IMHO, the biggest problem with the future of open weights models is that currently, open weights models are the result of philanthropy by some private org. (e.g. DeepSeek).
The spigot can be turned off at any time.
Until there's some sort of "community owned hardware", open weights models are always at risk of being discontinued.
Yeah, but the biggest plus for open models is that they can never be taken away. In other words, whatever capabilities they reach (even if there will never be another model), those stay forever. That can't be said for API-based models where a provider can sunset models whenever they feel like (i.e. gpt5-mini will soon be gone, and replaced by a more expensive 5.4-mini, same for goog, etc).
And there will always be incentivised parties that release models. Nvda for one has every incentive to keep the nemotron line going, as they're directly profiting from people running this. And the models aren't really far from open SotA anyway.
Goog will probably continue to release the small models, since they'll use them for browser stuff anyway, and know that they'll leak. So for them it's a win-win to release the small models and gain some dev market share.
And the chinese labs also have incentives to keep releasing models, and will likely continue to get gov support to do so (yay commercial wars between nations).
Your right to 3d print whatever you want is about to be taken away (in California).
What software you can run on your computer can already be restricted.
Absolutely everything can be taken away. The simplest way to remove open models is probably to declare them a tool that terrorists could use. Crazy? Yes, the world is totally crazy these days.
There's also, importantly, a distinction between what are told we can no longer use, and what can actually be taken away.
Open source and open hardware can be called illegal by a government, but, if we collectively invest our energy into open alternatives, they can't be taken away in the same sense. I can build a RepRap printer and I can use a local AI model. It's on all of us to make sure that the open alternatives are viable, maybe in the current global political reality now more than ever.
Making something illegal isn't a disincentive for everyone. When they start banning books, some of us start assembling printing presses.
Just like declaring piracy illegal stopped piracy and removed pirated materials from everyone's computers.
Everything cannot, in fact, be taken away. Don't propagandize yourself. Some things, like information, are free. Not even China can prevent all its citizens from accessing Western internet. USGov simply does not have the resources to find and audit every hard drive and USB stick in the country for illegal files. The internet cannot be censored 100% without literally cutting every cable and confiscating every radio.
The software that runs on my computer cannot, in fact, be restricted. It can be declared illegal, but there literally is no mechanism by which it can be enforced other than a government goon standing over my shoulder 24/7.
Some freedoms really cannot be removed without utterly implausible amounts of effort. Arguing otherwise is helping to erode freedom. So stop it.
On PCs, the best you could really do is restrict access to certain websites on certain boxes with TPMs the users can't disable. Remote attestation can lock people out of your stuff, but not out of their own stuff. For that you need control of the device. Of course, most mobile phones aren't easy for the user to have control of, but most PCs still are, so long as you scrub the rootkits (e.g. windows) off 'em
True, but the capabilities and knowledge of that model are also frozen in time, so the value of that model declines over time.
A model that writes code without knowledge of any language or library changes for half a decade is less useful. A 2021 era chatgpt would be quite quaint in 2026.
Right now the Chinese labs might have incentives to release their models for free, and maybe Google is happy to release open weights today, but I'm sure there are already bean counters at Google salivating at the idea of having Gemini in Chrome as part of a Google AI monthly subscription just like YouTube premium and other Google subscriptions.
> Nvda for one has every incentive to keep the nemotron line going
They're releases so far have been kind of lackluster compared to Qwen and other Chinese models. My suspicion is that Nvidia won't be releasing models that appear to compete with frontier models because that would upset their big customers.
I don't think that's the case, it's not philanthropy, they are getting something out of it. The labs are learning from one another from the shared models.
Plus I am certain it makes financial sense. I am guessing here but fully utilizing a subscriptions limits probably costs the operator more money than the subscription revenue, that is why anthropic is making such a big stink about the chinese data harvesting. By releasing the weights, you are relieving yourself from that burden because the competition does not need to hammer your subscription service they can just download your model and analyze it and run it all day.
Also for the largest models it makes no sense to run it yourself unless you are a major player. Renting the hardware is ludicrously more expensive than their subscription tens of thousands of dollars. And buying the hardware to run them is in the hundreds of thousands of dollars.
I think model training is pretty hard to do efficiently on a vastly distributed network. If the model cant fit into the VRAM of the node your performance becomes so bad its useless, so a distributed model could only be properly trained if the size of the model doesnt exceed the majority of the nodes VRAM sizes. Maybe there is a different way of doing training but this would be the only way I can see. And it would still be much worse than just using a big datacenter where everything is fully interconnected. BOINC projects work great because its usually just a lot of small compute and memory required so every old desktop and laptop can contribute. Training a model which can compete and is not tiny requires neither low compute or low memory amount. BOINC tasks take minutes usually or sometimes hours but not weeks or months like training a model from scratch. But something like 7B or lower could maybe be trained like this. Im not sure but I think someone is already working on something like this but I dont remember the name of the project.
