Why is it a problem that the Chinese labs are just distilling down Anthropic’s models? Aren’t Anthropic’s models not just distilling down other people’s work?
Feels like Anthropic crying do as I say not as I do.
What Anthropic is doing requires way more resources than what the Chinese labs are doing. So their complaint is that they do 95% of the work and the Chinese labs do the last 5% and call it their own.
An argument can be made that Anthropic is also only doing the last 5% of the work (because the content they are training on was the other 95%) but that's a bit more philosophical.
Yeah, the whole thing seems like a human centipede of rug pulling. Probably the same as it's always been. Curating AI knowledge should be something that we put our best researchers towards, but realistically I think we wind up with 2-3 highly biased nationalistic models that are constantly copying off each other's notes.
Thanks, that’s the nature of any business. Founders see a way to take existing knowledge and expertise, combine it in some novel or interesting way, and produce a new product
Napster was a fantastic and disruptive product, the likes of which arguably has no equal to this day. But eventually the hammer came down from the courts and it was replaced by streaming services like Netflix, which pay to license materials from their creators.
But if you take it deeper, didn't most of those authors rely on the work of others? Most of human knowledge is small advancements of things we already knew. Often by reorganizing what we already knew.
Is that not what the foundation models are? A new reorganization of existing knowledge?
So then the problem is that Anthropic seems hypocritical when they knowingly insert themselves into this chain, and then complain about people down-chain from them.
To remedy the negative impressions (if they even care to do so) they should do 1 of 2 things:
1) stop complaining about it
2) stop distilling other people's work
They insert themselves into this chain for profit and complain about it. I really think that adds a thick layer to the hypocrisy that people, or at least me, feel is especially distasteful.
No LLM products would exist without the avalanche of largely non-consensual use of IP to create them, full stop. Any of these companies doing this and then turning around and complaining when their IP is "breached" are going to met with a chorus of tiny violins.
It is an ancient practice, that when a human creates something, other humans will observe it and learn from it. Every group of humans living together has practiced this in some form for tens of thousands of years if not longer. Even animals do it. It's a natural assumption when making any form of art.
It is not a natural assumption that someone will digitize the artwork and use it to adjust a couple thousand matrix coefficients in a complex computer program. To most people that seems like copying with extra steps. The brain may in some ways resemble a computer, but what sets it apart is that we have always lived with brains. Everything a human does has already anticipated the presence of other brains, while etched circuits on ultrapure silicon crystals are something new.
To a degree. The human produced knowledge is the product of all humanity (no human is an island).
A comparable idea could be that an encyclopedia is only distilling the things that other people did, and how dare they sell them. But the "only" is doing quite a bit of work. LLMs do not just spawn into existence. There is a body of work that they feed on, and then there is also very attributable work they do around and on top of that. All labs are struggling around the first order question: Is it okay to use prior work like this? The second order issue is still entirely reasonable to separately have and enforce rules about.
SOTA models cost hundreds of millions to train. Did creating the contents of the text corpus they were trained on really cost an equivalent of 20x as much (~10 billions)? I honestly don’t know, but I could imagine it having been significantly less.
This isn’t meant as a moral argument, just musing about the relative cost comparison.
If you look at movies alone that would easily surpass 10s of billions. The cost of most books is probably more nebulous, but books, research, and more all have time and money spent to create them. I would guess the corpus of all media from the 20th century on would be minimally in the hundreds of billions of dollars.
I would argue that producing the complete written corpus on which they at least intend to train (even if some is still out of reach) cost literally everything to produce.
And the monetary cost doesn't even register when weighed against the blood, sweat and tears that went into capturing the authentic experiences of real human beings, whose honest expressions are now at least in some cases getting hoovered up, ingested, and then destroyed for all eternity, for fear that this specific work is the rounding error that might give an equally immoral competitor the edge in the bicycle-riding flamingo race that is currently consuming an absurd amount of the world's creativity and attention.
I get that angle but it’s a weak argument as Anthropic is doing the same to others. Also while there’s certainly a lot of computing power needed to do what Anthropic does, it’s increasingly clear there isn’t much secret sauce involved. Everyone knows how do to the core work it’s just a question of who wants to burn billions on compute to do it.
Anthropic’s anger here seems mostly rooted in their annoyance that this exposes they don’t really have core IP that’s not just easily replicated. And that’s clearly a problem for a deeply unprofitable company trying to convince people they’re worth $2 trillion.
Distillation doesn't "grab 95% of lab's work", that's ridiculous. At best it's icing on top of the cake that's already there. It's not even necessarily done on a better model (e.g. GLM 4.7 distilled Gemini 2.5, a weaker model), I'm pretty sure A\ and OAI could do (or even do) the same with greater efficiency since they have access to logits, weights, and internal state of open models.
>An argument can be made that Anthropic is also only doing the last 5% of the work (because the content they are training on was the other 95%) but that's a bit more philosophical.
How is this philosophical? They should release the unsupervised pretrains, at the very least.
And the communal work of humanity is orders of magnitude more work than what anthropic pays for their scraping of content. I got no check from them for my contributions
Indeed, if it isn't a crime to train on humanity's data, it isn't a crime to train on capitalism arranged frontier LLM provider models. Is that bad for shareholders and capitalism? Meh, sounds like a suboptimal socioeconomic systems issue. Burn up all the capital the unsophisticated are willing to provide. “We are selling to willing buyers at the current fair market price.”
With my apologies to Brewster Kahle, "Universal Access to All Knowledge." [2]
I use crime in the broad sense of "You shouldn't be allowed to do that" in this context. If you have a better word to capture that thought, let me know, I'll make the edit ("frowned upon" perhaps?). I don't have strong feelings other than "hah AI companies aren't going to be able to create a moat to capture the value they want to capture because we can collectively keep pulling it out of their models in perpetuity through ever improving model distillation methodologies". This is no different than Uber and DoorDash using VC dollars to subsidize services until they try to turn the knob to profitability once they've captured the market, except in this case, there are mechanisms to exfiltrate the model value into open models that can be distributed at very small marginal cost. They can never gate the golden goose money printer.
> What Anthropic is doing requires way more resources than what the Chinese labs are doing.
Oh that’s very sad.
Meanwhile Anthropic made a product from the work effort of millions of people without compensating them, sell that product on tap and unless I am mistaken do not even have their competitors’ cover of having released any sort of meaningful open weights model.
They have taken from culture (including very specifically their most direct customers’ specific culture — our culture), turned it into a machine to make themselves rich, appear likely to predicate their valuation on permanently removing people from the workforce, then want to dump themselves onto pensions funds and ordinary savers to carry the bag.
It is, I agree, philosophical, because karma is a philosophy as well as a bitch.
But that’s not more philosophical. It’s a perfect parallel! Enormous amounts of work, vacuumed up and resold. What’s the difference? If it’s ok to vacuum up all the knowledge in the world, then that includes knowledge of how to use all that to power an LLM.
It is kind of ironic that they scraped the web for publicly available data and used it freely to train their models and now their freely available models are being used to train other models.
I'm overall pro-Anthropic and pro-banning open-weights AI, but I agree with the parent commenter; distilling Claude models is not that different from pretraining on web data. It's all basically the same sort of thing.
I think a good litmus test here would be if Anthropic were to not care about distilling their models when the distillers keep the resulting models closed-source and sell tokens via an API. If they cared only about security concerns and not about people profiting off of their work, then they should be publicly fine with this and only protest against it going into open-weights models.
From one perspective, the 5% estimation is near-infinite orders of magnitude off, since they've trained on something approaching the sum-total of human knowledge.
I'd split it at 99.98% original, 0.015% Anthropic, 0.005% Chinese, and that's being exceedingly generous to the AI companies, there should be several more 9s and 0s in there.
The conversation here is mostly moral and ethical but the problem here seems to be financial.
Anthropic and OpenAi are spending a $$$$ to "distill" human output into an AI model, then others are spending $$ to distill their AI model into a near-equivalent model.
This is the same reason IP rights exist. On the surface, something like a patent feels ludicrious and even feels morally wrong. Some guy wrote down the recipe for arranging atoms or bits in a particular way, and now I can't!? However, it's designed to solve the same problem, figuring out and describing the process is much more costly than replicating it.
AI training, if viewed through the capitalist mindset, is plain theft. Anthropic can't morally defend copying someone else's IP, but denouncing others copying Anthropic's stolen IP.
That doesn't mean they won't try, and that also doesn't mean they won't succeed.
Anthropic and OpenAI are very happy to ignore the IP rights of others, so I'm not sure how they can ask for any kind of IP protection themselves. Live by the sword, die by the sword.
Model weights wouldn't be covered in a patent. You could patent a method of creating weights in a model, but you couldn't patent the weights themselves.
