domestic alternative was the Anthropic ToS. In general, I think that China is willing to agree on certain things relating to intellectual property. But not on this, it’s too big.
The US is already publicizing the way they are using Claude with Palantir for war gaming purposes. It’s a matter of national defense. Contract law has no meaning here.
Eh, I think you've done a pretty good job summarizing a collection of settlements with a few narrow bench rulings for seasoning. I'm not sure I follow you to it being a coherent legal theory. Buying a book in a bookstore is sure legal, and excerpting from it for e.g. literary criticism is pretty settled. Downloading every torrent of all e-books ever is pretty clearly illegal (or at least it fuckin would be if I did it). Pretty sure like, multiple labs have been popped for that though.
Situation right now seems more like a fragile detente: if you got a Hill staffer drunk and hounded him long enough he'd probably be like "God damnit the market will fucking tank if we don't get these two IPOs out north of a trillion. And don't even get me started on how I'm going to sell Chinese AI to a Senate that still calls people Nipponesians when no one is looking. We're doing the best we can alright, get off my back man."
We have a situation, but it's not exactly A&M Records, Inc. v. Napster.
> Downloading every torrent of all e-books ever is pretty clearly illegal (or at least it fuckin would be if I did it). Pretty sure like, multiple labs have been popped for that though.
Oh it is, and at least anthropic has paid $1.5 billion and deleted there torrented copies and not released any models derived from them as a consequence.
The thing is it turns out to be not that expensive to just buy a copy of every book legally and scan them. And there's even precedent that this is legal predating LLMs (Google books)
Great, let's go down to the courthouse and get some sworn testimony as to the ownership, value, condition, and so on and so forth of the bridge. And some document review and discovery run through professional legal firms under the same conditions. And perfectly reasonable and verifiable explanations as to why you own the bridge and are selling it (namely that you bought a copy of literally every book in existence in the meantime).
Facts are in fact knowable, and the US legal system is in fact not terrible at getting to them.
I think you're right to point out that historically the rule of law in the United States has been very robust by the standards of whatever era, it's been a tremendous advantage in attracting business and capital and talent, it's good stuff.
But we've gone through some pretty weird times too. Turn of the last century was pretty tech billionaire edits, reconstruction was uh, not smooth, it's a mixed bag.
And most takes I hear seem to acknowledge that this is one of those weirder times: serious election fraud rhetoric from most everybody from 2016 to the present, very politicized courts (on both sides to be clear), very soft on anti-trust, very soft on adventurous accounting. The Epstein files and like, no consequences (pretty much uniquely for a developed nation with Epstein people). It's weird right now.
And I think I would be hard pressed to think of a weirder part of this weird time than the rule of law meets AI. We can haggle on where laws end and norms begin (stare decis being maybe the midpoint), but in the 90s, the Justice Department got their brass knuckles on for a lot less.
I don't think it's a simple "the law works nothing to see here" story.
I broadly agree with your take on the state of the US - but this is a case where given the specific facts at hand I'm confident it still got to the truth.
I can understand why as someone who didn't follow it and the more corrupt legal developments closely you wouldn't be confident in that.
Knowlege should not have ownership. Training and distillation should be allowed
Granting people some form of control over knowledge only serves the public interest inasmuch it provides incentive to create more of it. Mass media, effortless duplication, and copyright extensions had already broken this to the point where control of knowledge was suppressing creation of new knowledge more than it facilitated.
The world has changed, we need a mechanism that works for the public interest that applies to the facts as they now are.
I strongly agree with the premise that distillation is not an “attack”.
But that said: K3 is not a distilled version of Fable or Sol. Fable has been barely available and Sol was just released! Moreover, K3 is superior to both models in some domains, according to user scoring on the Arena.
API distillation can’t give you these results anyway. All it is useful for is bootstrapping RL in new domains to get past the “cold start” problem faster. By far, what matters more is the quality and variety of RL environments the model learns from.
Look how hard Anthropic is to even be able scroll back on your conversation, or look at the thinking tokens or subagents. They want to keep everyone coming back to the watering hole but never to learn how to dig a well.
Well, there is precedence: Google can scrape the web, but you can't scrape Google. Laws around compiled databases exist for a reason: you can't just copy the phone book if effort has gone into compiling it, it is itself copyrightable
That varies by jurisdiction. In the United States, copying the phone book (or otherwise copying facts from someone else's collection) has been legal since 1991:
The fact that API based distillation is even a conversation right now makes me feel like the U.S. has their heads so far in the sand that it’s not really excusable.
These Chinese labs are producing novel models, publishing their techniques and sharing their open weights and the first topic of conversation is how they stole from U.S. AI labs.
Setting aside the fact that it doesn’t make any feasible sense to do API distillation, these models are outperforming frontier models on a number of benchmarks, and often times run more efficiently by several orders of magnitude.
We have to stop crying distillation, it’s getting embarrassing and at this point feels even a bit delusional.
There's little doubt that Kimi K3 was distilled off Claude. When Kimi was prompted to give its name, it answered "Claude" (according to this X post) https://x.com/denisewu/status/2077984660211269870 . This what happens when you include millions of Claude exchanges in your training data.
This line of thinking makes no sense because it assumes that labs that distill from frontier models are doing nothing else. It's the classic "the Chinese can only copy" mentality, and it's going to end poorly for American companies.
I'm pretty sure that all labs are distilling each others' LLMs, maybe apart from Anthropic and OpenAI. It would be stupid not to do it, because it's cheap and effective. But that's not the only thing they're doing. If you think K3 and GLM-5.2 got this good only from distilling frontier models, you're not paying attention to Chinese labs' publications.
i never assumed that, and i do keep up with the publications. i'm also not saying it's a dumb thing to do! what i am saying is that empirically, it appears that distillation of a more advanced model is a required first step for them to train a borderline competitive, cheaper model. in effect, their training is subsidized by the frontier labs.
if this were not the case, then we would be observing chinese models that far surpass frontier models in capabilities, rather than "almost as good, but much cheaper", and we would be having a very different conversation. what happens to these efforts when the subsidy is cut off?
> empirically, it appears that distillation of a more advanced model is a required first step
I see no evidence for that.
> if this were not the case, then we would be observing chinese models that far surpass frontier models
It's pretty clear that the primary reason for the difference is budget and compute availability. Chinese labs have at least an order of magnitude less money than Anthropic and OpenAI.
> what happens to these efforts when the subsidy is cut off?
They will continue making progress as they do now, minus the benefits of distillation.
Agentic reasoning and tool use
Coding and data analysis
Computer-use agent development
Computer vision
Moonshot (Kimi models) employed hundreds of fraudulent accounts spanning multiple access pathways. Varied account types made the campaign harder to detect as a coordinated operation. We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces.
I'm assuming you posted that as evidence for the claim that "empirically, it appears that distillation of a more advanced model is a required first step", but I don't think it is. It's just evidence that Moonshot distills Anthropic's models, which, yes, they do.
it is not a required first step for training a model, sure. but that's not what i claimed. what i claimed is that is how they are so significantly _reducing the cost_ of training one! how else do you think they are doing it?
the question was: what is the endgame for the stated "second class labs" strategy of distilling their frontier competitors then undercutting them on price?
Yes yes, we all understand the game-theoretic race-to-the-bottom you're describing here. Somehow despite linux being FOSS it still powers most of the important computing in the world. Can you explain how that works despite it being free? Once you understand that case I think you'll understand the game-theory behind how large projects can exist in the absence of traditional IP protection.
the obvious difference is the massive scale of data and compute required to develop and evolve these models, and the costs they impose on those building them.
Thanks for the models guys, sorry for your losses. Once this reality becomes mainstream and undeniable, surely the bubble pops and then what then. Future model development stops? Becomes private? Becomes a public effort?
The existing models are still going to exist. As hardware improves, there will be a day where it might cost a tenth of a penny to churn through 100M tokens a second of Opus 4.8. Established compute providers will invest in improving the models incrementally when margins drive them to look there.
If "distillation attacks" happen, we have to conclude there is some value add in what model labs do. Regardless of how we feel about using existing human knowledge in the way they currently do, it's simply impractical to infer that everything that happens downstream of LLMs can not be an attack on some IP because of it.
So both things can be true: a) People infringe on Anthropics IP and b) what Anthropic did to build their models is legally questionable (or might be ruled illegal, even though I doubt it).
Authors do not infringe on IP when they read another's book, nor should the lumber company be able to dictate how I use planks and if I can resell them if i'm done with them.
You're framing it as if the added value of the author or lumber company, awards them consideration when somebody uses the products to create more value.
IP law was always a big mess, and these questions cross far into ideology instead of law; but I do not understand people who think we need an ideology where more IP-law is good for society.
It's more simple: They infringe on the IP by way of violating the ToS. If you violate ToS and the company suffers financial harm, they usually can (usually) sue you in civil court for damages.
There are some quite interesting legal implications here. If Anthropic has IP over output produced by agents, do they somehow have legal rights to code and documents produced by such agents?
Unless someone literally stole the weights somehow (which is not out of the question, I doubt either oAI/Anthropic have the capabilities to prevent a state-level actor getting those weights), distillation from generations is not infringement on anyone's IP nor is it stealing nor is it an attack. It can't be. As long as you pay for tokens you get to do whatever you want with them. Someone saying you can't doesn't mean it's an attack or their IP or whatever. They either sell the tokens or not. They can decide to not sell them to anyone, but again that's not stealing.
And their ToS are a joke. Imagine how people would react if MS had ToS saying that you can't use MS software to develop solutions that compete with MS. They'd be laughed out of the room. Somehow it's ok for token sellers to decide what you do with the tokens? Why? If you pay for something you get to do whatever you want with that output. Train, distill, whatever.
Its definitely an attack. Thats established from anthropics perspective. No one has a right to use Anthropic’s services in ways that directly violate the ToS and user agreements.
> Its definitely an attack. Thats established from anthropics perspective.
How do things get "established" from someone's perspective, exactly?
By that logic it is established from my perspective that Anthropic has no right to train on anything I've written that is publicly available on the internet.
Of course, they don't care about my perspective, but then again I don't care about theirs.
Anthropic’s model outputs contain no IP. This is actually a simple legal proposition (rare in this field!) that derives from the fact that only specific classes of IP exist: copyrights, patents, trade secrets, and trademarks. Examining each, it is clear that API outputs do not qualify. Anthropic disclaims copyright in outputs; the outputs are not patented; the outputs are not secret (a prerequisite to having trade secrets); and trademarks are irrelevant in concept.
The output of Anthropic's models is not Anthropic's IP, as that would destroy their market, if Anthropic owned all the software it generated, and all the content. So distillation, which is just using those outputs is always going to exist.
I'm pretty sure that LLM output is not intellectual property. Nobody owns it, and it can't even be copyrighted. So using output from Anthropic's LLMs in ways Anthropic does not condone is not IP infringement.
Considering they were the original infringers, I don't know how anyone can expect tears to be shed here. The best we can hope for is for all these cancerous - and they really are the definition of a cancer - money burning entities to all fall apart to distillation attacks like these.
Regardless of whether it’s intellectual property or it isn’t intellectual property, it doesn’t actually matter. If AI doesn’t stop seeing diminishing returns in scaling up, and it hasn’t yet in the 10 years since the attention/transformers paper, the advent of AI will be the most important development in the history of humanity. Controlling that machine, or at least having one of your own, is an existential problem for nation states. It’s like a matter of national defense.
Do you really think intellectual property laws will prevent this in practice? It’s like as if we said, “hey, USSR, you can’t make a nuke, too! We patented that already.”
Asking China to not distill our models down is equally as ridiculous.
It's unlikely the USA would be granted an exclusive patent for the atomic bomb given the well-established existence of prior-art in the form of nuclear fission on the sun.
(I actually appreciated your analogy, despite my lark)
Anthropic’s IP is basically null and void for how they created it. And they might not want to try and challenge this in court, considering how they had to settle for using text books they had no right to use
Answer the question "how much does 5 cents of LLM computation in July 2026 cost in July 2005" and you'll have the answer to your question.
Don't forget to account for all the costs. It's not just that CPUs are X times slower. Memory is X times smaller, too, and networks are X time slower. And all this hardware is many times more expensive.
If I'm getting my mental estimation right, training a 2026-frontier-class LLM in 2005 would be somewhere on the order of all the computation power in the world at the time. It's not that many more factors of magnitude before you end up at "all the computation power in the world up to that point".
Is this some form of rage bait? 2005 we hadn't the GPUs, we have today. There are other factors, but I think this is the big one. The mathematics of building an LLM are really old, we just hadn't the hardware to do the needed calculations.
Right. Therefor it's not simply a derivative of information. The hardware is required to build the model. Software as well. The model uses information, it is not "distilled" from it.
"Distillation" literally means to separate and take some components out of something. You can distill how a model works from a model. You cant distill a model from information because the information does not contain the model.
People are happy to conflate distilling with building because they dont like how the information was used. You distill how the model works from the model, and you build a model with information. Both could be morally good or bad but its not the same thing.
Not really, what the information actually is, matters a great deal. It's harder to get good results going from "nothing > model+weights" than "nothing + traces from known good sessions of other good model > model+weights", this is what the "distillation" part is referring to. If "information is information", you wouldn't even need to separate good from bad sessions while doing the training, which leads to somewhat obvious results if you don't.