My understanding is that in addition to your comment and the development of a method to separate the training data for distributed learning, the latency/bandwidth of systems connected on the internet is a challenge, too. Information has to be sent around before and after any hypothetical number crunching.
This has been a (noble) goal of lots of different projects in the community for a long time. Federated learning projects like Flower have been chipping away at it for a long time. There are many many hurdles to be cleared before anything in this area is super feasible as an alternative, but I applaud everyone who works on it.
It's just a smart business decision that allows their models to compete and gain market-share against much pricier private models. No philanthropy there.
True. And it's possible that this has already happened at Alibaba Qwen - at least for the smaller models that people had a chance of running at home (122B and smaller).
We'll see. The qwen team has always released a few close to sota but proprietary models in between tgeir open releases. We did get 3.6 35B and 27B so its not all set in stone yet.
Its higley unlikely we get another open llama model though after the llama4 flop, even if their muse spark seems pretty good.
How is this a complaint? Once you have the model, you have the model. Download DeepSeek-R1 671B and you have it. You might not get improvements in the future, just like you may not ever get a future release of an open source project. Is that an indictment of open source?
But consider the alternative. OpenAI and Anthropic can shut off your account or API key at any time for any reason. How is this better? You have way more security when you're running your own model.
The Chinese models will not overtake the frontier US ones given the current way things are going. The US models derive their lead from incredible efforts to source more and higher quality (mostly synthetic data) via great feats (eg generating with humongous teacher models that could never feasibly serve interactive traffic). The Chinese models advance via heroic efforts to optimize models and great feats to secure more and higher quality training data from the US frontier models.
For an (Chinese) open weight model to surpass the (US lab) frontier models, this equation must flip and the Chinese labs must entirely retool from harvesting frontier model data to producing the data systems and efforts to produce novel data; as well as procuring latest generation hardware en masse for this. This does not happen easily. Also training a frontier scale model is actually not such an unimaginable feat: doing all the inference with the teacher models is where the hardware goes.
Unless you are working at one of these companies you don't know what they are doing.
You don't know what's happening in z.ai nor alibaba. And you don't know what's happening in anthropic and open ai.
I don't know what they are all doing, but I find it extremely unlikely that they are not all collecting data from one another. I am confident anthropic has a team going over GML 5.2 weights even if it's just to see where the competition is.
Just because some labs are getting data from Anthropic does not mean they are not also doing their own research.
They were focused on optimization because they could not get the best hardware.The only reason their top labs are behind may be because they did not have h200s and MI350s. And now they do.
Plus you are discounting other risks, Anthropic is currently sitting on "the best" models in the world because they got in a pissing match with the US administration.
btw: This could be the case in china as well, their administration has been surprisingly open on AI exports and open weight models, that we know of. There is a very small but not trivial chance they are hogging a better version of glm 5.2 for example, but no one is allowed to talk about it. Now I am not saying that is the case, I am saying the two cases (chinese labs are 6 months behind, they are forced to suppress their best models) are indistinguishable.
> Chinese labs must entirely retool from harvesting frontier model data to producing the data systems and efforts to produce novel data
Even if your characterization is accurate, they could do this tomorrow and are not so myopic that they wouldn’t have thought about it. I don’t see this as a barrier, and I see a lot of the same underestimation of Asia that’s been happening for 50 years. There’s not some innate American advantage to building LLMs, and personally I think whatever head start the US has is going to be squandered on delays from the export control “to dangerous for release” LARPing we’re seeing.
Exactly. If they wanted to they could produce the same amount of data. Companies like Scale, Mercor, Surge exists for a reason, a reason that doesn't need to exist in China if they mandate Chinese enterprises to provide all their real world data (or have them work inside RL environments) to the model companies for post training. There is no real advantage that US companies have except a head start, and as Jensen said, a ton of the research advantage is skewed since a lot of the best researchers in the US are Chinese nationals. I do think the model is just one piece of the pie (not to echo Jensen too much), and hopefully we will always be able to serve these bigger frontier models in a much more efficient way as well as building out the application layer faster which actually makes them useful and/or more dangerous/powerful.
The amount of data Anthropic has claimed was extracted for distillation is tiny in comparison to the entire internet, which is right there for the taking and holds most of the knowledge people expect models to have.
Distilling even with small amounts of data from a better model is still helpful, but not in the sense of transferring capabilities the raw internet-trained model doesn't have at all, but for identifying those capabilities that are compatible with the servile assistant persona and suppressing others that are undesirable (e.g. trolling). A primitive version of this were instruction-tuning datasets generated with ChatGPT, as used e.g. for Alpaca.
Without a clear target to emulate, competitors might have to rely more on human raters, but there are plenty of data labeling companies in China, so that's hardly a hurdle.