I wish people here could at least bother to inform themselves about the IP rights they are so quick to insist are abhorrent, when they seem to not even have a first clue as to what they actually cover.
>Aren’t Anthropic’s models not just distilling down other people’s work?
Can you elaborate on that? I mean my direct answer would be no, of course not. But why do you think frontier models are distilled? I think maybe there is an equivocation over the word “distillation.”
Frontier labs train on their own pretraining data, human feedback, synthetic data, and research. A distilled model is specifically optimized to reproduce another model's behavior.
Meanwhile R1-Distill-Qwen-32B was distilled from DeepSeek-R1.
If you want to say a frontier model is "distilled" from the world's data and R1-Distill-Qwen-32B is distilled from DeepSeek-R1 then you are equivocating two very different things.
Because cost of original training >> cost of distilling. It's the same thing that happens with Chinese knockoffs of physical products - it takes a lot of money and R&D time to design a new product, but it's basically free to buy the product, reverse engineer it, and resell it. All the data they originally trained on was available for free on the internet. If the original work was so valuable, it shouldn't be up on the internet for free in the first place imo.
Right, but the humans willingly released all those creations for free. I think that my issues with the "AI companies stole human creations" stance is that the information was freely available to everyone, and they put a lot of money and effort into transforming it into something useful.
It's different because you have to pay Anthropic and follow their TOS to get access to their model. If it was actually on the internet for free, anyone could do whatever they wanted with it.
The root comment was asking how it is a problem. If one considers it wrong, then it’s a problem regardless of whether Anthropic is complaining or not. Anthropic’s complaining or non-complaining should have no bearing on whether it’s considered a problem or not.
The difference is in the solution that would be proposed. I'm sure Anthropic wants to create some kind of IP protection regime for their model so it can't be distilled. I want their model to be public domain, since they trained on material that was not theirs to begin with.
Because no one outside the AI scientists understand what distilling means. They probably all think about Mash and a vodka still, and a completely unrelated association.
It ain't gonna happen. At work I have a dropdown menu in vscode with a dozen models to use interchangeably. They're all essentially commodities and will compete on price and squash almost all profit margin.
That's not their business model. They won't win on price, but they won't compete on price. Their business model is making the current state of the art.
If I'm a business and I need something done today, and bc Anthropic has the best model, there's a 99.9 chance it will be completed successfully for $1000. And using Deepseek there's a 70% chance it will, for $10 - you or me will go for the $10. Big businesses don't. Bc 1000 per task is nothing to them.
The business I work for is absolutely sensitive to 10 vs 1000, depending on the task. And it's a multi-billion dollar business. 1000/task may not be much on it's own, but there are a lot of tasks.
Actually opposite occurs. Big businesses are ok with a mediocre but cheaper result. Very few are willing to pay such cost. Just look at tech wages and the distributions
Big business doesn't pay more for better, but the do pay more for predictability, support and targeted outcomes. They will happily trade a chance at 100% better results for 10% less chance of unplanned outcomes
Except businesses are going the opposite direction here. The lack of stickiness makes the “premium” argument hard to play. Oracle won because swapping databases is a giant PITA. Swapping models requires almost no effort for most uses. And because of that enterprises are all building model marketplaces where providers have to compete on price performance.
Most folks I know can choose from any of the big labs or open weight models and they get billed internally for tokens against their budget. There’s little incentive to no switch to the lower cost closers.
This setup is a nightmare scenario for the big labs trying to execute the traditional enterprise sales plays. Those only work if your product is sticky and AI models are one of the least sticky things in the history of tech.
I'm also grateful to the Chinese labs for providing workarounds for the walled gardens that the US based AI companies are attempting to create.
Does anyone know if there are any distillation datasets available? I'd love to see these distributed on BitTorrent. I think it's critical that AI be democratized and not isolated in the hands of a few private companies.
As opposed to what? The US? You think the US is any different? Literally our pedo president publicly admits to insider trading. You think the US government gives a rat's ass about the American people?
Trump kicked and banned major news orgs from the White House just now. You’d have to be will fully ignorant or really naive to believe we have fair and just media.
I agree it would be very cool if they were open sourced, but the article is talking about the KV cache technique which if I understand correctly is open source (or at least the white paper is published).
Every workplace is an authoritarian regime where workers don't have a say and grovel with "please don't replace me" like that isn't the entire point.
The whole point of AI is to get rid of you so rich people can play with the planet like it's minecraft. Software engineers just think they're special because they're the ones building it like they'll get a pat on the head for being good little servants to the investor class. Or worse, that their portfolios will let them be the gods who rule over ashes.
It's crazy how articles like this get spawn-camped by people like this trying to throw this zinger in. Both AI conpanies in the US and the chinese companies with ccp desks in the corner can be bad. The good path forward is locally hosted AI models, that's known.
It's VERY clear that the US companies are trying to push for regulation to kill open models and open weights. I see this as much more hostile and authoritarian response than what we're seeing come out of China right now.
So is China going to always publish in the open? No clue. But right now they're modeling much better behavior.
It is in the interest of everybody-except-OpenAI and Anthropic that those companies have capable competitors; better still for the competitors to share their advances.
They don't have to be our friends to act in our interest.
You ask about distillation but I wonder, is there any training datasets (~ TB-order) available that startup folks in SV use or is it so that everyone has to create their own scraping pipeline ?
> I think it's critical that AI be democratized and not isolated in the hands of a few private companies.
While I'd like to agree with this, the fact is that pushing the frontier out has always taken (and folks expect to continue to take) hundreds of millions/billions of dollars. Open source and distilled models can follow on for much cheaper, but it's hard to imagine the frontier ever being "democratized" given the huge sums of money required. It was this realization that forced OpenAI to take tons of private investment in the first place.
There’s a libertarian sentiment that that doesn’t sit well with “AI is harming people” sentiment. If AI has harmful uses, and I think anyone sensible would have to agree that it does, then giving everyone unrestricted AI is likely to make it worse.
It’s sort of like gun nuts arguing that more guns is the answer. I mean, ok, maybe you’re a responsible gun owner or AI user but relying on personal responsibility doesn’t fix systemic problems. There are bad people out there.
If food has harmful uses, and I think any one sensible would have to agree that it does, then giving everyone unrestricted food is likely to make it worse.
You can do this with cars, tools, computers, ... whatever you want. So, no, I think your point is wrong.
Safety laws are the wrong category, the equivalent regulations would be those that forbid the use of cars for drive-by shootings or as robbery getaway vehicles, and regulations against the use of food to provide crime energy.
> Haven’t tried getting a gun in California. How bad is it? How could it be improved?
You have to pass a basic knowledge test, have a clean background, prove residency in the state, be 21 (or 18 for hunting rifles, IIRC) and then wait 10 days.
It’s not game over because AI doomers are wrong in their projections. LLMs are a transformative technology like the internet, but they’re also overhyped and the useful applications aren’t as broad as people think they are. They’re also not Skynet.
I’m also skeptical about some projections, particularly for robotics, but on the Internet at least, it does seem like AI is automating most things, more or less as predicted. We already have botnets and had one notorious AI botnet swarm, fortunately easily shut down without doing any real damage. I don’t think we’ve seen the last AI botnet.
AI-automated warfare is looking pretty scary too. I don’t think it will stay in Ukraine.
Time will tell. I think the problems will come from humans automating things that shouldn’t be automated, like the disaster of the Minab school targeting, not out-of-control superintelligence.
I didn't say I don't trust anyone. Don't put words in my mouth, it's a sign of bad faith.
Other countries have governments that have earned that level of trust. I believe the US could get there eventually, but it will take a very long time because it has a very long way to go.
All concerns balance against competing concerns, and in this case freedom of computing and knowledge wins over safety. Especially since it’s trivial to copy and share open models.
Okay, you’re asserting that but I disagree. Why should anyone else be convinced? Why can’t we get the good uses without the harms? It doesn’t seem like an unavoidable tradeoff.
Well then do your best, people like me will keep sharing models just fine. (Not like I even do anything with them, but I’m a compulsive data hoarder.) It’s not like the copyright industry has been able to stop sharing either, and there’s serious money at stake there.
I predict that the AI scaremongering will fizzle out when the bubble bursts. There will still be die-hard believers but the public will lose interest.
I don’t see how a stock market crash will make AI-related concerns go away. There was a dot-com crash but the Internet just kept getting bigger and causing more problems. In many ways security has improved, but we worry more than ever about social media, etc.
Gun owner enters the conversation. In high trust societies, armed people are very polite people. I don’t want the bad guys out there being the only ones with guns. Besides I like to shoot just like you like (whatever you like to do that is legal).
Replace “gun” with anything and you will see how your comment falls apart.
What’s next? A registry for food purchases? Your beer gut is starting to show.
If you mean high-trust societies like maybe Switzerland, I agree that it can work, but the US has lots of guns and doesn’t seem very polite, so how we’re doing it doesn’t seem to be working very well.