Can you be more specific? I have no idea what you are trying to say.
To succinctly restate my point, you cannot distill a model from information because the model is not contained within that information. You can distill a model from another model.
Their point is that "training" and "distillation" are essentially the same. The difference between the words is whether the source material is output from another model, vs being some original text.
That argument is moot as distillation also requires a lot of hardware and software, if copying models was as easy as that, we would have hundreds of competing models.
We didn't have the compute required (GPUs powerful enough to parallelize forward and backward pass). This compute is what allows us to train from human knowledge or distillation.
because you had neither the chips or the information in 2005. You have probably on the order of 5000x to 10000x more GPU compute today than you had in 2005 and three to four magnitudes more openly available data.
The first "L" in LLM does the work. In 2005 you had no Github, Stackoverflow, Youtube, common crawl and no archive of digital ebooks.
This was always where this was heading, but we got here much faster than expected.
Once western governments declare it to be a "national security" risk for citizens to have access to open-weight frontier models, and once they classify using these models as acts of terrorism, what will that world be like?
Will using Kimi K3 come to be like how napster was in the olden days? Everybody knew it was technically illegal, but come on -- any track at your fingertips? But surveillance is quite more evolved now.
Or it will be like cannabis, where a guy in the neighborhood will low key rent you metered access to the 8x5090 rig in his basement he cobbled together from parts on ebay? Or everyone will flock to VPNs?
Or will the oppressors actually succeed? The same way that napster is long gone, and everyone accepts that they must pay spotify for a homogenized collection, where artists must take only a minuscule cut (more than napster though)... We'll be stuck with nerfed Cohere or Mistral models for open-weight options, as if they need more lobotomizing. Or else we can pay through the nose for Anthropic/OpenAI for "American Frontier" models which will fall increasingly far behind China.
Or else, like how Kindle Fire was subsidized by ads, we'll have "Kindle AI" where influence is sold to the highest bidder, where the LLM will tell us that smoking is actually healthy if big tobacco can engineer its renaissance by turning its lobbying dollars to pay-to-play, pumping its propaganda into the training pipeline for Amazon's extra commercialized line of ultra budget LLMs.
Basically a new iron curtain didving the world into digiatl blocks.The era of open internet/science is on its last legs with the potential forr bifurcation into incompatible ecosystems high , the onger the exchange is disrupted.
As recently as this month the USgov has donce a Wolf Amendment style declaration for the Scientific collaboration NSF while shifting its purview under the military.
To add to that its trying to rope as many countries into its Pax Silica idea intentionally to exclude China while simultaneosly coercing its 'allies' into using its nerfed offerings [1]
So maybe some isolated switzrland/singapore type locales would exist for US/EUusers to be able to dip their toes across the curtain legally without reprucursions.
At this point, the United States will lose that battle most Countries in the world are going end up using electronics from Asia, that ship has sailed Japan, China, Korea, Singapore, Taiwan, Vietnam, dominate that area, China already dominates EVs, Drones and many other electronic devices, and with the way Donald Trump has picked fights, Europe, Canada, Australia, New Zealand, Mexico and many others are looking for other business partners.
If you need infrastructure done, China is dominating that area too. Rail, High-speed rail, Nuclear reactors, (near future Thorium reactors), Dams, Highway roads, bridges, Ocean ports, airports you name it, and they can roll it out, Transport ships, And if they don’t do it, Japan, Korea, Vietnam, and Taiwan do.
Is it too late? No, not necessarily, but America needs a regime change…
If they outlaw open source models that'll just handicap American companies, the rest of the world will be running open source and have an arbitrage against US companies.
EU countries are homeshoring their digital stacks as fast as they possibly can, and the reason is precisely because of the pressure that the US government is exerting via its (temporary) dominance in technology.
Those days are gone. Look no further than the occupant in the White House. IE the Swedish jet industry is about to get bigger, future drone expertise, if the Ukraine can hold on if you want to learn the ins and outs, you don’t need the United States. If you’re serious about learning and building drones.
It’s going to be a different world, a world where many former allies are not gonna look to the United States first they can no longer afford to.
Not sure if you noticed, European countries are distancing themselves from the US. They couldn't be pressured to offer logistic support to the US shitshow in Iran, why would they be pressured to help the US in its protectionism of its AI bubble?
I tried Kimi K3 on a task I've done with every other model I use regularly (https://swelljoe.com/post/i-let-every-agent-implement-its-ow...) and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan.
I only have the $20 plan from OpenAI and the same task, with a lot of the same implementation details as Kimi Code, only took a few minutes and consumed almost none of the 5 hour limit.
Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare, but when I sat down to add Kimi Code to flar, it was because I wanted to try it on some real work and then couldn't do any, because usage was nearly gone after the trivial task...no other ~$20 subscription I have has felt that tight before.
So, it was really slow to complete the task and seemingly much more expensive than every other model I'd tried. Maybe bad luck. Maybe it'll do better on other tasks. I wouldn't know as I was out of usage when I had time to try.
It did find a bug that Gemini 3.5 Flash introduced unprompted, though, so it has that going for it.
We really need to stop using $/M tokens as the pricing benchmark. I've found that the number of tokens used tends to be a bigger factor than the listed per token price. The cost per task vs. intelligence curve is really what you care about, and in my estimation Chinese models are just not there. They are focused on benchmaxing and getting the highest raw score they can, rather than efficiency.
In my experience, Kimi just tends to think a lot, with the main thing that takes up a lot of space is it constantly second-guessing itself. I've watched it do paragraph after paragraph of "Wait, actually..." while it stumbled and used a ton of tokens on one small detail of what it was asked to do. Though I also gave GLM 5.2 a task to port some JS code to Python to test it, and in my experience it doesn't second guess as bad as Kimi does, but it really did there. It kept doing web searches and second guessing tons of tiny little things, using $0.25 of API spend in total to port about ~50 lines of JavaScript. It did produce an error the first run, but on second run it gave me a program that ran.
I gave Claude Code/Fable the same task and it took significantly less time, but also stumbled on the same error as GLM. I didn't have it fix it though. I was mostly interested in timing differences.
I do like open models where I can, but I'm really hoping they get trained to second guess less. Or maybe I just need to prompt them differently. I'm not sure.
Earlier today I made Claude code implement a feature with fable. It worked roughly 60 minutes and used around 30% of my 100€ subs 5h sessions.
Then I typed /code-review in a second terminal/clean session after the analysis was done (no code changes) the usage was 99%. I then asked it to write that into a review.md so I could restart from that the next day. Sadly the last % wasn't enough for that.
Ymmv, these models behave very differently with no discernable reason. Usually reviews(even with fable) take like 10-20%... Yet suddenly you get it to burn through 65-69% in 15 minutes or so
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
AI subscription pricing is so goofy. You get some amount of usage that varies by models, is measured by opaque token usage, driven by how many tokens the (usually) vendor-provided interface (or model itself) wants to use. Then your usage is limited by time opaque time windows.
AI subscription pricing was fine when it was $100/month for some opaque 5 hour token budget I don't think I ever used, not even that one day where I coded for 14 hours non-stop using Fable. But like most people with low token usage, I had a human in the loop and and I didn't use workflows with swarms of agents.
Now, of course, the plan is to remove Fable from the subscription. To paraphrase Darth Vader, they have altered the deal. Pray they do not alter it further.
Gpt 5.6 is still like this at least for the $200/month option. It’s also always faster than fabel. Fabel might be able to do some things better but I don’t have time to constantly wait and find out.
Yeah I've noted this behavior with best in class open weight models. They said K3 would have token efficiency improvements and I was hoping especially solving the thinking loop issue that plagued K2.x but even if this release helped somewhat, it looks like we still have a long way to go here... I'm not sure what's up here but I suppose lacking finetuning quality.
What OpenAI in particular have done with reasoning efficiency in the past few months since ChatGPT 5.5 is nothing short of remarkable. It's overshadowed a bit by the benchmark game and the Fable hoopla.
Now is the time to focus less on token cost and intelligence, but tokens to solve a particular set of tasks in closed benchmarks for a variety of categories.
What is the use of grand intelligence if it either costs you a kidney or can't complete at all within a token budget? Even if there are niche uses where you truly want "maximum power" above all, we need to at least more severely penalize such models versus those that does it just as fine within a tenth of the token cost.
I'm aware of some benchmarks at the Artificial Intelligence site, but CLEARLY we are not focusing enough on these today and still leaving the fun surprises to the users.
Yeah, I'm finding I end up switching to Codex and GPT 5.6 a lot lately because I've either run out of Fable usage or Fable refused to do the task. Most recently it refused to work on a WiFi configuration UI for a robot. No idea why it thought that was related to security, biology, or some other sensitive topic. They've hobbled it with guardrails that are overzealous and now there's a big opening in the market. Fable may be the best, but if it won't do the job half the time, it stops being my go to model as I don't want to waste time only to find it refuses halfway through.
The more important question than subsidy is what is the tokenomics of running the model. If it's inefficient to run on an nvl72 cluster (or whatever the heck has enough vram to run a 3T parameter model), and k3 isn't very token efficient, then it might not be that compelling of an open weights model.
Subsidization could affect both of those. If you have $200B in the bank you can afford to throw massive compute at every single request; if you are less well funded, you might throttle more aggressively.
Additionally that same VC could be (read: is always) spent on developing the harness, and other infrastructure around the model, not just the model itself.
So it's apples-to-oranges when comparing a relatively new model to established competitors (i.e. OpenAI @ $900B funding vs Moonshot/Kimi's $30B FYI) because every new model they release is judged on "performance" which is not strictly speaking derived solely from the model.
It's possible Moonshot could get similar performance over time as the build out the rest of the infrastructure. We have no way of knowing how much of OpenAI/Anthropic's success is due to the model vs intelligent tooling built on top of it.
Yep, with Reasonix, DeepSeek is free real estate. Seems to just go and go for pennies.
And, DeepSeek is what I use for any task that works best with an API. It's cheap enough to where I don't think about cost, made even cheaper by DeepSeek having the most effective and cheap caching in the industry, and it's good enough to where I rarely have to follow up with a more expensive model or manually fix things. It's been alleged they're releasing an update to DeepSeek V4 Pro soon that improves it, which likely makes it a good fit for even more kinds of problems. It remains my favorite of the Chinese models, it's so cheap and cheerful. And, is also less aggressively censored than some of them.
> It's cheap enough to where I don't think about cost, made even cheaper by DeepSeek having the most effective and cheap caching in the industry
I use DeepSeek v4 Flash & MiMo v2.5 Pro. Prefer the latter over DeepSeek v4 Pro because it costs the same while being equally good & less chattier for coding workloads. Although, I've begun experimenting with Hy3 (as an in-between Flash & Pro) & GLM 5.2 (for long-horizon tasks).
It let's me choose different thinking levels in Kimi Code. Not sure if it actually works, yet, but it says "Thinking set to high." when I change it from max.
Absolutely do not pay for the kimi plans thinking they will be cheaper. If you sign up with a Chinese phone number, you can get the same plan for 200 yuan instead of 200 usd, it also only accepts Chinese payment methods iirc. So the plans are really made for Chinese userbase.
That sounds complicated. I'll just use my month of Kimi and then cancel. I have too many AI subscriptions to use them all, anyway. I subscribed mostly to test it. I mean, if it turned out to be competitive, I would keep it, but if it doesn't turn out to really excel and anything and also take longer than Claude or OpenAI models, I'll stick with them.
It would be really interesting to redo the public benchmarks for kimi k3 but token normalize the costs. Ok so maybe k3 beats fable on terminal bench, but how many tokens did it use?
> tried Kimi K3 on a task I've done with every other model I use regularly and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
I always put my coding subscriptions (that allow it) through "AI gateways" (Cloudflare & OpenRouter are free) which help track token use.
In my experience, Kimi & Qwen Cloud have opaque & restrictive limits, their "credits" drain faster. I now make it a point of subscribing (directly [0]) with providers that are transparent like MiniMax, DeepSeek, Xiaomi, & Z.ai.
[0] OpenCode Go, Cline, and AtlasCloud have generous limits for open weights, otherwise.
With the obligatory disclaimer that I’m impressed with what open weight models can do, I have the same experience with all of them.
The benchmarks come out and say they’re as good as Opus from N months ago, then I use it for a complex task and it doesn’t work as well as Opus from N months ago did when on similar problems.
There’s a real wow factor when you get an open weights model to do amazing things, but in my experience the gap to the frontier models has always been bigger than the benchmarks would lead me to believe.
There can be a lot of value in having the cheaper open weight models for chewing through lower complexity tasks (non-programming in my primary use case) at a cheaper rate than OpenAI or other frontier API costs. Even with those I can measure bigger gaps to the frontier models than the benchmarks suggest.
If the benchmarks aren’t being directly gamed, there’s at least some selection happening where training data or model structures are being picked in ways to maximize public benchmark performance. All of the labs know there’s immense value in having good benchmarks to show for your model because most LLM consumers are picking based on lab provided benchmark charts, not running their own evals. Running your own evals is hard and expensive.
Kimi K3 has 2.8 trillion parameters. We don't know the number of parameters of ChatGPT 5.6 or Opus 4.8, but it's probably in the same region. Fable/Mythos are rumored to be around 10 trillion.