We do have lots of food safety regulation, which has more to do with selling food.
I would say that there are very few people on earth that I trust less than Sam Altman, Dario Amodei and Elon Musk. Also my own government claims to have used Anthropic models to bomb a girls' school in Iran. If you combine US regulations with sociopathic private companies, you get into a worst case scenario for humanity imho. Again, I'm thankful to China or any other entity pushing open models, local models and even distribution of this technology.
The best explanation is that it's a goal of the CCP to generally commodotize LLMs, because LLMs will ultimately be a compliment to manufacturing (which China dominates), and you always want to "commodotize your compliments".
I think this explains why they are open sourcing broadly. It's not to be nice. It's a strategic play by the Chinese government to help ensure there are many players in this race and not too much power accumulates to American labs (even if American labs benefit in the process)
Totally agreed. People are so ready to praise China for their free models, but they aren't doing it because they believe in free open-source software. If China ever gets ahead, they're going closed source and weights immediately.
Please quote people you appear to be patronizing. China can't do anything about previous released self-hosted Chinese models. If you can show that local Chinese models funnel vast amounts us data home I'm sure you can move a lot of people to your side.
Comments like this also always fail to address why there aren't Western AI companies doing the same thing. Is it because they might get sued into oblivion by Big AI in the US?
It might be better for all of us if you solve that first instead of repeating something the government has been repeating for the last decade or more. It does this, mind you, while sabotaging itself in countless high-tech fields and leaving it all to China for the taking.
It's also possible that China has decided that this is ultimately going to be a race to the bottom, anyway, and values the soft power more highly than potential monetary profits.
Or perhaps they've looked at history and concluded that this historically hasn't been where the value is, anyway. It wouldn't be unprecedented - FAANG companies have a long tradition of publishing their algorithms and releasing open weight models. Because they saw the real value as being the training data and in proprietary special-purpose models. For example Google published the transformer architecture and released BERT as an open weight model, but doesn't really even talk in public about the (presumanbly) specialized internal models behind revenue-generating products.
I'm beyond thankful that Chinese AI models are so good. I desperately want us to cure the myriad of maladies that humans suffer needlessly with on a daily basis. We're going to need more powerful models than we have now if we're gonna do that and the Chinese are providing the competition needed to push this thing as fast as we can.
I realize "going as fast as we can" is not the most popular position atm. But I'm far more interested in what good we can do than 10% apocalypse scenarios. I volunteer with a charity for childhood brain cancer and I do not want to see another 4 year old die. I'm willing to risk anything to stop this.
Do you mean that you don't believe in the 10% apocalypse scenarios or that you think they're an acceptable risk? Only the latter is really "risk anything".
This really is a place where the guns-don't-kill-people argument applies, even moreso than guns themselves. The U.S. government massacred school children in Iran. Why does it matter how they targetted them?
It matters because the US military utilised Palantir’s Maven Smart System, a battlefield-management AI designed to compress the "kill chain". It matters because these kind of incidents are more likely to happen in the future.
If your argument is you can’t trust democratically elected governments to make decisions about technology, say that. This banal “oh the children” and “but they’ll use it to kill people” is baby think.
Over the last months I have seen news that OpenAI made breakthroughs in inference efficiency multiple times.
I have no reason to believe that the leading US labs don't have their own optimizations, or that they learned of this particular optimization from DeepSeek.
I've been running Qwen 3.8 27b (an opus 4.6 tier model), locally on a 5090 for just over two weeks @ 170 tokens/s. That's a frontier model from 9 months ago running on consumer hardware. Who knows where distillation and pruning gets us in another year.
Weird, I kinda gave up on 3.8 as anything other than a planner. I had it try to write some basic unit tests for an admittedly complex bit of code and it ran out of context thinking about the problem and exploring random parts of the code base repeatedly before it even wrote a single line. Toning down the thinking helped some, but then it wasn't much better than qwen coder.
V100S 32GB, I have had Claude optimizing it for about a week and it is already at around 900 t/s prefill, 90-100 t/s output in Pi on coding tasks. There is also a Ninfer fork for the v100 but it requires a custom format. I am working on upstream Unsloth with GGUF 4-bit quant.
(I also have flash next running even faster on this machine, something a single 5090 can do, with expert cache/pinning, but not quite as fast) :)
In all fairness there are probably a lot of folks who picked one up for around MSRP (even if one of the board partner cards with an MSRP 10-15% over the FE).
Local inference will have a boom of cheap, powerful, and available cards at some point (even if it isn’t until 2028/2029). At some point the hyperscalers, and frontier labs, will face the capex problems that everyone talks about, and NVidia, AMD, Apple, and Intel will want to keep selling products.
Powerful, by today’s standard, local inference needs to be accessible to really unlock the “AI” economy long term. It’s just like how the move from mainframes to the PC 40ish years ago unlocked the “computer revolution.”
$9k is a small price to pay to experience the rapturous glory of AGI. I'd easily pay up to 3 times that to comfortably run the superintelligent models released in this post RSI world.
I’m not an AI researcher, but it seems like there’s a ton of waste having a universal model that knows everything when any individuals use case requires like generously 10% of what’s stored in the model. Does it even need to have memorized knowledge stored in the model or could it just look up info and docs like humans do? If all you need is the language and intelligence, I think Opus5.5 equivalent intelligence will run on an iPhone within 5 years.
Indistinguishable or very mildly better. But it's considerably faster. Some portion of that is also probably down to improvements in model harnesses, I've been using opencode.
“So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.”
I don’t think it is intentional but this is actually quite bad for the western labs.
The entire booster narrative has been “look at how their revenue is growing! $10bn to $100bn ARR in under a year! This’ll be a multi-trillion IPO!” and the extrapolated future growth from $100bn to $500bn and $500bn to $1tn justified future investment… but that revenue was just because inference was expensive.
The revenue growth story is all that matters pre-IPO. If revenue falls from $100bn to $50bn that’s very very bad optics for OpenAI and Anthropic even if they are now profitable, it completely destroys the growth narrative.
I don't see why revenue has to fall even if marginal costs drop off a cliff. As long as they have the best models (perceived or otherwise) and can make security and IP guarantees that satisfy enterprise, and no firm with similar guarantees undercuts them on price (why would they want a race to the bottom?) they can have high revenue and high margin.
unless the major players collude they don’t get decide if they are in a race to the bottom. The best model was compelling 6 months ago when everyone was too impressed to care about price but that has worn off now and clients are paying attention to price. The best model is no longer a license to charge any amount.
I was about to say, one take I've heard is that the party ideology considers profit a kind of "rent" in a derogatory way, and consequently seeks to undermine the ability of western companies to collect large profit margins
It is funny that so many comments vascilate between "It is so expensive these companies can't make money and will go bankrupt in seconds" and "Inference is so cheap that these companies can't make money and will go bankrupt in seconds".
I never take them seriously, I just assume they are coming from countries that don't understand how capitalism works or are operating out of bad faith. The underlying reality of the market is always changing and needs are always changing. Some AI companies will fail, that is a given. Remember alta-vista? Yahoo? Did search go away? How about Microsoft phones? Nokia? Motorola?
OpenAI and Anthropic are not in the inference business. That is a commodity. They need to sell products and solutions.
They’re not contradictory positions. Inference is too expensive now to make money because the industry is immature and hasn’t yet optimized for financial success while customers don’t care much about price because they’re more concerned about not missing out.
Inference will be too cheap long term to make money because it is being commoditized and customers will start to care about results and not just be wowed by impressive technology.
And of course this technology will continue to exist but that is irrelevant to the business. OpenAI investors don’t care if LLMs exist in 10 years, they care if their investment in OpenAI has made money.
How much conclusive evidence people need to stop parroting that inference is expensive? Literally this article is another example showing it's cheap and just got massively cheaper.
The article is about how a month ago a new approach allowed inference costs to be cut substantially. Anthropic currently spend over $5bn per month on compute and have over $400bn in committed spend over the next 5 years. Inference is, by any measure, expensive, it’s just now getting less expensive.
Relative to traditional software margins, the type of margins we are all used to, inference is obscene.
> For nearly all tasks, I want the fastest and smartest AI model.
This is not nearly true for everyone else in the world.
For example, think about the world in ~2021 pre-LLM. Would anyone say the sentence "I only want the fastest and smartest humans working on my project"?
No of course not. Most people don't want to pay $10 million dollar salary to the best programmers in the world. They prefer to pay $200k salary to a median programmer and that's good enough for their ecommerce website.
Right now you are right -- even if I ran my local LLM all day, the quality is not nearly as great, and it runs slowly -- so I use the tier one AI subscription services as they are faster and smarter. But that might only be true for a limited amount of time, and a limited number of circumstances.