So, K3 is directly comparable with ChatGPT 5.6 and Opus 4.8, and the price is not so much lower:
K3: $3/$15 per 1 Mtok input/output
ChatGPT 5.6 Sol: $5/$30
Opus 4.8: $5/$25
This is not a watershed moment. It's a competitor converging to the same capability and trying to undercut your prices, but not by a lot.
As for the open weights? For now, Kimi K3's weights are closed, and I don't expect the situation would change.
Even in this very thread the feedback on Kimi's actual efficacy is debated. I personally feel its worse than both Fable and 5.6 Sol, but I feel like the conversation isn't really about whether its good or not, but a backlash against the U.S governments foray into regulation. So I think people _want_ it to be superior out of anger/frustration with the current situation.
This seems like a replay of what happened with DeepSeek. They put out v3, or whichever one it was, and everyone said it was over for US companies... then everything continued on.
It's really good. I'd put it between Sol and Fable. I'm not super impressed by Sol's UI design skills, something K3 is strong at. Fable is still overall the fastest, most consistently well-performing model, though.
This does depend heavily on the kind of work you do and how you use these models, but the idea that K3 isn't right up there with US SOTA models doesn't match my experience.
When you net out across benchmarks and firsthand reviews it seems like it's maybe a little behind. There seems to be a consensus it's token hungry and a little slower. So maybe it's a point release behind.
That's weeks maybe months behind, not months maybe a year behind. It's "would my life really change if Claude was gone, not really" behind.
I actually haven't used it much, because Claude started kicking ass again the last few days. Like, way too much of a difference to be normal load-based variance. I got more done in the last 48 hours than week before that.
This line made me think by 'normal coding work' the author means doing something they don't understand well enough to be able to distinguish the models' output.
This is comparing Fable High with K3 High. I'm mostly using these models for game development. The tasks I usually send are ambiguous visual bugs, changing the look of a scene or models, or adding a large feature. The wording wasn't accurate there. I don't use Fable or K3 most of the time. I'm usually working on smaller scoped tasks that I review myself afterwards.
It was all distillation up to this point anyway. And I agree with what Suhail said on twitter: "Make the margins next to zero for all these AI models. It was trained on humanity's data, it should be gift to ourselves. Doing so will save us from a few in control of our species."
Pricing is actually far cheaper than that. There's two tiers of pricing: Chinese and US.
If you sign up with non-Chinese phone number, you're bucketed into US, you get US prices, can pay only in USD and with American credit card network.
Chinese prices are about 9x cheaper than the US prices, which are already far cheaper than Claude or other American provider. If you can somehow get hold of a Chinese phone number, keep in mind that you can save ~90% of the bill.
My assumption is that Anthropic, OpenAI, Kimi, etc all have a similar cost structure when serving models. The same size model roughly generates the same GPU usage whether you’re American or Chinese. I’d also guess that the model sizes across all SOTA models is similar, we just only see data for open models. The difference is most likely that American companies simply charge more because they have the dominant market position.
Remember not too long ago when Anthropic was charging $75/mt for Opus? Now that many models are in “opus tier”, their pricing is $25 - higher than competitors but close. The newest Kimi is $15. 40% lower to forgo “made in America” with American enterprise support staff is not crazy. Compare AWS to Hetzner or any other flagship enterprise service to the foreign and discount option. I assume that over time, we’ll see the commodification of models reducing prices even towards the raw GPU costs.
Well, there is the small issue of privacy policy: Kimi will train their models on your interactions if you use their subscriptions, and only with direct API usage (billed at API prices) they say they won't. Whether you trust that is another matter.
Those things do make a difference to some of us, even though nothing is black and white. In my case, I'll probably want to wait until other providers appear through OpenRouter and then I'll try to judge how much I trust them. But even if I don't trust them much, they don't train models anyway, so the likelihood of my data being used that way is smaller.
> Kimi will train their models on your interactions
I find these kinds of concerns increasingly silly: most of the input to these models will be ... previous output from the very same models, alongside the occasional half-assed human command to fix something and "make zero mistakes". Who cares if they train on that? Let them, if it makes their future models better!
99% of users are not working on any special IP to worry about that.
Indeed, the B2B / no-data-retention market is still going to provide plenty of business for American companies even if every hobbyist uses open-weight models.
I can see the economics of open vs. frontier models turning out similarly to pharmaceuticals, where generic drugs cost a fraction what the name brands do and Americans end up paying the highest prices in the world partly as a consequence of propping up drug discovery research.
I think the biggest problem with Chinese models is that they seems to overthink for most of the tasks, especially for smaller ones. The OpenAI models have in my experience only gotten better in terms of efficiency.
Yes, this (imo) is a clear result of benchmaxxing. You can get a much better score on most "intelligence" benchmarks by massively over-saturating reasoning. This looks good on those, but for actual daily usage makes the models much less effective: I don't want a model I use for coding to burn a bunch of reasoning (read: time) on trivial tasks.
It's undeniable that some of these models generate a ton of thinking tokens, but it's arguable whether that makes them "much less effective."
For example, Kimi 2.7 has been really effective for me despite having verbose thinking blocks, simply because it runs so fast. Speed-wise, it feels about like Sonnet, possibly faster.
I strongly suspect the flip side is that in the future it enables you to train smarter models by "distilling" the end result of the super duper heavily thinking models.
The current administration's immigration policy isn't helping. This wouldn't have happened 10 years ago because the US was this city on the hill that everyone wanted to immigrate to. Talented Asian researchers would have immigrated to the US and China would be deprived of talent.
No it is H-1B visa. Right out of the university it is hard to recognize extraordinary talent. People like Sundar Pichai were not recognized as extraordinary right out of the university, he had to start at the bottom and rise up the ranks.
To be fair, that was clearly well deserved. Marrying Trump and then becoming first lady is definitely an extraordinary ability; I doubt I could have done it.
This makes even less sense, Trump admin has been here for 1 year, the implication here is a university grad on H1-B in January would become a world class researcher capable of building a frontier model in <18mo
This is why VC is actually hard. Everyone’s instinct is always “Man, once the company has demonstrated it’s awesome I would love to have been in the seed round”. The tendency to want “proven performers” is the default belief.
When people demonstrate their capability thoroughly, the Chinese government takes away their passports. You’re not exactly going to get them here with an O-1.
This is essentially the point of the visa, it feels wrong especially as YC drops standards and increases cohort sizes, but the same power laws that keep them winning also apply here in maximizing economic value of each O-1 approval
Basically of all visas O-1 is virtually guaranteed to have highly positive economic value
Your comment feels like an outdated brain drain model where talented Chinese researchers naturally want to leave China and the only question is whether the US lets them in.
That may have been closer to reality 10-20 years ago, China is a different country now, what I mean by that is they offer research funding, they have huge digital behemoths (alibaba, tencent, huawei, bytedance etc), large scale deployment opportunities and prestigious careers. Many graduates return because the opportunity set is attractive and they want to return, it's not just because US immigration policy pushed them out. Some also want to contribute to their own country's technological progress (which is a normal motivation btw), like probably you are also a patriot and want your country to succeed.
So, really, China's AI progress is not mainly the result of America failing to absorb every talented Chinese researcher. China has built a domestic ecosystem capable of producing and keeping top talent itself. I feel like a lot of Americans do not understand this.
Chinese students still want to attend US universities [1]. While it is true that the progress made by China is a factor, this administration's policies are the bigger deterrent [2] [3].
It doesn't really address the point. Chinese students wanting to attend US universities is evidence that US universities remain attractive, not that those students would otherwise permanently immigrate to the US or that China lacks attractive careers for them afterward.
US immigration policy may be unnecessarily pushing away talent but the assumption that talented Chinese researchers would naturally remain in America unless prevented from doing so ignores the growth of Chinese universities/labs, companies, their funding, national prestige etc.
I mean, don't get me wrong, US is still highly attractive, it is just no longer the only place where an ambitious Chinese researcher can do important work and grow.
"China can draw on a talent pool of 1.3 billion people, but the United States can draw on a talent pool of 7 billion and recombine them in a diverse culture that enhances creativity in a way that ethnic Han nationalism cannot." --Lee Kuan Yew, former prime minister of Singapore.
- china’s homegrown tech industries already achieved escape velocity from it a long time ago, after China fenced off its market for Alibaba and Baidu in the ‘00s. some of their AI innovation at the edges was already top class 10 years ago
Correct: can't opt out of training. This is well documented.
"Can't use for commercial purposes" - incorrect AFAICT. In what sense do you mean this? The open weight MIT version obviously allows for commercial use, but I don't think that's what you're referring to, because training data is irrelevant on the open weight version. Pretty sure the API allows commercial use too. Maybe the free version doesn't? But who cares?
GPT 5.6 Sol comes out ahead of Kimi K3 on price/task (but not significantly so). You're probably thinking, "Why use Kimi K3? Isn't an open model supposed to beat the closed one on price?", but you need to consider that the closed models are completely hobbled when trying to do anything security-related. For my use-case, I can't risk getting pwned because I'm using a model that refuses to secure my app while there is now an open model that obliges to obliterate any app that isn't protected.
I never truly understood what the intended business model around LLMs was. Get them widespread through cheap pricing and then jacking it up? Being the only ones that had a viable product so to get the ability to extract as much value as you want from AI?
I don't understand how a product that:
- is interfaced with and is deeply linked to natural language, so everything you produce (sessions, history, etc) is in Markdown and you can literally install a second model and tell it "hey import all of Claude's memory into yours" and that's it
- is based on well understood technology, the real constraints are how much money you put into training the models, but the theory has all been developed in the open
- clearly has a threshold where it quickly commoditises and turns from "I want the best" to "hey the best is a bit too expensive. The second best is half the price and works close enough".
was ever supposed to be a money printing machine. The fact something is extremely useful doesn't imply it's extremely profitable.
IMHO we're clearly speedrunning the process of turning AI into a commodity. Dario Amodei knows pretty well that when or if Anthropic cuts people off Fable, the vast majority of them will definitely not pay for it because Opus 4.8 is good enough for almost everybody that _knows_ what they're doing, and so are basically half of the most recent models. If I already have good baking skills I don't become more productive with an automatic bread machine, I just need a better dough mixer and oven
There is no business model. That’s not a joke, the idea is to be the one that survives the race, then figure out how to be profitable. If you look at the level of capex and money raised, that’s not something you do if you have an actual business plan. We are very far from business fundamentals
> I never truly understood what the intended business model around LLMs was.
A closely related question is “what do the American labs need to do in order to justify their enormous market valuations?”
It seems like the answer cannot possibly be “gradually improve model capability while figuring out how to better monetize inference.” The valuations are just way too high for that to be sufficient.
Surely the answer has to be “continually achieve large leaps in capability comparable to the first consumer releases of ChatGPT while also maintaining a significant capability lead over open models and new competitors.”
And does anyone think that’s going to happen? Even with state-level protection from competition (which incidentally would significantly harm the American economy), the large leaps in capability seem to be coming fewer and farther between.
> I never truly understood what the intended business model around LLMs was
What appeared initially to be a huge innovation was later easily duplicated by many. There are no platform-lockins or network effects. Switching costs for users are zero, and there are low barriers to entry, with vast numbers of models to choose from and more appearing every day. As a business a token will be a commodity like an electron. Doesnt matter who produces it, or how (solar, wind, coal, nuclear etc) as long as it powers my toaster.
It's fairly simple. Sell GPU compute + extra margins as only some GPUs can load the models + extra margins based on how much better closed source models are from open source ones + hopefully reduced cost due to batching
The valuation is based on one lab getting a decisive first advantage, and turning that into a durable self-improving advantage that can never be caught up to. If any can pull it off (a gigantic if), they will effectively own most AI value, and the people who own their shares will live happily ever after. Divide your investment between the labs that could plausibly do this, and your EV may not be dreadful.
This is clearly not how it's going though. Any advancement from any lab has been quickly (< 6 months) matched up by basically everybody else. Even Grok nowadays is decent, and that's something. When something like you've described actually happened historically you generally had quite fast a clear frontrunner and a bunch of copycats that failed miserably; in 2026 we are very far from that. we are heading face-first into towards a pricing war because all models are easily interchangeable nowadays - AI is turning into a commodity more or less
If you have multiple metrics to evaluate goodness of a design, one would normally need to decide which metrics they care the most about in order to find the "best" design.
The Pareto frontier tells you which designs are the best in at least one of your metrics (non-dominated by another design). For example if you're selecting a car and you care about both speed and mpg, a Formula 1 car and a Prius might lie on the Pareto frontier, but a Model T Ford would not.
The set of models that are pareto-optimal, IE for some set of variables, no other model strictly dominates them = no other model is better than them on every variable.
So like, on a cost-intelligence graph, the cheapest and most intelligent models are pareto optimal. Then in-between those if you have
- cost $3 intelligence 6
- cost $1 intelligence 5
- cost $2 intelligence 4
The 1st and 2nd are pareto optimal, the 3rd is not, because it's dominated by the 2nd (2nd is cheaper AND more intelligent at the same time)
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.
It’s not that the Chinese firms are any less likely to misuse your data, it’s that you don’t live in china, so their abuse of your data is unlikely to directly impact your day-to-day life in the same way
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.
Are you comfortable sending it to US ones? Especially if installing Claude Code or another tool on your PC and it can collect all the data it wants..
On Openrouter Kimi K3 says it does not retain data or train on it, which is better than what US hosts claim for Claude, ChatGPT, etc.. as they collect and retain data even if you disable training on it.
Opencode or similar open source tool + a zero data retention provider is about the best option aside from running a smaller fully local model on your own PC.
For open weight models, you can choose from a few providers. Each have their own caveats, none of ToS'/Privacy Policies I entirely trust, nor do many make renewable energy claims.
> I think I can see where this goes. The government will try to regulate AI and open source in particular, and it will run the playbook it ran for the auto industry. Decades of subsidies, bailouts, and protective tariffs produced American carmakers that sell trucks at home and barely register anywhere else in the world.
Here's the thing about this though, the auto industry directly employed hundreds of thousands of people.
The AI labs are small, only few benefit directly from their wealth and there's already immense opposition to AI, data centers, etc...
I cannot imagine wanting the product of the entire intellectual output of humanity since recorded history to be an expensive, paywalled commercial product owned by 5 or so of the most insufferable, detestable people who ever lived. Why would you want to live in that world. We should all be rooting for open source AI to win.
In my experience GLM 5.2 is a pretty good Opus replacement. But K3 has not given me an experience on par with Sol or Fable. The price/intelligence ratio might still make sense. But it’s not very inspiring when it comes to my real world tasks. I’m doing pretty mundane web stuff.
Its worthwhile to have a quote from the article as some comment without reading:
"...I’ve been running Kimi K3 alongside Claude on my normal coding work, and for all practical purposes I can’t tell them apart. Same tasks, same quality of output, and near identical token counts to get there. I expected an open model to be sloppier or to grind through more tokens on the way to the same answer, and neither turned out to be true.
The prices are nowhere near each other. K3’s API runs $3 per million input tokens and $15 per million output. Claude’s top model costs $10 and $50 for the same units. The subscription side is even more lopsided..."
I don't know. You can't just rely on looking for em dashes or other obvious tells because anyone who cares can get the AI to avoid those.
> When the headline model on your plan can be switched off because the economics don’t work, the plan was never really selling you the headline model. Kimi’s tiers don’t come with that asterisk.
This line has a certain smug, punchy cleverness that I associate with AI. To me, the vibes are ~30% AI writing.
> The prices are nowhere near each other. K3’s API runs $3 per million input tokens and $15 per million output. Claude’s top model costs $10 and $50 for the same units.
And this is the point where your internal compiler should have started shouting 'Type Error'
Notice the trick here?
> Then there’s the fine print. Claude couldn’t sustain Fable access on the twenty dollar plan, so they turned it off, and the plan quietly falls back to Opus.
Ask Claude whether a man can be a woman. No matter what political side you're on, models are censored. I'd argue American models are a lot more censored/biased on a lot more topics, and especially on things that come up a lot more than some highly specific Chinese politics. Claude even refuses to translate song lyrics because of copyright.. I wish we could simply have uncensored models from all sides, but it's pretty clear that's not happening right now.
I can't recall the last time that it was useful knowledge when writing code. Reservoir sampling, online softmax, Otsu, sure, Tianenmen square not really.
Regardless of whether they achieved parity via distillation, or whether they got here via independently constructing a model from scratch, it was always going to end this way for the frontier American labs. Distillation “attacks” are not attacks. The frontier labs “distilled” all existing human written knowledge into their models, there was always going to be a second class lab that would distill that model into a cheaper version of it. There was never any plausible explanation for why this wouldn’t happen. There was never any practical mechanism to prevent someone from saving a conversation and using it to train their own model.
Even if it didn’t happen here, it was still the case that it was going to happen going forward. It was always going to end like this. Invest in the hardware companies, not the model companies.
> Distillation “attacks” are not attacks.
Say it louder for the people in the back. All these complaints about "distillation" from frontier labs are bordering on felony contempt of business model at this point. It's great for us.
Almost all markets depend on some form of regulation whether its as simple as "leave everyone alone but no stealing" or "every participant has to source every object through mountains of red tape."
Thus far the US has not really chosen to go the Chinese rare-earth method yet. The problem with distillation attacks is the end result is everyone who is not doing them is going to deal with some kind of regulation whether it's complete loss of access, or the amount of control you'll have to give up to access them will be ridiculous.
Sort of like the "stealing music is fine" but "lets freak out now that it's producing visual art", in the end the entire thing is a social construct. Whether this is treated as theft or "business as usual" is entirely societal.
Eventually the gap will close, unless there's a major breakthrough that hasn't been made yet.
Given these models could not have been trained in the first place if they had to license every line of random fan fiction on the internet, I think distillation also being fair game is a tradeoff everyone should be willing to take (unless they want to decelerate, but that's a different conversation).
Us models didnt pay for licenses too
We're still in the early days of the AI industry timeline(relative to traditional industries). Not everything has yet been litigated.
Taxes on AI subscriptions or AI capable hardware, to financially compensate IP holders for (potential) IP theft, could very well arrive in the near future, once the industry is mature.
If this shocks you and sounds preposterous, I'll remind you that in several EU countries, we still pay extra taxes on any and all storage mediums and on devices with built-in storage (tapes, CDs, DVDs, HDDs, SSDs, tablets, phones, etc) simply because they can be used to store pirated content, decisions based on laws from 50-100 years ago, and the money goes to the national unions and associations of music and arts IP holders. It's basically a lobby pushed and government legalized extortion racket that no voter agrees with or can change but has no choice but to conform either way.
So I guarantee you in the future, it will be the same for AI subscriptions and hardware capable of running LLMs locally. Every time you purchase a Claude or ChatGPT subscription, an Nvidia GPU, Intel/AMD SoC PC or an Apple/Qualcomm powered smartphone, you'll pay a government enforced tax to the likes of Sony, Axel Springer, etc. for licensing their IP, whether you want to or not. In the EU at least. US maybe not.
I think we are going to direction where AI corps will have stronger lobby compared to IP holders.
That is incorrect. Anthropic paid $1.5 billion in compensation to copyright holders for use of their content in training data. OpenAI pays hundreds of millions per year across 150+ licensing deals for access to copyrighted data. Meta and Alphabet have similar arrangements.
Under the settlement, Anthropic was forced to delete the pirated data they were training on.
Chinese labs can still train on pirated data. I doubt the Chinese models operate under similar licensing agreements.
Anthropic paid $1.5 billion in compensation to copyright holders for use of their content in training data.
The payment was for illegally downloading copyrighted material, not training. Training was explicitly ruled to be fair use.
The court explicitly ruled that training on pirated data, which is what Anthropic was doing, is not considered fair use.
Training on legally acquired / licensed data is potentially fair use.
they didn't pay yet, because court challenged settlement as inadequate.
> I doubt the Chinese models operate under similar licensing agreements.
US corps likely pay licenses when afraid to be sued, or have troubles getting that data, otherwise they just take data, which was demonstrated many times. The same apply to Chinese corps, alibaba totally can be sued in US.
China is infamous for weakly enforcing copyright law. Even when it is completely obvious that Chinese labs are training models on pirated data, US copyright holders face a virtually impossible task of proving it in court. Those lawsuits won't go anywhere.
There are tons of lawsuites which resulted in banning Chinese companies from doing business in US, those lawsuits totally have consequences.
They settled with a subset of copyright holders. Guarantee they violated lots of others' rights in the process
They only paid when they got caught. And not to everyone.
But they still paid. I don't see any Chinese labs paying billion dollar infringement settlements.
Chinese labs can freely train on pirated material, which is a structural advantage.
really!? nobody paid me anything for my comments on HN.
The only ones getting paid this time around had registered copyrights (in the US at that.)
Let’s not forget that Anthropic only paid that to settle a class action lawsuit.
That's like saying someone is a big proponent of community law and order, and they donated $1000 to the county sheriff when actually they got caught drunk speeding in a school zone.
A false equivalence. A more correct example is: Anthropic was speeding, got caught by the county sheriff, and paid the fine. Anthropic stopped speeding.
Meanwhile, Chinese labs are speeding in a different county. Everyone knows they are speeding, yet the sheriff won't pull them over, so they just keep doing it.
This lax enforcement gives Chinese labs a structural advantage over American ones.
After the fact. They did the same thing Youtube, Uber and Airbnb did: Break the law, eventually get caught, cut some deal where they pay a pittance and keep doing the same thing but now with leverage on their side.
Because they got caught
there is much less intellectual property in China so it’s not ‘theft’ (as you can’t put property on information)
How is distillation an "attack" but gigascraping the Internet to the point of crashing servers and everyone needs Cloudflare and Anubis now not an "attack"?
I'm not aiming for a what about kickflip here: I'm saying we need to either agree on some rules or stop crying foul. Maybe the coherent legal theory is that neural networks and intellectual property don't interact. That would be weird but it would be consistent, a market could price it, I could do coding stuff and know if I was illegaling.
But this weird gerrymander that no judge will really rule on in an emphatic way is like, bad for the planet, bad for markets, bad business.
There are a lot of reasons to look forward to DeepSeek Huggingface drop kicking the unambiguous frontier weights in like, November, but I think my favorite one will be "who's distilling now bitch?"
I think you've basically got the legal theory. Training a neural network isn't prohibited by copyright law so if you can legally get your hands on something (e.g. by sending a GET request to someone with rights to serve the contents of their web page, or by buying a book) without signing a contract to not train on it, you can train on it.
But the American AI companies only let you query their models if you first sign a contract to not train on the output.
It's hypocrisy and unfair, but I think there's a strong legal argument for it.
Of course China can simply decline to assist in enforcing that contract... But I would expect US courts to do their best to.
> to someone with rights to serve the contents
Now THAT'S doing some heavy lifting lmao. The vast, vast, VAST majority of the original datasets were from pirated books and the like. Also, arguably a robots.txt is the exact mechanism to follow to do the mass GET-ing, yet the AI cos choose time and time and time again to simply ignore it and be as abusive as they possibly fucking can
> The vast, vast, VAST majority of the original datasets were from pirated books and the like
And there's been significant legal consequences as a result
> Also, arguably a robots.txt is the exact mechanism to follow to do the mass GET-ing
You're free to argue this of course, but the courts have largely rejected it already pre LLMs. See for example hiQ Labs v. LinkedIn
Yes, anthropic and openai have really been brought to their knees and ipos cancelled because of the legal consequences of obtaining their training data.
This would have been a problem but it turns out that Anthropic is actually valued multiple orders of magnitude more than a copy of all the books in the world. So they survived the significant legal consequences.
> It's hypocrisy and unfair, but I think there's a strong legal argument for it.
That right there is the problem.
> but I think there's a strong legal argument for it.
Maybe today. I doubt it tomorrow. Legal and not legal, largely, has to answer to the population sooner or later. Ultimately, humanity decides legality. And I don't think the frontier labs will get a pass from humanity in the midterm, let alone the long term. I think you'll see the rules change towards something more "intent" driven. And there's absolutely no difference in intent between Frontier labs and everyone chasing them.
Frontier labs just want the door closed behind them, as do their investors, because they know the money will never be recouped if others can do the same magic tricks.
Contract law is never going to prevent this.
Why not? Seems like a perfectly normal contact term to me.
Or do you just mean that US courts don't have enough teeth to prevent Chinese companies from violating contracts? On that I agree.
I mean the latter, but more narrowly: China would never allow the United States to have a monopoly on machine intelligence if the only thing standing in the way of a domestic alternative was the Anthropic ToS. In general, I think that China is willing to agree on certain things relating to intellectual property. But not on this, it’s too big.
The US is already publicizing the way they are using Claude with Palantir for war gaming purposes. It’s a matter of national defense. Contract law has no meaning here.
Eh, I think you've done a pretty good job summarizing a collection of settlements with a few narrow bench rulings for seasoning. I'm not sure I follow you to it being a coherent legal theory. Buying a book in a bookstore is sure legal, and excerpting from it for e.g. literary criticism is pretty settled. Downloading every torrent of all e-books ever is pretty clearly illegal (or at least it fuckin would be if I did it). Pretty sure like, multiple labs have been popped for that though.
Situation right now seems more like a fragile detente: if you got a Hill staffer drunk and hounded him long enough he'd probably be like "God damnit the market will fucking tank if we don't get these two IPOs out north of a trillion. And don't even get me started on how I'm going to sell Chinese AI to a Senate that still calls people Nipponesians when no one is looking. We're doing the best we can alright, get off my back man."
We have a situation, but it's not exactly A&M Records, Inc. v. Napster.
> Downloading every torrent of all e-books ever is pretty clearly illegal (or at least it fuckin would be if I did it). Pretty sure like, multiple labs have been popped for that though.
Oh it is, and at least anthropic has paid $1.5 billion and deleted there torrented copies and not released any models derived from them as a consequence.
The thing is it turns out to be not that expensive to just buy a copy of every book legally and scan them. And there's even precedent that this is legal predating LLMs (Google books)
> and deleted there torrented copies and not released any models derived from them as a consequence.
I have a bridge to sell you
Great, let's go down to the courthouse and get some sworn testimony as to the ownership, value, condition, and so on and so forth of the bridge. And some document review and discovery run through professional legal firms under the same conditions. And perfectly reasonable and verifiable explanations as to why you own the bridge and are selling it (namely that you bought a copy of literally every book in existence in the meantime).