To borrow your steam engine analogy, if local LLMs get as good as a Toyota Prius, even if OpenAI / Anthropic offer Ferraris, most people will be happy with their Prius as their daily driver.
Similarly, if the big labs start raising prices or cutting usage, you won't be able to use it as much as you want -- whereas a local LLM will run all day every day without costing you any extra money.
So right now you are right, but who knows how long that will last.
I find this too. In fact, recently I've been pushing more and more to the latest and greatest model every time there is an update. It just saves me so much headache.
The margins on NVidia datacenter hardware are ... high. At least one order of magnitude larger than a consumer chip.
Given the recent deepseekv4.1 advances - how good of a 3B model can we make to run on an iphone natively? is it good enough to match common muse/dot use cases for consumers? the phone is already always on.. no need for a cloud server.
Most phones are not really "always on" in any real sense, a phone on active standby uses very little power and most of it is for its mobile connection. Local AI is best run in a stationary homelab environment, even running it on laptops has its very real problems.
> So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.
"Thus the expert in battle moves the enemy, and is not moved by him."
They figured out a clever method for avoiding excessive training costs via distillation. That forces the hand of frontier labs to move faster, produce better models, etc. (to avoid embarrassment and 'falling behind'—all the while shouldering most of the cost), which they can just keep distilling—or applying other techniques against—much to the dismay of said frontier labs.
The doublethink required to simultaneously believe "our safeguards prevent our models from doing unsanctioned cybersecurity tasks" and "distillation is why Chinese models are getting better at cybersecurity tasks" is genuinely quite funny.
I started using a zero data retention(at least claimed) deepseek v4.1 flash this month
I still use claude and openai right now, but I can see that not long in the future I won't bother with them, still waiting for a model good enough with computer use and a good enough computer use agent
There are 1000 Chinese labs. They are involuting, they cannot coordinate and the state won't let them coordinate because the state wants domination, not actual profits for anybody. So the market forces are allowed to dominate.
Because of the basic huge recession going on in China, you can't actually make money in China doing China things. So they gotta gird up their export stuff and try to export. That entails strong relations with American companies, American PR, English stuff, etc.
If you want an essay about this from a VC, read this one
What a strange way to put it. Market forces are a good thing in a market economy. Only someone who secretly wants or hopes for monopolies would you say something to the contrary (a VC).
The party has done something about enormous involution in solar panels, for example (https://www.csis.org/analysis/chinas-solar-industry-upheaval...) and previously steel. They're planning something for cars. They don't on LLM because of the newness and the wish for preeminence.
Pouring 100% of your tech optimism, and probably portfolio, into a product that some say is about to get Wile E Coyote flattened by market dynamics would probably inspire serious market skepticism.
China has a different perspective on it, they believe that there is such a thing as harmful competition and they are willing to step in to stop it.
Enshittification and related problems can be a result of market forces just as much as they can be a result of monopoly/duopoly or a small cartel. Excess competition sometimes results in all firms scraping the barrel to squeeze out pennies, especially with technology (such as large online marketplaces) making pricing more transparent.
Marx actually predicted that ever-intensifying competition would destroy markets through overproduction, although he did not use the term involution.
Price wars are good even if they lead to more bad products on the market. If quality is important people will pay more for it and ignore the bad ones, if not then everybody saves some money.
The reason it doesn't work like that IRL is centralized marketplaces. If the winning strategy on Alibaba is low prices, bad quality and botted reviews to compensate then every seller has to do it to survive, because they can't get buyers outside the platform. That's not excess competition. It's a lack of competition just on a different level.
> The reason it doesn't work like that IRL is centralized marketplaces.
Well yes, that’s why I mentioned those specifically. But even if the marketplaces were split up, someone could create an aggregator to comparison shop and the same effects would apply. The problem for producers is that the Internet erases information asymmetry.
It’s also not just affecting low end goods, it’s a constant pressure on everyone, which is why many formerly upscale brands are seeing the same problems. It’s also a general problem with public companies, as large shareholders demand constant growth, as well as many private companies owned by PE where brands are stripped for short-term profits.
My experience is that the best low cost mid-tier products right now are coming from fronts like Vevor and Fanttik who do sourcing from noname factories in China. I’m not sure if their position is sustainable; it’s not like they have much of a moat. (I guess Fanttik has a team that adds some slick design to their otherwise utilitarian items.) If that model holds up then maybe that’s the future, but I suspect that they just have a temporary advantage thanks to a dual presence and connections in both the US and China.
I'd call it overcapacity instead. A lot of investment without capital discipline making sure there is actually return of investment leading to too much supply.
Not that OpenAI, Anthropic or SpaceX aren't doing the same.
Idk, but it's doing a lot for my view of China. Maybe that's a point? Maybe they just want their own innovation to go as fast as possible and they don't care that other countries also benefit? A rising tide lifts all boats? They are already known for the best manufacturing, they're just adding software dev to the list? Maybe they just want to undermine the US in a non-aggressive way?
Why did we (the west) ever start open sourcing anything? Maybe we just like sharing? Maybe humanity only grows on pre-competitive layers like Linux and clean water. Maybe, the chinese government is closer to their people, and does not let large companies influence them and just doesn't like closed private hyperscalers with a lot of power?
(Some points assume the government has a role in the openness, which I think is likely)
My immediate midwit take is: doesn't matter if it helps Anthropic/OpenAI if it helps DeepSeek more, relatively. Making open-weights models even cheaper and easier to run expands that "market" and increases competitive pressure on the Big Two, who still have to charge money.
This looks like a speed run of the history of analytics DBMS’s.
Once upon a time, everyone had a secret sauce in network or data encoding or query optimization, but in the last ~10 years computational physics and economics have basically decided the “correct” architecture and everyone (including OSS) has converged.
I think it’s fairly simple: they’re forced into this situation by being late and worse in terms of capabilities. They’re not far behind, but as long as they’re behind they’ve needed to give people some reason to try and use their models. Cost is one factor. But it probably wasn’t enough. Being open has given them a lot of attention. Free marketing. Good will.
Put another way: if they were not cheaper and open, they would simply not be competitive. They would already be dead.
I don’t think this ends well for the Chinese labs. This is going pretty much like I thought. Western labs is just copying their improvements (I don’t think publishing the techniques matter here.. they’d just hire to gain the knowledge or figure it out themselves), and they have access to more GPUs and have better branding, so in the end where can the Chinese labs compete? Even lower cost? Open weights? I’m not sure open is a sustainable way to compete either. Eventually there will be some fully open source AI models that cuts out that avenue of competition as well.
It's just their usual national strategy like what they did to solar pane and EV. The solar panel industry is mostly dominated by China and their profit rate is basically ... negative. The EV industry in China is in similar condition, where the average profit rate is only 1.5%. Their upstream suppliers are also hold as hostages that most of them won't get their money back within 6 months.
The weird ideology here is to dominate the market at ANY COST, even it benefits the opponents.
The post makes an assumption that US labs did not already possess similar optimization. It's also very possible that they did, but are simply not telling anyone in order to maintain obscene margins on cached read, similar to AWS' absurd pricing on bandwidth.
Yeah, would have been nice if the author put a bit more thought into this article to come up with something rather than to just give up once they've reached the point of their article lol
Yes - the Chinese labs serve a market that rely on open-weight models and managed deployments, and the labs gain competitive relevance by releasing those models. The cache optimization feature they came up with required new software to utilize on the inference-end, meaning that open source software would need to be specifically updated to work with these models. It wasn't the type of advancement that they could even theoretically keep secret.
The chinese don't view AI as metaphysical, they view it as an engineering challenge they consider good for their state and want to dominate the global market like they do with batteries/EVs/photovoltaics. They want to proliferate them as much as possible and traditionally they don't care much for IP. They also want hardware makers to make optimized chips specifically for these models.
My guess would be that if they "help" western labs becoming better, then any break-throughs they (western labs) make after that, is also a benefit to the Chinese labs - if they can distill the models.
Basically, western labs are in it for the money / commercial monopoly. Chinese labs are in it for the tech? As long as they can keep distilling models, and get access to research other ways, they benefit. And if they can push western labs forward, they'll benefit from that themselves.
Because it's entirely possible that Western labs already did this optimization but didn't publish it, and then the Chinese figured it out and decided to brag about it.
We don't know either way, so I find the whole thing silly to speculate on.
All the comments here are very US/Western-centric. Maybe they are a different culture, having a completely different economic model. Maybe they are not Capitalists and not thinking in pure Capitalistic terms, such as winning, growth, market domination, IPO, market value or competition. Maybe they are not obsessed with US labs. Maybe they are ideologically different than you. Maybe they have different priorities. Maybe they have a different playbook. Maybe they never thought of it as throwing a lifeline to American labs. Maybe they don't care. Maybe they're just different people.