Facts are in fact knowable, and the US legal system is in fact not terrible at getting to them.
I think you're right to point out that historically the rule of law in the United States has been very robust by the standards of whatever era, it's been a tremendous advantage in attracting business and capital and talent, it's good stuff.
But we've gone through some pretty weird times too. Turn of the last century was pretty tech billionaire edits, reconstruction was uh, not smooth, it's a mixed bag.
And most takes I hear seem to acknowledge that this is one of those weirder times: serious election fraud rhetoric from most everybody from 2016 to the present, very politicized courts (on both sides to be clear), very soft on anti-trust, very soft on adventurous accounting. The Epstein files and like, no consequences (pretty much uniquely for a developed nation with Epstein people). It's weird right now.
And I think I would be hard pressed to think of a weirder part of this weird time than the rule of law meets AI. We can haggle on where laws end and norms begin (stare decis being maybe the midpoint), but in the 90s, the Justice Department got their brass knuckles on for a lot less.
I don't think it's a simple "the law works nothing to see here" story.
I broadly agree with your take on the state of the US - but this is a case where given the specific facts at hand I'm confident it still got to the truth.
I can understand why as someone who didn't follow it and the more corrupt legal developments closely you wouldn't be confident in that.
Knowlege should not have ownership. Training and distillation should be allowed
Granting people some form of control over knowledge only serves the public interest inasmuch it provides incentive to create more of it. Mass media, effortless duplication, and copyright extensions had already broken this to the point where control of knowledge was suppressing creation of new knowledge more than it facilitated.
The world has changed, we need a mechanism that works for the public interest that applies to the facts as they now are.
American labs have ripped everything out of the internet. And now they cry someone else is “stealing” from them. Cry me a river.
This is nothing like music piracy.
I strongly agree with the premise that distillation is not an “attack”.
But that said: K3 is not a distilled version of Fable or Sol. Fable has been barely available and Sol was just released! Moreover, K3 is superior to both models in some domains, according to user scoring on the Arena.
API distillation can’t give you these results anyway. All it is useful for is bootstrapping RL in new domains to get past the “cold start” problem faster. By far, what matters more is the quality and variety of RL environments the model learns from.
Look how hard Anthropic is to even be able scroll back on your conversation, or look at the thinking tokens or subagents. They want to keep everyone coming back to the watering hole but never to learn how to dig a well.
Did you enable the flicker-free TUI mode?
Why is it hard to scroll?
It really is not, not sure what OP is on about.
Well, there is precedence: Google can scrape the web, but you can't scrape Google. Laws around compiled databases exist for a reason: you can't just copy the phone book if effort has gone into compiling it, it is itself copyrightable
That varies by jurisdiction. In the United States, copying the phone book (or otherwise copying facts from someone else's collection) has been legal since 1991:
https://en.wikipedia.org/wiki/Feist_Publications,_Inc._v._Ru....
And funnily enough, most laws about compiled databases might not apply, for one.
And then there's new updates related to AI that fully take out LLMs from protection.
The fact that API based distillation is even a conversation right now makes me feel like the U.S. has their heads so far in the sand that it’s not really excusable.
These Chinese labs are producing novel models, publishing their techniques and sharing their open weights and the first topic of conversation is how they stole from U.S. AI labs.
Setting aside the fact that it doesn’t make any feasible sense to do API distillation, these models are outperforming frontier models on a number of benchmarks, and often times run more efficiently by several orders of magnitude.
We have to stop crying distillation, it’s getting embarrassing and at this point feels even a bit delusional.
There's little doubt that Kimi K3 was distilled off Claude. When Kimi was prompted to give its name, it answered "Claude" (according to this X post) https://x.com/denisewu/status/2077984660211269870 . This what happens when you include millions of Claude exchanges in your training data.
Anthropic claimed in February that Moonshot AI (the creator of Kimi) distilled ~3.4 million exchanges from Claude models, as explained in their press release https://www.anthropic.com/news/detecting-and-preventing-dist...
or the application layer - which will capture majority of the value.
yeah hardware companies make for nice stories or green numbers on Wall Street - but value will be captured by application layer.
look at history.
That’s true up until the point where you can ask the hardware you made to make its own application layer.
assume you are a "second class lab" and you are in fact making progress by distilling the results of the frontier labs' efforts.
what is the end game for this strategy?
if the frontier labs shut down, or stop releasing to the public, and there's noting left to distill, how will you progress?
This line of thinking makes no sense because it assumes that labs that distill from frontier models are doing nothing else. It's the classic "the Chinese can only copy" mentality, and it's going to end poorly for American companies.
I'm pretty sure that all labs are distilling each others' LLMs, maybe apart from Anthropic and OpenAI. It would be stupid not to do it, because it's cheap and effective. But that's not the only thing they're doing. If you think K3 and GLM-5.2 got this good only from distilling frontier models, you're not paying attention to Chinese labs' publications.
i never assumed that, and i do keep up with the publications. i'm also not saying it's a dumb thing to do! what i am saying is that empirically, it appears that distillation of a more advanced model is a required first step for them to train a borderline competitive, cheaper model. in effect, their training is subsidized by the frontier labs.
if this were not the case, then we would be observing chinese models that far surpass frontier models in capabilities, rather than "almost as good, but much cheaper", and we would be having a very different conversation. what happens to these efforts when the subsidy is cut off?
> empirically, it appears that distillation of a more advanced model is a required first step
I see no evidence for that.
> if this were not the case, then we would be observing chinese models that far surpass frontier models
It's pretty clear that the primary reason for the difference is budget and compute availability. Chinese labs have at least an order of magnitude less money than Anthropic and OpenAI.
> what happens to these efforts when the subsidy is cut off?
They will continue making progress as they do now, minus the benefits of distillation.
https://www.anthropic.com/news/detecting-and-preventing-dist...
Moonshot AI Scale: Over 3.4 million exchanges
The operation targeted:
Agentic reasoning and tool use Coding and data analysis Computer-use agent development Computer vision Moonshot (Kimi models) employed hundreds of fraudulent accounts spanning multiple access pathways. Varied account types made the campaign harder to detect as a coordinated operation. We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces.
I'm assuming you posted that as evidence for the claim that "empirically, it appears that distillation of a more advanced model is a required first step", but I don't think it is. It's just evidence that Moonshot distills Anthropic's models, which, yes, they do.
it is not a required first step for training a model, sure. but that's not what i claimed. what i claimed is that is how they are so significantly _reducing the cost_ of training one! how else do you think they are doing it?
>request metadata, which matched the public profiles of senior Moonshot staff
Translation: we have the machinery in place to identify our users, and actively do so.
In public with budgets that don't risk destroying the American economy presumably. Yes it may be slower.
> with budgets
and what will fund these budgets exactly? inference is cheap, distillation is cheap, training is what's expensive.
Same people who fund linux kernel development. A coalition of companies that find it useful.
Presumably the US military / NSA.
the USG/NSA will fund chinese labs? to what end?
I was more thinking they would be funding US labs.
the question was: what is the endgame for the stated "second class labs" strategy of distilling their frontier competitors then undercutting them on price?
Yes yes, we all understand the game-theoretic race-to-the-bottom you're describing here. Somehow despite linux being FOSS it still powers most of the important computing in the world. Can you explain how that works despite it being free? Once you understand that case I think you'll understand the game-theory behind how large projects can exist in the absence of traditional IP protection.
the obvious difference is the massive scale of data and compute required to develop and evolve these models, and the costs they impose on those building them.
Making lots of money?
There doesn't need to be progress at this point. Some models even from 1 or more years ago are useful for some purposes
Thanks for the models guys, sorry for your losses. Once this reality becomes mainstream and undeniable, surely the bubble pops and then what then. Future model development stops? Becomes private? Becomes a public effort?
The existing models are still going to exist. As hardware improves, there will be a day where it might cost a tenth of a penny to churn through 100M tokens a second of Opus 4.8. Established compute providers will invest in improving the models incrementally when margins drive them to look there.
> Distillation “attacks” are not attacks.
If "distillation attacks" happen, we have to conclude there is some value add in what model labs do. Regardless of how we feel about using existing human knowledge in the way they currently do, it's simply impractical to infer that everything that happens downstream of LLMs can not be an attack on some IP because of it.
So both things can be true: a) People infringe on Anthropics IP and b) what Anthropic did to build their models is legally questionable (or might be ruled illegal, even though I doubt it).
>People infringe on Anthropics IP
No.
Authors do not infringe on IP when they read another's book, nor should the lumber company be able to dictate how I use planks and if I can resell them if i'm done with them.
You're framing it as if the added value of the author or lumber company, awards them consideration when somebody uses the products to create more value.
IP law was always a big mess, and these questions cross far into ideology instead of law; but I do not understand people who think we need an ideology where more IP-law is good for society.
It's more simple: They infringe on the IP by way of violating the ToS. If you violate ToS and the company suffers financial harm, they usually can (usually) sue you in civil court for damages.
Violating terms of service and violating IP rights are two independent violations. Neither implies the other.
That’s not what “IP” means. You’re describing breach of contract.
You can't violate ToS you never agreed to. If I use pirate Claude through a third-party reseller, I have entered no agreement with Anthropic.
Api key?
I guess you could steal them but thats a whole other issue.
>Authors do not infringe on IP when they read another's book
Are the distillers reading books or are they building models?
If anthropic is providing no value they can just build from scratch. But obviously distilling is easier. Hes saying thats the value they add.
There are some quite interesting legal implications here. If Anthropic has IP over output produced by agents, do they somehow have legal rights to code and documents produced by such agents?
This would demolish agent usage by corporations.
General consensus is that neither the model nor its outputs can be protected IP
You say “if”. How did Anthropic obtain this IP, if the model serves ripped internet and all human knowledge?
> People infringe on Anthropics IP
Unless someone literally stole the weights somehow (which is not out of the question, I doubt either oAI/Anthropic have the capabilities to prevent a state-level actor getting those weights), distillation from generations is not infringement on anyone's IP nor is it stealing nor is it an attack. It can't be. As long as you pay for tokens you get to do whatever you want with them. Someone saying you can't doesn't mean it's an attack or their IP or whatever. They either sell the tokens or not. They can decide to not sell them to anyone, but again that's not stealing.
And their ToS are a joke. Imagine how people would react if MS had ToS saying that you can't use MS software to develop solutions that compete with MS. They'd be laughed out of the room. Somehow it's ok for token sellers to decide what you do with the tokens? Why? If you pay for something you get to do whatever you want with that output. Train, distill, whatever.
Its definitely an attack. Thats established from anthropics perspective. No one has a right to use Anthropic’s services in ways that directly violate the ToS and user agreements.
> Its definitely an attack. Thats established from anthropics perspective.
How do things get "established" from someone's perspective, exactly?
By that logic it is established from my perspective that Anthropic has no right to train on anything I've written that is publicly available on the internet.
Of course, they don't care about my perspective, but then again I don't care about theirs.
> People infringe on Anthropics IP
Anthropic’s model outputs contain no IP. This is actually a simple legal proposition (rare in this field!) that derives from the fact that only specific classes of IP exist: copyrights, patents, trade secrets, and trademarks. Examining each, it is clear that API outputs do not qualify. Anthropic disclaims copyright in outputs; the outputs are not patented; the outputs are not secret (a prerequisite to having trade secrets); and trademarks are irrelevant in concept.
The output of Anthropic's models is not Anthropic's IP, as that would destroy their market, if Anthropic owned all the software it generated, and all the content. So distillation, which is just using those outputs is always going to exist.
I'm pretty sure that LLM output is not intellectual property. Nobody owns it, and it can't even be copyrighted. So using output from Anthropic's LLMs in ways Anthropic does not condone is not IP infringement.
Whether or not its legal to distil models, it is obviously morally permissible to do so.
Anthropic, OpenAI, etc do not deserve legal protection.
Considering they were the original infringers, I don't know how anyone can expect tears to be shed here. The best we can hope for is for all these cancerous - and they really are the definition of a cancer - money burning entities to all fall apart to distillation attacks like these.
anthropic model output is not their IP
that would be existential doom for them because then they have a case to claim ownership of their users' codebases
no corporation would sign off on that
The value is simply that it is easier. The same way it is easier to ask someone who has experience for advice than reading hundreds of textbooks.
Regardless of whether it’s intellectual property or it isn’t intellectual property, it doesn’t actually matter. If AI doesn’t stop seeing diminishing returns in scaling up, and it hasn’t yet in the 10 years since the attention/transformers paper, the advent of AI will be the most important development in the history of humanity. Controlling that machine, or at least having one of your own, is an existential problem for nation states. It’s like a matter of national defense.
Do you really think intellectual property laws will prevent this in practice? It’s like as if we said, “hey, USSR, you can’t make a nuke, too! We patented that already.”
Asking China to not distill our models down is equally as ridiculous.
It's unlikely the USA would be granted an exclusive patent for the atomic bomb given the well-established existence of prior-art in the form of nuclear fission on the sun.
(I actually appreciated your analogy, despite my lark)
In fact, China stealing fire from the gods is essential to the future balance of power, so long as they keep making the results freely available.
Anthropic’s IP is basically null and void for how they created it. And they might not want to try and challenge this in court, considering how they had to settle for using text books they had no right to use
>Distillation “attacks” are not attacks. The frontier labs “distilled” all existing human written knowledge into their models
So why didnt we have these LLMs in 2005?
Answer the question "how much does 5 cents of LLM computation in July 2026 cost in July 2005" and you'll have the answer to your question.
Don't forget to account for all the costs. It's not just that CPUs are X times slower. Memory is X times smaller, too, and networks are X time slower. And all this hardware is many times more expensive.
If I'm getting my mental estimation right, training a 2026-frontier-class LLM in 2005 would be somewhere on the order of all the computation power in the world at the time. It's not that many more factors of magnitude before you end up at "all the computation power in the world up to that point".
Is this some form of rage bait? 2005 we hadn't the GPUs, we have today. There are other factors, but I think this is the big one. The mathematics of building an LLM are really old, we just hadn't the hardware to do the needed calculations.
Right. Therefor it's not simply a derivative of information. The hardware is required to build the model. Software as well. The model uses information, it is not "distilled" from it.
"Distillation" literally means to separate and take some components out of something. You can distill how a model works from a model. You cant distill a model from information because the information does not contain the model.
People are happy to conflate distilling with building because they dont like how the information was used. You distill how the model works from the model, and you build a model with information. Both could be morally good or bad but its not the same thing.
Information is information. Why is some information considered different than others in your estimation?
> Information is information
Not really, what the information actually is, matters a great deal. It's harder to get good results going from "nothing > model+weights" than "nothing + traces from known good sessions of other good model > model+weights", this is what the "distillation" part is referring to. If "information is information", you wouldn't even need to separate good from bad sessions while doing the training, which leads to somewhat obvious results if you don't.
Can you be more specific? I have no idea what you are trying to say.
To succinctly restate my point, you cannot distill a model from information because the model is not contained within that information. You can distill a model from another model.
Their point is that "training" and "distillation" are essentially the same. The difference between the words is whether the source material is output from another model, vs being some original text.
That argument is moot as distillation also requires a lot of hardware and software, if copying models was as easy as that, we would have hundreds of competing models.
No. Building models and distilling models both require the hardware and software. It doesn't mean building models is distillation.
Because the transformer architecture that enabled modern LLMs wasn't invented until 2017[1]?
1: That's the "T" in GPT fyi, even though Google is the author of the research paper that changed everything
Right. So we had enough information to train LLMs but not the technology to build it.
So the initial models arent just distilled from information. We’ve always had the information.
Moore's Law or something. Were you alive in 2005? The Nintendo DS getting the Opera browser was a big deal. THAT 2005 with today's LLMs? Hilarious.
We didn't have the compute required (GPUs powerful enough to parallelize forward and backward pass). This compute is what allows us to train from human knowledge or distillation.
because you had neither the chips or the information in 2005. You have probably on the order of 5000x to 10000x more GPU compute today than you had in 2005 and three to four magnitudes more openly available data.
The first "L" in LLM does the work. In 2005 you had no Github, Stackoverflow, Youtube, common crawl and no archive of digital ebooks.
This was always where this was heading, but we got here much faster than expected.
Once western governments declare it to be a "national security" risk for citizens to have access to open-weight frontier models, and once they classify using these models as acts of terrorism, what will that world be like?
Will using Kimi K3 come to be like how napster was in the olden days? Everybody knew it was technically illegal, but come on -- any track at your fingertips? But surveillance is quite more evolved now.
Or it will be like cannabis, where a guy in the neighborhood will low key rent you metered access to the 8x5090 rig in his basement he cobbled together from parts on ebay? Or everyone will flock to VPNs?
Or will the oppressors actually succeed? The same way that napster is long gone, and everyone accepts that they must pay spotify for a homogenized collection, where artists must take only a minuscule cut (more than napster though)... We'll be stuck with nerfed Cohere or Mistral models for open-weight options, as if they need more lobotomizing. Or else we can pay through the nose for Anthropic/OpenAI for "American Frontier" models which will fall increasingly far behind China.
Or else, like how Kindle Fire was subsidized by ads, we'll have "Kindle AI" where influence is sold to the highest bidder, where the LLM will tell us that smoking is actually healthy if big tobacco can engineer its renaissance by turning its lobbying dollars to pay-to-play, pumping its propaganda into the training pipeline for Amazon's extra commercialized line of ultra budget LLMs.
Basically a new iron curtain didving the world into digiatl blocks.The era of open internet/science is on its last legs with the potential forr bifurcation into incompatible ecosystems high , the onger the exchange is disrupted. As recently as this month the USgov has donce a Wolf Amendment style declaration for the Scientific collaboration NSF while shifting its purview under the military. To add to that its trying to rope as many countries into its Pax Silica idea intentionally to exclude China while simultaneosly coercing its 'allies' into using its nerfed offerings [1]
So maybe some isolated switzrland/singapore type locales would exist for US/EUusers to be able to dip their toes across the curtain legally without reprucursions.
[1] https://nitter.net/RnaudBertrand/status/2069574934972797089
At this point, the United States will lose that battle most Countries in the world are going end up using electronics from Asia, that ship has sailed Japan, China, Korea, Singapore, Taiwan, Vietnam, dominate that area, China already dominates EVs, Drones and many other electronic devices, and with the way Donald Trump has picked fights, Europe, Canada, Australia, New Zealand, Mexico and many others are looking for other business partners.
If you need infrastructure done, China is dominating that area too. Rail, High-speed rail, Nuclear reactors, (near future Thorium reactors), Dams, Highway roads, bridges, Ocean ports, airports you name it, and they can roll it out, Transport ships, And if they don’t do it, Japan, Korea, Vietnam, and Taiwan do.
Is it too late? No, not necessarily, but America needs a regime change…
If they outlaw open source models that'll just handicap American companies, the rest of the world will be running open source and have an arbitrage against US companies.
The disappearance of high ram Mac studio rigs is probably just a coincidence, right? :/
even 8x rtx pro 6000 is only 768GB of VRAM. IDK how anyone is going to run k3
This is why God gave us 1.58-bit ternary quants?
1-trit*
Free server racks for everyone when the bubble bursts!
No they will ship it to China as ewaste and then sell it back to us.
Almost free server racks for everyone when the bubble bursts!
I just want an Oxide server at home for sub $1000. Or this, whichever's cheaper.
> western governments
Are you talking about the US, specifically?
Why would other countries, that don't share the same anxiety about China as the US, would be troubled with the this?
Well, many EU countries like Italy and Germany officially freaked out about DeepSeek, ordering it to be banned from app stores etc.
That's about the mobile apps specifically, and it is related to personal data of EU citizens being transferred to servers in China.
It has nothing to do with running open models, especially in hardware within Europe.
The US might pressure them?
EU countries are homeshoring their digital stacks as fast as they possibly can, and the reason is precisely because of the pressure that the US government is exerting via its (temporary) dominance in technology.
Those days are gone. Look no further than the occupant in the White House. IE the Swedish jet industry is about to get bigger, future drone expertise, if the Ukraine can hold on if you want to learn the ins and outs, you don’t need the United States. If you’re serious about learning and building drones.
It’s going to be a different world, a world where many former allies are not gonna look to the United States first they can no longer afford to.
Not sure if you noticed, European countries are distancing themselves from the US. They couldn't be pressured to offer logistic support to the US shitshow in Iran, why would they be pressured to help the US in its protectionism of its AI bubble?
I tried Kimi K3 on a task I've done with every other model I use regularly (https://swelljoe.com/post/i-let-every-agent-implement-its-ow...) and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan.
I only have the $20 plan from OpenAI and the same task, with a lot of the same implementation details as Kimi Code, only took a few minutes and consumed almost none of the 5 hour limit.
Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare, but when I sat down to add Kimi Code to flar, it was because I wanted to try it on some real work and then couldn't do any, because usage was nearly gone after the trivial task...no other ~$20 subscription I have has felt that tight before.
So, it was really slow to complete the task and seemingly much more expensive than every other model I'd tried. Maybe bad luck. Maybe it'll do better on other tasks. I wouldn't know as I was out of usage when I had time to try.
It did find a bug that Gemini 3.5 Flash introduced unprompted, though, so it has that going for it.
Do models know when they're being benchmarked? I also used K3 briefly and noticed it spent a very long time thinking and obviously a lot of tokens.
However I've seen some benchmarks say it uses fewer than fable which hasn't been my experience.
We really need to stop using $/M tokens as the pricing benchmark. I've found that the number of tokens used tends to be a bigger factor than the listed per token price. The cost per task vs. intelligence curve is really what you care about, and in my estimation Chinese models are just not there. They are focused on benchmaxing and getting the highest raw score they can, rather than efficiency.
In my experience, Kimi just tends to think a lot, with the main thing that takes up a lot of space is it constantly second-guessing itself. I've watched it do paragraph after paragraph of "Wait, actually..." while it stumbled and used a ton of tokens on one small detail of what it was asked to do. Though I also gave GLM 5.2 a task to port some JS code to Python to test it, and in my experience it doesn't second guess as bad as Kimi does, but it really did there. It kept doing web searches and second guessing tons of tiny little things, using $0.25 of API spend in total to port about ~50 lines of JavaScript. It did produce an error the first run, but on second run it gave me a program that ran.
I gave Claude Code/Fable the same task and it took significantly less time, but also stumbled on the same error as GLM. I didn't have it fix it though. I was mostly interested in timing differences.
I do like open models where I can, but I'm really hoping they get trained to second guess less. Or maybe I just need to prompt them differently. I'm not sure.
Earlier today I made Claude code implement a feature with fable. It worked roughly 60 minutes and used around 30% of my 100€ subs 5h sessions.
Then I typed /code-review in a second terminal/clean session after the analysis was done (no code changes) the usage was 99%. I then asked it to write that into a review.md so I could restart from that the next day. Sadly the last % wasn't enough for that.
Ymmv, these models behave very differently with no discernable reason. Usually reviews(even with fable) take like 10-20%... Yet suddenly you get it to burn through 65-69% in 15 minutes or so
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
AI subscription pricing is so goofy. You get some amount of usage that varies by models, is measured by opaque token usage, driven by how many tokens the (usually) vendor-provided interface (or model itself) wants to use. Then your usage is limited by time opaque time windows.
AI subscription pricing was fine when it was $100/month for some opaque 5 hour token budget I don't think I ever used, not even that one day where I coded for 14 hours non-stop using Fable. But like most people with low token usage, I had a human in the loop and and I didn't use workflows with swarms of agents.
Now, of course, the plan is to remove Fable from the subscription. To paraphrase Darth Vader, they have altered the deal. Pray they do not alter it further.
They're probably not going to remove Fable. They just extended it for another month.
Gpt 5.6 is still like this at least for the $200/month option. It’s also always faster than fabel. Fabel might be able to do some things better but I don’t have time to constantly wait and find out.
You call it goofy, in a different context we would call that a dark pattern, shady, prone to fraud
Yeah I've noted this behavior with best in class open weight models. They said K3 would have token efficiency improvements and I was hoping especially solving the thinking loop issue that plagued K2.x but even if this release helped somewhat, it looks like we still have a long way to go here... I'm not sure what's up here but I suppose lacking finetuning quality.
What OpenAI in particular have done with reasoning efficiency in the past few months since ChatGPT 5.5 is nothing short of remarkable. It's overshadowed a bit by the benchmark game and the Fable hoopla.
Now is the time to focus less on token cost and intelligence, but tokens to solve a particular set of tasks in closed benchmarks for a variety of categories.
What is the use of grand intelligence if it either costs you a kidney or can't complete at all within a token budget? Even if there are niche uses where you truly want "maximum power" above all, we need to at least more severely penalize such models versus those that does it just as fine within a tenth of the token cost.
I'm aware of some benchmarks at the Artificial Intelligence site, but CLEARLY we are not focusing enough on these today and still leaving the fun surprises to the users.
Yeah, I'm finding I end up switching to Codex and GPT 5.6 a lot lately because I've either run out of Fable usage or Fable refused to do the task. Most recently it refused to work on a WiFi configuration UI for a robot. No idea why it thought that was related to security, biology, or some other sensitive topic. They've hobbled it with guardrails that are overzealous and now there's a big opening in the market. Fable may be the best, but if it won't do the job half the time, it stops being my go to model as I don't want to waste time only to find it refuses halfway through.
Is Kimi K3 subsidized as hard as the other models out there?
From the end of the month it will be served profitably by other providers around the world, like Kimi K2.6
Does it matter? As an end user I really only care about 1) how much I can do in a week, and 2) how long each task takes.
Subsidies would affect 1, but not 2. But if some VC wants to subsidize my Claude or Codex or whatever, awesome.
The more important question than subsidy is what is the tokenomics of running the model. If it's inefficient to run on an nvl72 cluster (or whatever the heck has enough vram to run a 3T parameter model), and k3 isn't very token efficient, then it might not be that compelling of an open weights model.
It doesn't matter if you can switch easily. It might matter if there are barriers to switching.
Subsidization could affect both of those. If you have $200B in the bank you can afford to throw massive compute at every single request; if you are less well funded, you might throttle more aggressively.
Additionally that same VC could be (read: is always) spent on developing the harness, and other infrastructure around the model, not just the model itself.
So it's apples-to-oranges when comparing a relatively new model to established competitors (i.e. OpenAI @ $900B funding vs Moonshot/Kimi's $30B FYI) because every new model they release is judged on "performance" which is not strictly speaking derived solely from the model.