Unpopular, maybe, but what about the normal reasons? The researchers are looking to make a name for themselves, and/or they genuinely care about AI advancement.
AI fundamentally insecure. Vulnerable to forcing hallucinations via search results. Vulnerable to invocation of commands in data stream. More AI means more vulnerabilities.
I can't even fathom the trend these days of "we don't review the code" from security team perspective.
They get to make US labs look dumb and provide an open alternative that anyone can host themselves.
They're building bridges over the moats that companies with far too much US investment are trying to build, and if they do it continually it can help destabilize the US economy.
Open releases are just the obvious move when you're not the incumbent. You commoditize the thing your competitors charge for and get distribution you could never buy.
> So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.
Because contrarily to the author's assumption, all labs, Western or not, have sufficient skills to discover the optimizations anyway, and publishing or not is not actually that important?
Watch any interview with the Chinese AI leaders and compare it to the unending doomer word salads from America's mightiest paper billionaires. We're losing because we have a loser late stage capitalist scarcity mindset. They're winning because they're sharing notes and one-upping each other just like we used to until 2015 or so. They have a healthy ecosystem of competing small AI startups. We have two bloated unprofitable pigs both striving to be too big to fail. My money's on China for the immediate future.
How co-designed are these optimizations with the model itself? I'd imagine you can't just stick post-training adapters onto existing architectures for these things, or am I wrong?
I really want to explore the inference space, but it seems like many of the inference optimizations are coming from model-hardware codesign. I don't seem to recall many generic "inference engine" optimizations since prefill/decode disagg a year ago.
This matters for me since I want to break in but the bar seems to be understanding the actual theory of the training process now too given the codesign happening, and I'm not the richest guy on the block lol
Did private frontier models use sparse embedding and ngram first? The article claims sparse attention was copied from open weight but we can't know that. We could just as easily argue that OAI and ANT had these improvements for years and decided to slash their margins only now to stay competitive with open weight neoclouds.
Second, sparse attention is an old area of active research. Offloaded N-gram tables are the next big open weight technological leap.
The O & A strategy has until very recently been to use brute force and just throw more money at the problem.
Deepseek was the company that invented some and improved some other ideas and got them to workreliably in production. Before that Sam and Dario were basically competing in who has the most expensive training.
A random accusation of western lab using DeepSeek caching technique being top of HackerNews, this article feels more awkward than the AI race for me. It's a shame for all HackerNews viewers.
https://liorsinai.github.io/machine-learning/2025/02/22/mla.... I wrote an article on MLA last year. In short, I found the main idea of MLA as very innovative but it was documented in a strange paper with other ideas I couldn't comprehend as being useful, and they hadn't properly reported the positives or negatives of MLA.
The whole confusion expressed by this article is resolved by refusing the "arms race" framing. Chinese AI labs are acting really normal for researchers. Researchers in academia collaborate and share breakthroughs. This was until very recently the norm in American ML/AI research as well. The hawk-brained reasoning that this is some kind of "fate of the world" style arms race is as far as I can tell a narrative entirely pushed by American megacorps (and their cultural orbiters) who want to hold on to a business model of proprietary control of technology at all costs, and this is lapped up by political actors who clearly mostly just want to keep public perception in a cold war framing, which also seems mostly in the interest of consolidating power through the classic FUD method. If we think of this as normal research and development on a normal technology, it makes a lot more sense. I already see little reason to use proprietary models, but these companies insisting that they're in a war about which their supposed opponents have not seemingly gotten that memo makes me want to do so even less.
To answer the author’s question of why Chinese labs give away their work for less than cost, the answer is involution. China is struggling with overcompetition in other areas of its economy as well, such as electric cars, and perhaps ironically its labor share of income is substantially lower than the U.S.
Yeah, and Anthropic probably got inspired to this new fast read-in thing for making agentic stuff make more sense from the latest DeepSeek model. Maybe it was in the pipeline, but it clearly has the same effect and DeepSeek had published it by the point Anthropic dropped their prices for reading tokens in, so they may well have copied it.
it's funny to me how the success/not total implosion of htese companies is predicated on profitable, revolutionary-tier success, and that it's increasingly possible that the profits will never really materialize. pretty interesting move on China's part.
> If you read the news headlines these days, you would be forgiven for thinking that the Western labs are getting spawn-camped by Chinese labs en masse.
As far as I can tell, neither of the frontier US labs have referred to distillation as "stealing", but someone please provide a link if I'm wrong.
They do claim that it violates their ToS, which we can assume is simply correct, since they get to put whatever they want in their ToS.
Given all that, I don't know what the fuss is. Are they supposed to not use the advances that were openly published by Chinese labs? The entire industry is built on a discovery made at Google, which was published openly. Should Chinese labs therefore not use transformers? Should US labs not try to prevent distillation of their models?
I don’t think there is fuss, just the author sharing the information and mentioning how they find it a bit ironic that US labs expenses can be reduced drastically thanks to the Chinese companies they continuously frame as adversaries
I’m incredibly skeptical that OpenAI is spinning up custom ASICs for improved inference performance, but they never thought of optimizing KV cache until a tiny Chinese lab did it? Give me a break.
>The new game in town is adopting Chinese labs’ advances. Note how I call this adoption instead of the more vitriol-infused “stealing” that Anthropic tends to use.
I mean, there's a pretty big difference between labs publishing their research openly and a competitor utilizing it versus a lab breaking TOS to... hmmm, what's the word? steal data from a competitor?
>That’s because, unlike the Western companies, the Chinese are pretty much giving away their recipes.
Yeah, Western AI companies have never published their research. It's crazy how the Chinese had to independently develop the foundational technology that powers LLMs because Western companies simply never publish their research (I mean, as long as you ignore stuff like this <https://arxiv.org/abs/1706.03762>).
>The latest one shamelessly copied without acknowledgement is the breakthrough in KV cache optimizations that DeepSeek has generously shared with the world.
Thank you, generous corporation. I'm sorry that other corporations don't provide you free publicity for your selfless contributions to the world.
>Now I don’t know why they would freely give away such a breakthrough, but they just did
Well I'm glad the author finally got to their point. A very insightful analysis.
>They do seem to be a little embarrassed by the copying. Hence the silent releases without much pre-announcement for both Claude Opus 5.5 and GPT-6.1 Sol.
You have to be in pretty deep to infer this kind of emotion to these kinds of corporate activities.
>So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.
Then why write this article? Why point out these things just to have no conclusion?
This article sucks. Even if you hate US AI labs and are all aboard Chinese labs producing open models, there's nothing of substance here. This is the loose draft that you hand to your LLM to finish for you, but it seems the author just forgot to do so.
Even if you're willing to characterize US AI labs as evil and selfish and Chinese AI labs as righteous and generous (which is already completely trivializing these dynamics to the extent that anybody over the age of 14 can likely identify is lacking nuance), you can at least put some effort into producing some hypotheses about why these dynamics are occurring. Of course, odds are if the author did try to articulate some hypothesis, they'd likely quickly realize that the narrative they're painting just doesn't hold up.
It's actually a little funny that this whole thing is predicated on "beating" the Chinese, when (as they have been for the past 4 decades, and the Japanese before them) they're perfectly happy to let us do the bulk of the work and then swoop in with a svelte, cheap, user-friendly version right after. One part Apple, one part Dollar Store.
Is the thinking that the day or so between US systems achieving ASI and Chinese systems doing the same, we'll figure out a way to neutralize them indefinitely? Because otherwise, none of this makes much sense. And it only starts to swerve back to sanity if the assumption is that this isn't a race or competition, but instead a joint effort to achieve something good for humanity. But you can't really delta profit off that, can you?
Color me skeptical that OpenAI and Anthropic’s researchers had never thought to dig into these optimizations, and instead are just spending hundreds of billions on data centers and custom ASICs.
This is an extremely thin analysis that has obviously been voted to the top of the homepage because HN hates the big labs.
Due to shifting definitions, I believe that makes you a boomer[0] so you'd be readily excused for not knowing.
Sarcasm aside, gaming and FPS terminology are so tightly coupled with online culture that it's just assumed everyone knows it. Spawn camping is among the oldest examples of gaming terms that broke out into common online usage and it dates back to Quake some time around 1997.
> FPS terminology are so tightly coupled with online culture
What is online culture? Is someone who is heavily into instagram for fashion, facebook for family contact and news, maybe Google for mail and search, part of online culture? Because I know people who are like that and there's no way they know what "spawn" or "camping" mean in gamer context and certainly wouldn't be able to piece together what "spawn camping" is.
> if you've been online a little bit you'd know this expression
I've been online since circa 1995 (earlier if you count BBSs), and I can't say I did. It's possible to infer its meaning but assuming everyone is on the same circles as one is, is silly.