It's possible Moonshot could get similar performance over time as the build out the rest of the infrastructure. We have no way of knowing how much of OpenAI/Anthropic's success is due to the model vs intelligent tooling built on top of it.
Not sure how the economics work for the Chinese models, but DeepSeek did the same task for a dime.
In my opinion, for the vast majority of use cases, DeepSeek is still the most cost-effective model by a mile. $10 feels like it lasts forever.
Yep, with Reasonix, DeepSeek is free real estate. Seems to just go and go for pennies.
And, DeepSeek is what I use for any task that works best with an API. It's cheap enough to where I don't think about cost, made even cheaper by DeepSeek having the most effective and cheap caching in the industry, and it's good enough to where I rarely have to follow up with a more expensive model or manually fix things. It's been alleged they're releasing an update to DeepSeek V4 Pro soon that improves it, which likely makes it a good fit for even more kinds of problems. It remains my favorite of the Chinese models, it's so cheap and cheerful. And, is also less aggressively censored than some of them.
> It's cheap enough to where I don't think about cost, made even cheaper by DeepSeek having the most effective and cheap caching in the industry
I use DeepSeek v4 Flash & MiMo v2.5 Pro. Prefer the latter over DeepSeek v4 Pro because it costs the same while being equally good & less chattier for coding workloads. Although, I've begun experimenting with Hy3 (as an in-between Flash & Pro) & GLM 5.2 (for long-horizon tasks).
Aren't they still locking reasoning to "max" pending adjustments to support shorter reasoning levels.
It let's me choose different thinking levels in Kimi Code. Not sure if it actually works, yet, but it says "Thinking set to high." when I change it from max.
Why would they do that? Sounds terrible
Absolutely do not pay for the kimi plans thinking they will be cheaper. If you sign up with a Chinese phone number, you can get the same plan for 200 yuan instead of 200 usd, it also only accepts Chinese payment methods iirc. So the plans are really made for Chinese userbase.
That sounds complicated. I'll just use my month of Kimi and then cancel. I have too many AI subscriptions to use them all, anyway. I subscribed mostly to test it. I mean, if it turned out to be competitive, I would keep it, but if it doesn't turn out to really excel and anything and also take longer than Claude or OpenAI models, I'll stick with them.
It is complicated, but paying for the cheaper usd plans really don't get you much usage.
Wow! Does it accept a foreign alipay/wechat pay account?
Those exist?
No idea lol, didn't even know those exist..
how do we get chinese phone numbers?
esim
Nah that won't work. I don't know tbh, I just used someone else's number.
It would be really interesting to redo the public benchmarks for kimi k3 but token normalize the costs. Ok so maybe k3 beats fable on terminal bench, but how many tokens did it use?
> tried Kimi K3 on a task I've done with every other model I use regularly and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan
ArtificialAnalysis puts Kimi K3 just below DeepSeek v4 & GLM 5.2 in token use per task, which is about 2x to 3x more tokens than Grok 4.5: https://x.com/ArtificialAnlys/status/2077832879187620192 / https://archive.vn/zBbFi 2 other open weights MiMo v2.5 & MiniMax M3 are comparatively thrifty.
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
I always put my coding subscriptions (that allow it) through "AI gateways" (Cloudflare & OpenRouter are free) which help track token use.
In my experience, Kimi & Qwen Cloud have opaque & restrictive limits, their "credits" drain faster. I now make it a point of subscribing (directly [0]) with providers that are transparent like MiniMax, DeepSeek, Xiaomi, & Z.ai.
[0] OpenCode Go, Cline, and AtlasCloud have generous limits for open weights, otherwise.
With the obligatory disclaimer that I’m impressed with what open weight models can do, I have the same experience with all of them.
The benchmarks come out and say they’re as good as Opus from N months ago, then I use it for a complex task and it doesn’t work as well as Opus from N months ago did when on similar problems.
There’s a real wow factor when you get an open weights model to do amazing things, but in my experience the gap to the frontier models has always been bigger than the benchmarks would lead me to believe.
There can be a lot of value in having the cheaper open weight models for chewing through lower complexity tasks (non-programming in my primary use case) at a cheaper rate than OpenAI or other frontier API costs. Even with those I can measure bigger gaps to the frontier models than the benchmarks suggest.
If the benchmarks aren’t being directly gamed, there’s at least some selection happening where training data or model structures are being picked in ways to maximize public benchmark performance. All of the labs know there’s immense value in having good benchmarks to show for your model because most LLM consumers are picking based on lab provided benchmark charts, not running their own evals. Running your own evals is hard and expensive.
OpenAI measures token efficiency. Look at the API cost charts in their announcement: https://openai.com/index/gpt-5-6/
I think it's the opposite.
Kimi K3 has 2.8 trillion parameters. We don't know the number of parameters of ChatGPT 5.6 or Opus 4.8, but it's probably in the same region. Fable/Mythos are rumored to be around 10 trillion.
So, K3 is directly comparable with ChatGPT 5.6 and Opus 4.8, and the price is not so much lower:
K3: $3/$15 per 1 Mtok input/output ChatGPT 5.6 Sol: $5/$30 Opus 4.8: $5/$25
This is not a watershed moment. It's a competitor converging to the same capability and trying to undercut your prices, but not by a lot.
As for the open weights? For now, Kimi K3's weights are closed, and I don't expect the situation would change.
> As for the open weights? For now, Kimi K3's weights are closed, and I don't expect the situation would change.
It'll change on July 27 (based on https://www.kimi.com/blog/kimi-k3):
> The full model weights will be released by July 27, 2026
Even in this very thread the feedback on Kimi's actual efficacy is debated. I personally feel its worse than both Fable and 5.6 Sol, but I feel like the conversation isn't really about whether its good or not, but a backlash against the U.S governments foray into regulation. So I think people _want_ it to be superior out of anger/frustration with the current situation.
This seems like a replay of what happened with DeepSeek. They put out v3, or whichever one it was, and everyone said it was over for US companies... then everything continued on.
It's really good. I'd put it between Sol and Fable. I'm not super impressed by Sol's UI design skills, something K3 is strong at. Fable is still overall the fastest, most consistently well-performing model, though.
This does depend heavily on the kind of work you do and how you use these models, but the idea that K3 isn't right up there with US SOTA models doesn't match my experience.
That would make sense, what the US government has done this year with regards to AI is unacceptable
Agree completely.
When you net out across benchmarks and firsthand reviews it seems like it's maybe a little behind. There seems to be a consensus it's token hungry and a little slower. So maybe it's a point release behind.
That's weeks maybe months behind, not months maybe a year behind. It's "would my life really change if Claude was gone, not really" behind.
I actually haven't used it much, because Claude started kicking ass again the last few days. Like, way too much of a difference to be normal load-based variance. I got more done in the last 48 hours than week before that.
So, fuck yeah competition.
> I’ve been running Kimi K3 alongside Claude on my normal coding work, and for all practical purposes I can’t tell them apart
When you say "Claude", do you mean Opus? Fable? What effort level?
This line made me think by 'normal coding work' the author means doing something they don't understand well enough to be able to distinguish the models' output.
Steve Yegge calls this is the "discernment horizon" - https://steve-yegge.medium.com/the-flat-curve-society-36c8b0...
This is comparing Fable High with K3 High. I'm mostly using these models for game development. The tasks I usually send are ambiguous visual bugs, changing the look of a scene or models, or adding a large feature. The wording wasn't accurate there. I don't use Fable or K3 most of the time. I'm usually working on smaller scoped tasks that I review myself afterwards.
It was all distillation up to this point anyway. And I agree with what Suhail said on twitter: "Make the margins next to zero for all these AI models. It was trained on humanity's data, it should be gift to ourselves. Doing so will save us from a few in control of our species."
Pricing is actually far cheaper than that. There's two tiers of pricing: Chinese and US.
If you sign up with non-Chinese phone number, you're bucketed into US, you get US prices, can pay only in USD and with American credit card network.
Chinese prices are about 9x cheaper than the US prices, which are already far cheaper than Claude or other American provider. If you can somehow get hold of a Chinese phone number, keep in mind that you can save ~90% of the bill.
It’s 100 yuan per million output tokens in China. That’s $14.7 USD - not “far cheaper”.
He's talking about the plans, you are talking about API prices.
Most of this hand-wringing on price will go away.
My assumption is that Anthropic, OpenAI, Kimi, etc all have a similar cost structure when serving models. The same size model roughly generates the same GPU usage whether you’re American or Chinese. I’d also guess that the model sizes across all SOTA models is similar, we just only see data for open models. The difference is most likely that American companies simply charge more because they have the dominant market position.
Remember not too long ago when Anthropic was charging $75/mt for Opus? Now that many models are in “opus tier”, their pricing is $25 - higher than competitors but close. The newest Kimi is $15. 40% lower to forgo “made in America” with American enterprise support staff is not crazy. Compare AWS to Hetzner or any other flagship enterprise service to the foreign and discount option. I assume that over time, we’ll see the commodification of models reducing prices even towards the raw GPU costs.
Well, there is the small issue of privacy policy: Kimi will train their models on your interactions if you use their subscriptions, and only with direct API usage (billed at API prices) they say they won't. Whether you trust that is another matter.
Those things do make a difference to some of us, even though nothing is black and white. In my case, I'll probably want to wait until other providers appear through OpenRouter and then I'll try to judge how much I trust them. But even if I don't trust them much, they don't train models anyway, so the likelihood of my data being used that way is smaller.
> Kimi will train their models on your interactions
I find these kinds of concerns increasingly silly: most of the input to these models will be ... previous output from the very same models, alongside the occasional half-assed human command to fix something and "make zero mistakes". Who cares if they train on that? Let them, if it makes their future models better!
99% of users are not working on any special IP to worry about that.
Indeed, the B2B / no-data-retention market is still going to provide plenty of business for American companies even if every hobbyist uses open-weight models.
Exactly; this is a no-go for me, I will wait for an independent provider to sell the service, which is possible thanks to the open weights.
I can see the economics of open vs. frontier models turning out similarly to pharmaceuticals, where generic drugs cost a fraction what the name brands do and Americans end up paying the highest prices in the world partly as a consequence of propping up drug discovery research.
Kimi K3 is really good, but it’s obviously worse than Fable, usually worse than Opus, in my experience.
Definitely not my experience. Fable is better but I'd prefer K3 to Opus based my experience with both.
That’s not obvious to me at all. Especially your claim around opus.
Agreed. I think benchmarks are pretty much right in pegging it somewhere in between Opus and Fable.
I think the biggest problem with Chinese models is that they seems to overthink for most of the tasks, especially for smaller ones. The OpenAI models have in my experience only gotten better in terms of efficiency.
Yes, this (imo) is a clear result of benchmaxxing. You can get a much better score on most "intelligence" benchmarks by massively over-saturating reasoning. This looks good on those, but for actual daily usage makes the models much less effective: I don't want a model I use for coding to burn a bunch of reasoning (read: time) on trivial tasks.
It's undeniable that some of these models generate a ton of thinking tokens, but it's arguable whether that makes them "much less effective."
For example, Kimi 2.7 has been really effective for me despite having verbose thinking blocks, simply because it runs so fast. Speed-wise, it feels about like Sonnet, possibly faster.
I strongly suspect the flip side is that in the future it enables you to train smarter models by "distilling" the end result of the super duper heavily thinking models.
But these models already distill the smarter American ones ;)
It's turtles all the way down
The current administration's immigration policy isn't helping. This wouldn't have happened 10 years ago because the US was this city on the hill that everyone wanted to immigrate to. Talented Asian researchers would have immigrated to the US and China would be deprived of talent.
The visa that would correlate to this is the O-1 visa
20k O-1 visas were issued last FY which was mostly under the Trump admin, up from 19.5k the previous FY under the Biden admin
No it is H-1B visa. Right out of the university it is hard to recognize extraordinary talent. People like Sundar Pichai were not recognized as extraordinary right out of the university, he had to start at the bottom and rise up the ranks.
Melania got a EB-1 "extraordinary ability" immigrant visa
To be fair, that was clearly well deserved. Marrying Trump and then becoming first lady is definitely an extraordinary ability; I doubt I could have done it.
This makes even less sense, Trump admin has been here for 1 year, the implication here is a university grad on H1-B in January would become a world class researcher capable of building a frontier model in <18mo
This is why VC is actually hard. Everyone’s instinct is always “Man, once the company has demonstrated it’s awesome I would love to have been in the seed round”. The tendency to want “proven performers” is the default belief.
When people demonstrate their capability thoroughly, the Chinese government takes away their passports. You’re not exactly going to get them here with an O-1.
This admin and its policies on immigrant visas have been around for 1 year and Biden was famously pro immigration
The O-1 has also been abused for a long time, basically any software engineer kid who gets into Y Combinator has been getting an O-1
This is essentially the point of the visa, it feels wrong especially as YC drops standards and increases cohort sizes, but the same power laws that keep them winning also apply here in maximizing economic value of each O-1 approval
Basically of all visas O-1 is virtually guaranteed to have highly positive economic value
Your comment feels like an outdated brain drain model where talented Chinese researchers naturally want to leave China and the only question is whether the US lets them in.