In first-person shooters, when you die you regenerate (“respawn”) somewhere on the map. If those regeneration points are know to opposing players, they can wait next to them, and kill you again the moment you respawn.
The premise in this article is: Western companies do a ton of expensive work building new models, meanwhile the Chinese companies just wait for a Western release and then they immediately grab and distill it and announce it as their own model. That’s the spawn-camp.
If you tend to stop reading every time you encounter something you don't understand, I can't imagine you learn very much
"spawn-camping" is the process of taking out your enemies at the point they spawn (or appear) in a game without giving them a chance to regroup. In this case I think the writer is saying that the news implies that western models are getting distilled on release. Not the perfect analogy but it gives some color.
Deepseek's innovations are published as research. There's nothing 'assumed' about this. The common slur repeated in the West that Chinese labs are parasitic distillers is totally absurd when so many genuinely valuable advances and contributions to the field are published openly by China's labs.
I am talking about correlating Anthropic/OpenAI cache prices going down with Deepseek publication - neither of those labs have said that's what they used for example.
And the only data they are showing is that cache prices went down for new Claude/OpenAI models but that's proving nothing, IMO.
Feels like Anthropic crying do as I say not as I do.
An argument can be made that Anthropic is also only doing the last 5% of the work (because the content they are training on was the other 95%) but that's a bit more philosophical.
The original authors of all the text, creators of the media and developers of the software did far more work than Anthropic.
Did Anthropic put work in? Yes. Did they derive their value from Humanity being open with knowledge then try to sell it back? Also yes.
Did they even steal the tech? Also yes.
Is that not what the foundation models are? A new reorganization of existing knowledge?
So then the problem is that Anthropic seems hypocritical when they knowingly insert themselves into this chain, and then complain about people down-chain from them.
To remedy the negative impressions (if they even care to do so) they should do 1 of 2 things: 1) stop complaining about it 2) stop distilling other people's work
No LLM products would exist without the avalanche of largely non-consensual use of IP to create them, full stop. Any of these companies doing this and then turning around and complaining when their IP is "breached" are going to met with a chorus of tiny violins.
It is not a natural assumption that someone will digitize the artwork and use it to adjust a couple thousand matrix coefficients in a complex computer program. To most people that seems like copying with extra steps. The brain may in some ways resemble a computer, but what sets it apart is that we have always lived with brains. Everything a human does has already anticipated the presence of other brains, while etched circuits on ultrapure silicon crystals are something new.
A comparable idea could be that an encyclopedia is only distilling the things that other people did, and how dare they sell them. But the "only" is doing quite a bit of work. LLMs do not just spawn into existence. There is a body of work that they feed on, and then there is also very attributable work they do around and on top of that. All labs are struggling around the first order question: Is it okay to use prior work like this? The second order issue is still entirely reasonable to separately have and enforce rules about.
This isn’t meant as a moral argument, just musing about the relative cost comparison.
Image/video models are, but those weren’t the topic.
The totality of the content on internet is worth several orders of magnitude more.
And the monetary cost doesn't even register when weighed against the blood, sweat and tears that went into capturing the authentic experiences of real human beings, whose honest expressions are now at least in some cases getting hoovered up, ingested, and then destroyed for all eternity, for fear that this specific work is the rounding error that might give an equally immoral competitor the edge in the bicycle-riding flamingo race that is currently consuming an absurd amount of the world's creativity and attention.
A training set of 15 trillion tokens is 10 trillion words.
A penny a word is cheaper than the cheapest beginner freelance writer.
That makes a training set of 10 trillion words cost $100B.
Lots of assumptions there for sure, but we're certainly in the ballpark you are describing.
$500B for all kinds including TV and online
$300B for newsrooms including all staff
$140B for newsroom reporters only
So yeah, I think the price of the information ingested is way higher than training costs
How is this even a question.
“bro like, what if we could price the sum total of human knowledge? That wouldn’t be that much, right?”
It’s like FTL. Until someone realizes it, it’s just talk.
Anthropic’s anger here seems mostly rooted in their annoyance that this exposes they don’t really have core IP that’s not just easily replicated. And that’s clearly a problem for a deeply unprofitable company trying to convince people they’re worth $2 trillion.
>An argument can be made that Anthropic is also only doing the last 5% of the work (because the content they are training on was the other 95%) but that's a bit more philosophical.
How is this philosophical? They should release the unsupervised pretrains, at the very least.
With my apologies to Brewster Kahle, "Universal Access to All Knowledge." [2]
[1] https://wiki.archiveteam.org/index.php/ArchiveTeam_Warrior
[2] https://www.youtube.com/watch?v=RV_ALlJGU_c
"The Spice must flow."
Oh that’s very sad.
Meanwhile Anthropic made a product from the work effort of millions of people without compensating them, sell that product on tap and unless I am mistaken do not even have their competitors’ cover of having released any sort of meaningful open weights model.
They have taken from culture (including very specifically their most direct customers’ specific culture — our culture), turned it into a machine to make themselves rich, appear likely to predicate their valuation on permanently removing people from the workforce, then want to dump themselves onto pensions funds and ordinary savers to carry the bag.
It is, I agree, philosophical, because karma is a philosophy as well as a bitch.
I see absolutely no distinction between the two, aside from minor technical approaches to gathering the content.
It sounds exactly the same, not more philosophical to me, except one is more inconvenient.
I think a good litmus test here would be if Anthropic were to not care about distilling their models when the distillers keep the resulting models closed-source and sell tokens via an API. If they cared only about security concerns and not about people profiting off of their work, then they should be publicly fine with this and only protest against it going into open-weights models.
that's an interesting way to describe reddit posts
And writing a book requires many more resources than what anthropic does
Anthropic and OpenAi are spending a $$$$ to "distill" human output into an AI model, then others are spending $$ to distill their AI model into a near-equivalent model.
This is the same reason IP rights exist. On the surface, something like a patent feels ludicrious and even feels morally wrong. Some guy wrote down the recipe for arranging atoms or bits in a particular way, and now I can't!? However, it's designed to solve the same problem, figuring out and describing the process is much more costly than replicating it.
That doesn't mean they won't try, and that also doesn't mean they won't succeed.
I wish people here could at least bother to inform themselves about the IP rights they are so quick to insist are abhorrent, when they seem to not even have a first clue as to what they actually cover.
Can you elaborate on that? I mean my direct answer would be no, of course not. But why do you think frontier models are distilled? I think maybe there is an equivocation over the word “distillation.”
Frontier labs train on their own pretraining data, human feedback, synthetic data, and research. A distilled model is specifically optimized to reproduce another model's behavior.
Meanwhile R1-Distill-Qwen-32B was distilled from DeepSeek-R1.
If you want to say a frontier model is "distilled" from the world's data and R1-Distill-Qwen-32B is distilled from DeepSeek-R1 then you are equivocating two very different things.
You could say Anthropic distilled human knowledge and art.
The word itself is the pivot, not anything else.
I don't know anyone with even a passing understanding of how LLM training works that thinks that is the appropriate analogy.
it's not complex. there's hundreds of billions of investor dollars counting on vendor lock in and walled gardens
If I'm a business and I need something done today, and bc Anthropic has the best model, there's a 99.9 chance it will be completed successfully for $1000. And using Deepseek there's a 70% chance it will, for $10 - you or me will go for the $10. Big businesses don't. Bc 1000 per task is nothing to them.
Also, is it really 99.9% vs 70%, or 99.9% vs 99%?
The real issue is that Deepseek has a 99.7% chance. So I can run it 10 times until it works and still pay 1/10 the money.
Big businesses might pay $1000 vs $60 for certain tasks, but that won't work out well at scale.
Most folks I know can choose from any of the big labs or open weight models and they get billed internally for tokens against their budget. There’s little incentive to no switch to the lower cost closers.
This setup is a nightmare scenario for the big labs trying to execute the traditional enterprise sales plays. Those only work if your product is sticky and AI models are one of the least sticky things in the history of tech.
why should sammie or darigold have the keys to the kingdom?
Does anyone know if there are any distillation datasets available? I'd love to see these distributed on BitTorrent. I think it's critical that AI be democratized and not isolated in the hands of a few private companies.
Is it going to shut the laptop's lid when i'm not looking and pinch my fingers?
Love the arguments btw. Your position is so cut and dry that you can’t even support it with evidence.
China has zero independent media and has the largest and most sophisticated censorship and surveillance state in the world.
Trump would love to have the absolute power that Xi has, but thankfully doesn’t.
Tired of this lazy whataboutism.
Trump kicked and banned major news orgs from the White House just now. You’d have to be will fully ignorant or really naive to believe we have fair and just media.
The whole point of AI is to get rid of you so rich people can play with the planet like it's minecraft. Software engineers just think they're special because they're the ones building it like they'll get a pat on the head for being good little servants to the investor class. Or worse, that their portfolios will let them be the gods who rule over ashes.
https://www.anthropic.com/research/glm-5-3-and-the-spread-of...
It's VERY clear that the US companies are trying to push for regulation to kill open models and open weights. I see this as much more hostile and authoritarian response than what we're seeing come out of China right now.
So is China going to always publish in the open? No clue. But right now they're modeling much better behavior.
They don't have to be our friends to act in our interest.
The point is get what you can from both to develop open models, data and tools.
Absolutely. In fact, the authoritarian regime already did try to force export controls on major frontier labs not long go so this isn't a theoretical.
Or did they not pull back when their models allegedly became highly capable, with the whole mythos debacle ?
You ask about distillation but I wonder, is there any training datasets (~ TB-order) available that startup folks in SV use or is it so that everyone has to create their own scraping pipeline ?
While I'd like to agree with this, the fact is that pushing the frontier out has always taken (and folks expect to continue to take) hundreds of millions/billions of dollars. Open source and distilled models can follow on for much cheaper, but it's hard to imagine the frontier ever being "democratized" given the huge sums of money required. It was this realization that forced OpenAI to take tons of private investment in the first place.
It’s sort of like gun nuts arguing that more guns is the answer. I mean, ok, maybe you’re a responsible gun owner or AI user but relying on personal responsibility doesn’t fix systemic problems. There are bad people out there.
You can do this with cars, tools, computers, ... whatever you want. So, no, I think your point is wrong.
Now, what I want to regulate are accordions.
Haven’t tried getting a gun in California. How bad is it? How could it be improved?
You have to pass a basic knowledge test, have a clean background, prove residency in the state, be 21 (or 18 for hunting rifles, IIRC) and then wait 10 days.
AI-automated warfare is looking pretty scary too. I don’t think it will stay in Ukraine.
Other countries have governments that have earned that level of trust. I believe the US could get there eventually, but it will take a very long time because it has a very long way to go.
I predict that the AI scaremongering will fizzle out when the bubble bursts. There will still be die-hard believers but the public will lose interest.
I don’t see how a stock market crash will make AI-related concerns go away. There was a dot-com crash but the Internet just kept getting bigger and causing more problems. In many ways security has improved, but we worry more than ever about social media, etc.
People did use global agreements and regulation to fix the ozone hole, though, so I think there’s a chance.
Replace “gun” with anything and you will see how your comment falls apart.
What’s next? A registry for food purchases? Your beer gut is starting to show.
Surely you could name three such societies?
We do have lots of food safety regulation, which has more to do with selling food.
Worse for whom?
The only effective defense against predatory corporate and government AI is personal protective AI.
Anything else is unilateral disarmament. It's the only way individuals can survive in the worse case scenario.
> gun nuts
Guns are different. They can't protect you against the government, contrary to gun nut claims.
I think this explains why they are open sourcing broadly. It's not to be nice. It's a strategic play by the Chinese government to help ensure there are many players in this race and not too much power accumulates to American labs (even if American labs benefit in the process)
Please quote people you appear to be patronizing. China can't do anything about previous released self-hosted Chinese models. If you can show that local Chinese models funnel vast amounts us data home I'm sure you can move a lot of people to your side.
Comments like this also always fail to address why there aren't Western AI companies doing the same thing. Is it because they might get sued into oblivion by Big AI in the US?
It might be better for all of us if you solve that first instead of repeating something the government has been repeating for the last decade or more. It does this, mind you, while sabotaging itself in countless high-tech fields and leaving it all to China for the taking.
Or perhaps they've looked at history and concluded that this historically hasn't been where the value is, anyway. It wouldn't be unprecedented - FAANG companies have a long tradition of publishing their algorithms and releasing open weight models. Because they saw the real value as being the training data and in proprietary special-purpose models. For example Google published the transformer architecture and released BERT as an open weight model, but doesn't really even talk in public about the (presumanbly) specialized internal models behind revenue-generating products.
Commoditizing one’s compliments is a different strategy.
I realize "going as fast as we can" is not the most popular position atm. But I'm far more interested in what good we can do than 10% apocalypse scenarios. I volunteer with a charity for childhood brain cancer and I do not want to see another 4 year old die. I'm willing to risk anything to stop this.
> So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.
That inference wasn't profitable is a widespread myth.
Analysis based on Kimi K3 suggests that OpenAI and Anthropic have margins well north of 95%: https://inferencex.semianalysis.com/run/kimi-k3-on-b200
Over the last months I have seen news that OpenAI made breakthroughs in inference efficiency multiple times.
I have no reason to believe that the leading US labs don't have their own optimizations, or that they learned of this particular optimization from DeepSeek.
(I also have flash next running even faster on this machine, something a single 5090 can do, with expert cache/pinning, but not quite as fast) :)
Local inference will have a boom of cheap, powerful, and available cards at some point (even if it isn’t until 2028/2029). At some point the hyperscalers, and frontier labs, will face the capex problems that everyone talks about, and NVidia, AMD, Apple, and Intel will want to keep selling products.
Powerful, by today’s standard, local inference needs to be accessible to really unlock the “AI” economy long term. It’s just like how the move from mainframes to the PC 40ish years ago unlocked the “computer revolution.”
I don’t think it is intentional but this is actually quite bad for the western labs.
The entire booster narrative has been “look at how their revenue is growing! $10bn to $100bn ARR in under a year! This’ll be a multi-trillion IPO!” and the extrapolated future growth from $100bn to $500bn and $500bn to $1tn justified future investment… but that revenue was just because inference was expensive.
The revenue growth story is all that matters pre-IPO. If revenue falls from $100bn to $50bn that’s very very bad optics for OpenAI and Anthropic even if they are now profitable, it completely destroys the growth narrative.
I never take them seriously, I just assume they are coming from countries that don't understand how capitalism works or are operating out of bad faith. The underlying reality of the market is always changing and needs are always changing. Some AI companies will fail, that is a given. Remember alta-vista? Yahoo? Did search go away? How about Microsoft phones? Nokia? Motorola?
OpenAI and Anthropic are not in the inference business. That is a commodity. They need to sell products and solutions.
Inference will be too cheap long term to make money because it is being commoditized and customers will start to care about results and not just be wowed by impressive technology.
And of course this technology will continue to exist but that is irrelevant to the business. OpenAI investors don’t care if LLMs exist in 10 years, they care if their investment in OpenAI has made money.
Relative to traditional software margins, the type of margins we are all used to, inference is obscene.
If local LLMs get "good" enough, people will soon paying for subscriptions to ChatGPT and Claude, which hurts their revenue.
It is vanishingly rare I ask an older model to do any task. Newer bigger and smarter models will just do the task better.
Therefore, I believe we are nowhere near 'good enough'.
I never drive my steam engine to work these days. It isn't good enough.
This is not nearly true for everyone else in the world.
For example, think about the world in ~2021 pre-LLM. Would anyone say the sentence "I only want the fastest and smartest humans working on my project"?
No of course not. Most people don't want to pay $10 million dollar salary to the best programmers in the world. They prefer to pay $200k salary to a median programmer and that's good enough for their ecommerce website.
To borrow your steam engine analogy, if local LLMs get as good as a Toyota Prius, even if OpenAI / Anthropic offer Ferraris, most people will be happy with their Prius as their daily driver.
Similarly, if the big labs start raising prices or cutting usage, you won't be able to use it as much as you want -- whereas a local LLM will run all day every day without costing you any extra money.
So right now you are right, but who knows how long that will last.
Given the recent deepseekv4.1 advances - how good of a 3B model can we make to run on an iphone natively? is it good enough to match common muse/dot use cases for consumers? the phone is already always on.. no need for a cloud server.
Think you missed a word there.
"Thus the expert in battle moves the enemy, and is not moved by him."
They figured out a clever method for avoiding excessive training costs via distillation. That forces the hand of frontier labs to move faster, produce better models, etc. (to avoid embarrassment and 'falling behind'—all the while shouldering most of the cost), which they can just keep distilling—or applying other techniques against—much to the dismay of said frontier labs.
Checkmate.
I still use claude and openai right now, but I can see that not long in the future I won't bother with them, still waiting for a model good enough with computer use and a good enough computer use agent
Any ideas?
Because of the basic huge recession going on in China, you can't actually make money in China doing China things. So they gotta gird up their export stuff and try to export. That entails strong relations with American companies, American PR, English stuff, etc.
If you want an essay about this from a VC, read this one
https://earnedintuition.substack.com/p/involution-without-ex...
Enshittification and related problems can be a result of market forces just as much as they can be a result of monopoly/duopoly or a small cartel. Excess competition sometimes results in all firms scraping the barrel to squeeze out pennies, especially with technology (such as large online marketplaces) making pricing more transparent.
Marx actually predicted that ever-intensifying competition would destroy markets through overproduction, although he did not use the term involution.