That may have been closer to reality 10-20 years ago, China is a different country now, what I mean by that is they offer research funding, they have huge digital behemoths (alibaba, tencent, huawei, bytedance etc), large scale deployment opportunities and prestigious careers. Many graduates return because the opportunity set is attractive and they want to return, it's not just because US immigration policy pushed them out. Some also want to contribute to their own country's technological progress (which is a normal motivation btw), like probably you are also a patriot and want your country to succeed.
So, really, China's AI progress is not mainly the result of America failing to absorb every talented Chinese researcher. China has built a domestic ecosystem capable of producing and keeping top talent itself. I feel like a lot of Americans do not understand this.
Chinese students still want to attend US universities [1]. While it is true that the progress made by China is a factor, this administration's policies are the bigger deterrent [2] [3].
[1] https://www.latimes.com/world-nation/story/2025-02-21/why-ch...
[2] https://www.wsj.com/world/china/americas-allure-fades-in-chi...
[3] https://www.theguardian.com/world/2025/jun/06/chinese-studen...
It doesn't really address the point. Chinese students wanting to attend US universities is evidence that US universities remain attractive, not that those students would otherwise permanently immigrate to the US or that China lacks attractive careers for them afterward.
US immigration policy may be unnecessarily pushing away talent but the assumption that talented Chinese researchers would naturally remain in America unless prevented from doing so ignores the growth of Chinese universities/labs, companies, their funding, national prestige etc.
I mean, don't get me wrong, US is still highly attractive, it is just no longer the only place where an ambitious Chinese researcher can do important work and grow.
China never allow US AI in China, so they HAVE to build Chinese equivalents...
US immigration policy isn't a big factor.
China's got 1.8B people. If you don't think they've got the talent to pull this off, even if a lot of it leaves to live elsewhere, you're naive.
No one uses Baidu, but they built their own Google, and it's good.
They built their own Facebooks and Instagrams.
The US isn't the only place in the world where people can build software...
"China can draw on a talent pool of 1.3 billion people, but the United States can draw on a talent pool of 7 billion and recombine them in a diverse culture that enhances creativity in a way that ethnic Han nationalism cannot." --Lee Kuan Yew, former prime minister of Singapore.
- thats not a sustainable strategy
- china’s homegrown tech industries already achieved escape velocity from it a long time ago, after China fenced off its market for Alibaba and Baidu in the ‘00s. some of their AI innovation at the edges was already top class 10 years ago
It has been a sustainable strategy for the tech industry for decades.
Claude is not reliable anymore with their sudden Fable access drops etc tbh and I am happy there are good alternatives coming out
We're having so many moments! Every day a new moment.
Terms of use are very broad and not friendly for most things.
Can’t use for commercial purposes. Can’t opt out of training. Data retained.
Correct: can't opt out of training. This is well documented.
"Can't use for commercial purposes" - incorrect AFAICT. In what sense do you mean this? The open weight MIT version obviously allows for commercial use, but I don't think that's what you're referring to, because training data is irrelevant on the open weight version. Pretty sure the API allows commercial use too. Maybe the free version doesn't? But who cares?
GPT 5.6 Sol comes out ahead of Kimi K3 on price/task (but not significantly so). You're probably thinking, "Why use Kimi K3? Isn't an open model supposed to beat the closed one on price?", but you need to consider that the closed models are completely hobbled when trying to do anything security-related. For my use-case, I can't risk getting pwned because I'm using a model that refuses to secure my app while there is now an open model that obliges to obliterate any app that isn't protected.
A lot of these open source models do look good on public benchmark but not sure if they are that trustworthy with production workloads.
Is anyone using open source models for anything major ?
According to OpenAI's "head of strategic futures":
1) Kimi 3 is a "very good model"
2) It's performance can NOT be explained by distillation
3) The US government should create FUD to stop US corporations from using it (so they use OpenAI instead)
https://x.com/deanwball/status/2078133895766114412
If Kim was "distilled" from Claude, how much were the token costs, assuming Kim got everything it can out of Claude?
I never truly understood what the intended business model around LLMs was. Get them widespread through cheap pricing and then jacking it up? Being the only ones that had a viable product so to get the ability to extract as much value as you want from AI?
I don't understand how a product that:
- is interfaced with and is deeply linked to natural language, so everything you produce (sessions, history, etc) is in Markdown and you can literally install a second model and tell it "hey import all of Claude's memory into yours" and that's it
- is based on well understood technology, the real constraints are how much money you put into training the models, but the theory has all been developed in the open
- clearly has a threshold where it quickly commoditises and turns from "I want the best" to "hey the best is a bit too expensive. The second best is half the price and works close enough".
was ever supposed to be a money printing machine. The fact something is extremely useful doesn't imply it's extremely profitable.
IMHO we're clearly speedrunning the process of turning AI into a commodity. Dario Amodei knows pretty well that when or if Anthropic cuts people off Fable, the vast majority of them will definitely not pay for it because Opus 4.8 is good enough for almost everybody that _knows_ what they're doing, and so are basically half of the most recent models. If I already have good baking skills I don't become more productive with an automatic bread machine, I just need a better dough mixer and oven
There is no business model. That’s not a joke, the idea is to be the one that survives the race, then figure out how to be profitable. If you look at the level of capex and money raised, that’s not something you do if you have an actual business plan. We are very far from business fundamentals
> I never truly understood what the intended business model around LLMs was.
A closely related question is “what do the American labs need to do in order to justify their enormous market valuations?”
It seems like the answer cannot possibly be “gradually improve model capability while figuring out how to better monetize inference.” The valuations are just way too high for that to be sufficient.
Surely the answer has to be “continually achieve large leaps in capability comparable to the first consumer releases of ChatGPT while also maintaining a significant capability lead over open models and new competitors.”
And does anyone think that’s going to happen? Even with state-level protection from competition (which incidentally would significantly harm the American economy), the large leaps in capability seem to be coming fewer and farther between.
> I never truly understood what the intended business model around LLMs was
What appeared initially to be a huge innovation was later easily duplicated by many. There are no platform-lockins or network effects. Switching costs for users are zero, and there are low barriers to entry, with vast numbers of models to choose from and more appearing every day. As a business a token will be a commodity like an electron. Doesnt matter who produces it, or how (solar, wind, coal, nuclear etc) as long as it powers my toaster.
It's fairly simple. Sell GPU compute + extra margins as only some GPUs can load the models + extra margins based on how much better closed source models are from open source ones + hopefully reduced cost due to batching
It seems like the endgame is to amass absurd amounts of hardware and produce something that will replace you the baker entirely
everything else we see today is just preparing for it.
Same business model as always: build cool tech because it is cool and figure it out later.
The valuation is based on one lab getting a decisive first advantage, and turning that into a durable self-improving advantage that can never be caught up to. If any can pull it off (a gigantic if), they will effectively own most AI value, and the people who own their shares will live happily ever after. Divide your investment between the labs that could plausibly do this, and your EV may not be dreadful.
This is clearly not how it's going though. Any advancement from any lab has been quickly (< 6 months) matched up by basically everybody else. Even Grok nowadays is decent, and that's something. When something like you've described actually happened historically you generally had quite fast a clear frontrunner and a bunch of copycats that failed miserably; in 2026 we are very far from that. we are heading face-first into towards a pricing war because all models are easily interchangeable nowadays - AI is turning into a commodity more or less
And then you wake up from the dream…
Half-OT: can anyone recommend a LLM cost calculator that's up to date?
https://deepswe.datacurve.ai/ or https://artificialanalysis.ai/ pareto frontier graph.
Thanks!
What is the parento frontier?
If you have multiple metrics to evaluate goodness of a design, one would normally need to decide which metrics they care the most about in order to find the "best" design.
The Pareto frontier tells you which designs are the best in at least one of your metrics (non-dominated by another design). For example if you're selecting a car and you care about both speed and mpg, a Formula 1 car and a Prius might lie on the Pareto frontier, but a Model T Ford would not.
The set of models that are pareto-optimal, IE for some set of variables, no other model strictly dominates them = no other model is better than them on every variable.
So like, on a cost-intelligence graph, the cheapest and most intelligent models are pareto optimal. Then in-between those if you have
- cost $3 intelligence 6
- cost $1 intelligence 5
- cost $2 intelligence 4
The 1st and 2nd are pareto optimal, the 3rd is not, because it's dominated by the 2nd (2nd is cheaper AND more intelligent at the same time)
try PARETO
considering token efficiency as well I presume?
I'm struggling to decide whether I feel comfortable sending my data to these Chinese models
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.
How do we know that Chinese models would not do the same? What makes you so sure that China is less likely to abuse my data?
It’s not that the Chinese firms are any less likely to misuse your data, it’s that you don’t live in china, so their abuse of your data is unlikely to directly impact your day-to-day life in the same way
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.
really weird that they would download every SF-86 file the government had and Equifax credit records of every American then.
Are you comfortable sending it to US ones? Especially if installing Claude Code or another tool on your PC and it can collect all the data it wants..
On Openrouter Kimi K3 says it does not retain data or train on it, which is better than what US hosts claim for Claude, ChatGPT, etc.. as they collect and retain data even if you disable training on it.
Opencode or similar open source tool + a zero data retention provider is about the best option aside from running a smaller fully local model on your own PC.
For open weight models, you can choose from a few providers. Each have their own caveats, none of ToS'/Privacy Policies I entirely trust, nor do many make renewable energy claims.
> I think I can see where this goes. The government will try to regulate AI and open source in particular, and it will run the playbook it ran for the auto industry. Decades of subsidies, bailouts, and protective tariffs produced American carmakers that sell trucks at home and barely register anywhere else in the world.
Here's the thing about this though, the auto industry directly employed hundreds of thousands of people.
The AI labs are small, only few benefit directly from their wealth and there's already immense opposition to AI, data centers, etc...
20 years ago we used to pay a lot for things that are now practically free. I don't think AI is an exception.
We also used to get for free things that people now routinely pay for. Remember maps mash ups?
Has anyone tried Kimi K3 against gpt-5.6-sol on real projects?
I cannot imagine wanting the product of the entire intellectual output of humanity since recorded history to be an expensive, paywalled commercial product owned by 5 or so of the most insufferable, detestable people who ever lived. Why would you want to live in that world. We should all be rooting for open source AI to win.
USA GOVT. BE LIKE: BAN ALL CHINESE LLM
But damn Moonshot.
In my experience GLM 5.2 is a pretty good Opus replacement. But K3 has not given me an experience on par with Sol or Fable. The price/intelligence ratio might still make sense. But it’s not very inspiring when it comes to my real world tasks. I’m doing pretty mundane web stuff.
The dumb efforts by the US AI industry to use fear mongering for regulatory capture will hand dominance to China and others.
In a few years there will be Mythos level open weight models hosted by the lowest bidder anyway.
> In a few years there will be Mythos level open weight
At the rate things are moving I'd expect that to happen much sooner.
In fact: Somebody, right now as we speak, is most likely already working on training the next best open source model.
I just thought about that recently too, then Kimi K3 came out, and I thought: Yea, I'm not surprised. Just a matter of time now...
Related:
Kimi K3: Open Frontier Intelligence
https://news.ycombinator.com/item?id=48935342
Kimi K3, and what we can still learn from the pelican benchmark
https://news.ycombinator.com/item?id=48947717
Its worthwhile to have a quote from the article as some comment without reading:
"...I’ve been running Kimi K3 alongside Claude on my normal coding work, and for all practical purposes I can’t tell them apart. Same tasks, same quality of output, and near identical token counts to get there. I expected an open model to be sloppier or to grind through more tokens on the way to the same answer, and neither turned out to be true.
The prices are nowhere near each other. K3’s API runs $3 per million input tokens and $15 per million output. Claude’s top model costs $10 and $50 for the same units. The subscription side is even more lopsided..."
Was this written by Kimi K3 or Fable?
Doesn't read like AI writing whatsoever.
I don't know. You can't just rely on looking for em dashes or other obvious tells because anyone who cares can get the AI to avoid those.
> When the headline model on your plan can be switched off because the economics don’t work, the plan was never really selling you the headline model. Kimi’s tiers don’t come with that asterisk.
This line has a certain smug, punchy cleverness that I associate with AI. To me, the vibes are ~30% AI writing.
> The prices are nowhere near each other. K3’s API runs $3 per million input tokens and $15 per million output. Claude’s top model costs $10 and $50 for the same units.
And this is the point where your internal compiler should have started shouting 'Type Error'
Notice the trick here?
> Then there’s the fine print. Claude couldn’t sustain Fable access on the twenty dollar plan, so they turned it off, and the plan quietly falls back to Opus.
Where is the Fable-class Kimi model at all?
> Where is the Fable-class Kimi model at all?
6 months away
People pay a premium for the best.
Try asking Kimi about Tianenmen square...
Ask Claude whether a man can be a woman. No matter what political side you're on, models are censored. I'd argue American models are a lot more censored/biased on a lot more topics, and especially on things that come up a lot more than some highly specific Chinese politics. Claude even refuses to translate song lyrics because of copyright.. I wish we could simply have uncensored models from all sides, but it's pretty clear that's not happening right now.
i just prompted Kimi & it replied with an uncensored version.
i posted the transcript as a reply to you, and HN automatically flagged it.
but go on.
I can't recall the last time that it was useful knowledge when writing code. Reservoir sampling, online softmax, Otsu, sure, Tianenmen square not really.