The reason it doesn't work like that IRL is centralized marketplaces. If the winning strategy on Alibaba is low prices, bad quality and botted reviews to compensate then every seller has to do it to survive, because they can't get buyers outside the platform. That's not excess competition. It's a lack of competition just on a different level.
Well yes, that’s why I mentioned those specifically. But even if the marketplaces were split up, someone could create an aggregator to comparison shop and the same effects would apply. The problem for producers is that the Internet erases information asymmetry.
It’s also not just affecting low end goods, it’s a constant pressure on everyone, which is why many formerly upscale brands are seeing the same problems. It’s also a general problem with public companies, as large shareholders demand constant growth, as well as many private companies owned by PE where brands are stripped for short-term profits.
My experience is that the best low cost mid-tier products right now are coming from fronts like Vevor and Fanttik who do sourcing from noname factories in China. I’m not sure if their position is sustainable; it’s not like they have much of a moat. (I guess Fanttik has a team that adds some slick design to their otherwise utilitarian items.) If that model holds up then maybe that’s the future, but I suspect that they just have a temporary advantage thanks to a dual presence and connections in both the US and China.
Not that OpenAI, Anthropic or SpaceX aren't doing the same.
Do you do business in China?
I'm curious what you mean by this? Because in my experience, you can only do business in China by doing "China" things.
I'd be interested in picking your brain as to how you get around those issues?
I'm talking like, getting 100x, VC sized returns. Of course you can sell widgets in China, it's a major world economy.
That is how actual capitalist market should work and what anti-monopoly legislation should ensure.
Why did we (the west) ever start open sourcing anything? Maybe we just like sharing? Maybe humanity only grows on pre-competitive layers like Linux and clean water. Maybe, the chinese government is closer to their people, and does not let large companies influence them and just doesn't like closed private hyperscalers with a lot of power?
(Some points assume the government has a role in the openness, which I think is likely)
Once upon a time, everyone had a secret sauce in network or data encoding or query optimization, but in the last ~10 years computational physics and economics have basically decided the “correct” architecture and everyone (including OSS) has converged.
Put another way: if they were not cheaper and open, they would simply not be competitive. They would already be dead.
I don’t think this ends well for the Chinese labs. This is going pretty much like I thought. Western labs is just copying their improvements (I don’t think publishing the techniques matter here.. they’d just hire to gain the knowledge or figure it out themselves), and they have access to more GPUs and have better branding, so in the end where can the Chinese labs compete? Even lower cost? Open weights? I’m not sure open is a sustainable way to compete either. Eventually there will be some fully open source AI models that cuts out that avenue of competition as well.
The weird ideology here is to dominate the market at ANY COST, even it benefits the opponents.
Basically, western labs are in it for the money / commercial monopoly. Chinese labs are in it for the tech? As long as they can keep distilling models, and get access to research other ways, they benefit. And if they can push western labs forward, they'll benefit from that themselves.
We don't know either way, so I find the whole thing silly to speculate on.
Unpopular, maybe, but what about the normal reasons? The researchers are looking to make a name for themselves, and/or they genuinely care about AI advancement.
I can't even fathom the trend these days of "we don't review the code" from security team perspective.
Just my guess though.
They're building bridges over the moats that companies with far too much US investment are trying to build, and if they do it continually it can help destabilize the US economy.
Because contrarily to the author's assumption, all labs, Western or not, have sufficient skills to discover the optimizations anyway, and publishing or not is not actually that important?
I'm glad people are saying this out loud, because that is what they want. Not for the good of the world, but for the good of their pockets.
How co-designed are these optimizations with the model itself? I'd imagine you can't just stick post-training adapters onto existing architectures for these things, or am I wrong?
I really want to explore the inference space, but it seems like many of the inference optimizations are coming from model-hardware codesign. I don't seem to recall many generic "inference engine" optimizations since prefill/decode disagg a year ago.
This matters for me since I want to break in but the bar seems to be understanding the actual theory of the training process now too given the codesign happening, and I'm not the richest guy on the block lol
Second, sparse attention is an old area of active research. Offloaded N-gram tables are the next big open weight technological leap.
Deepseek was the company that invented some and improved some other ideas and got them to workreliably in production. Before that Sam and Dario were basically competing in who has the most expensive training.
Had western labs figured that out before, they would have used it to make kv caching cheaper before and not only now.
The burden of proof here is on western labs. But I doubt they'll try to lie that much.
Ah brings back Halo 2 memories
They do claim that it violates their ToS, which we can assume is simply correct, since they get to put whatever they want in their ToS.
Given all that, I don't know what the fuss is. Are they supposed to not use the advances that were openly published by Chinese labs? The entire industry is built on a discovery made at Google, which was published openly. Should Chinese labs therefore not use transformers? Should US labs not try to prevent distillation of their models?
I’m incredibly skeptical that OpenAI is spinning up custom ASICs for improved inference performance, but they never thought of optimizing KV cache until a tiny Chinese lab did it? Give me a break.
timeline suggests not.
I mean, there's a pretty big difference between labs publishing their research openly and a competitor utilizing it versus a lab breaking TOS to... hmmm, what's the word? steal data from a competitor?
>That’s because, unlike the Western companies, the Chinese are pretty much giving away their recipes.
Yeah, Western AI companies have never published their research. It's crazy how the Chinese had to independently develop the foundational technology that powers LLMs because Western companies simply never publish their research (I mean, as long as you ignore stuff like this <https://arxiv.org/abs/1706.03762>).
>The latest one shamelessly copied without acknowledgement is the breakthrough in KV cache optimizations that DeepSeek has generously shared with the world.
Thank you, generous corporation. I'm sorry that other corporations don't provide you free publicity for your selfless contributions to the world.
>Now I don’t know why they would freely give away such a breakthrough, but they just did
Well I'm glad the author finally got to their point. A very insightful analysis.
>They do seem to be a little embarrassed by the copying. Hence the silent releases without much pre-announcement for both Claude Opus 5.5 and GPT-6.1 Sol.
You have to be in pretty deep to infer this kind of emotion to these kinds of corporate activities.
>So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.
Then why write this article? Why point out these things just to have no conclusion?
This article sucks. Even if you hate US AI labs and are all aboard Chinese labs producing open models, there's nothing of substance here. This is the loose draft that you hand to your LLM to finish for you, but it seems the author just forgot to do so.
Even if you're willing to characterize US AI labs as evil and selfish and Chinese AI labs as righteous and generous (which is already completely trivializing these dynamics to the extent that anybody over the age of 14 can likely identify is lacking nuance), you can at least put some effort into producing some hypotheses about why these dynamics are occurring. Of course, odds are if the author did try to articulate some hypothesis, they'd likely quickly realize that the narrative they're painting just doesn't hold up.
Is the thinking that the day or so between US systems achieving ASI and Chinese systems doing the same, we'll figure out a way to neutralize them indefinitely? Because otherwise, none of this makes much sense. And it only starts to swerve back to sanity if the assumption is that this isn't a race or competition, but instead a joint effort to achieve something good for humanity. But you can't really delta profit off that, can you?
This is an extremely thin analysis that has obviously been voted to the top of the homepage because HN hates the big labs.
It's not that niche, if you've been online a little bit you'd know this expression.
No way. You'd need to be pretty well versed in gamer lingo. Even more specifically, combative, likely FPS gamer lingo.
Sarcasm aside, gaming and FPS terminology are so tightly coupled with online culture that it's just assumed everyone knows it. Spawn camping is among the oldest examples of gaming terms that broke out into common online usage and it dates back to Quake some time around 1997.
[0] anyone older than 39 at this point
What is online culture? Is someone who is heavily into instagram for fashion, facebook for family contact and news, maybe Google for mail and search, part of online culture? Because I know people who are like that and there's no way they know what "spawn" or "camping" mean in gamer context and certainly wouldn't be able to piece together what "spawn camping" is.
I've been online since circa 1995 (earlier if you count BBSs), and I can't say I did. It's possible to infer its meaning but assuming everyone is on the same circles as one is, is silly.
The premise in this article is: Western companies do a ton of expensive work building new models, meanwhile the Chinese companies just wait for a Western release and then they immediately grab and distill it and announce it as their own model. That’s the spawn-camp.
Do you have many mini tantrums like this per day? Probably makes you very difficult to work with Mr I was a CTO.
"spawn-camping" is the process of taking out your enemies at the point they spawn (or appear) in a game without giving them a chance to regroup. In this case I think the writer is saying that the news implies that western models are getting distilled on release. Not the perfect analogy but it gives some color.
In a PvP (player vs player) game, if you kill a player the moment they spawn into the game arena, that's called "spawn-camping".
Edit: Apt domain.
And the only data they are showing is that cache prices went down for new Claude/OpenAI models but that's proving nothing, IMO.