Coding a case where it's possible to programmatically generate large amounts of data relatively cheaply. China could realistically surpass the US in coding while still being behind in many other areas.
How so? You'll soon have your choice of a very old OAI model or a new Chinese model, because the USG has no interest in letting you access the newest models without explicit permission.
Apologies - I was too quick in my response. I was speaking from a "how the users will perceive it" point of view. China's pretty good at the internet reputation thing.
Chinese frontier models don't need to catch up in every category. They just need to win in coding and that's exactly where they are going. The gap went from 12+ months to 1-2 months with the latest release of GLM 5.2 and coding is a task that you don't need heroic efforts to find rare and long-tail training data, you can just outsmart your competitor by optimizing algorithms and training recipes. This is something they can do at scale with the money and talent pool.
> They just need to win in coding and that's exactly where they are going.
They don't even need to 'win' in the sense of maxing the benchmark. They can be 20% worse/50% cheaper and many of us (and our managers who approve our token budgets) will be in.
Deepseek is 30x cheaper for input/75x cheaper for output than sonnet on openrouter, and it's not a whole lot worse for many things.
Yeah, this is, to be perfectly blunt, cope, for several reasons:
1. It's unclear if there is a law of diminishing returns with ever-larger models. They're more expensive to run and for many applications, you'll probably find smaller models are sufficient;
2. There's an inbuilt market for local LLMs. This is an effective limit on how large models can get. Case law hasn't been established yet on, for example, if a law firm using ChatGPT breaks privilege. Specifically, chat logs may be discoverable. Medical applications have this issue too and I think you'll find that financial firms are going to be leery about this as well;
3. Better, larger models will bleed into smaller, open source models. The chat logs themselves are training data. There's a whole market in China for Claude tokens around this;
4. China has a national security interest in not being beholden to US tech giants when it comes to AI. China has a history of being able to commit to large-scale long-term projects and Anthropic just won't be able to compete with a national project by one of the world's superpowers, if it comes down to it;
5. Winning doesn't necessarily mean being the best. Often it's just being good enough;
6. As an example of a national project, China is busy replicating EUV because of the US ban on ASML and NVidia exporting their best stuff. I don't think many in the West are prepared for how rapid this will be. I'm reminded of the policy debate in 1945 when many in American policy and militarey circles thought the USSR would never catch up with atomic bomb or, if they did, it would take 20+ years. It took 4 years. For the hydrogen bomb, it took 1. The US hardware advantage is a lot more tenuous than many realize.
I wonder if a lot of the companies and governments that seem to think it is essential to be on the forefront of applying leading edge LLMs to the point of starting to become dependent on them are going to find themselves in a situation like that from the Arthur C. Clarke short story "Superiority"? [1] [2].
Interesting to consider this inline with recent us export bans, could the US be squandering its lead by giving the open source, largely Chinese labs catch up (in terms of model quality available to masses), will US labs be able to maintain the lead without users being able to use their latest models?
For now, "training on frontier model conversations" are just allegations for which no evidence has been provided, while their research publications are certain evidence about their innovations.
The Americans should wake up to reality because their fantasies that are repeated continuously in all Internet media, that supposedly the Chinese copy the US technology so they will not be able to surpass it, were true many years ago, but there are already many years since this theory has become false and now there are many domains where USA would have to copy the Chinese technology if they do not want to remain behind.
Among other "sanctions", USA has forbidden the export to China of high-performance computing devices, but this has backfired as China has just demonstrated a supercomputer that is faster than any US supercomputer and which uses custom CPUs designed in China, apparently by Huawei, the company that was the main target of the US efforts to sabotage the Chinese competitors.
The US "sanctions" have hurt China for a few years, but they have convinced them that they must allocate resources to become able to make themselves everything that they previously bought from USA. The result is that now China has become stronger and USA weaker.
USA should have never sold technology to China a quarter of century ago and then the power relationship between the 2 countries would have been very different. But even 5 years ago it was already too late for any US "sanctions" to have lasting effects. Nowadays any hopes that US "sanctions" will keep China in the dark ages are pathetic.
With the kind of policies that are promoted by the US government, the chances that USA will keep its leading position in AI are minimal.
> It would be interesting to know how much of the "distillation" boost is helping the open weight models keep up.
Some people in China surely know.
> Like if the closed models stop improving will all the closed models also stop improving?
Seems extremely unlikely, unless the models all hit some kind of wall soon. The Chinese companies may be behind the US in compute capacity, but they have excellent researchers [0] who are probably approximately as good as their US counterparts at the kind of problem generation and RL that is currently working so well.
I would be very surprised, though, if the models cannot continue to be improved rapidly in any area that allows a tight feedback loop like programming, at least up to the point where we puny humans lose the ability to define objective functions.
(And, conversely, I don’t expect magic in fields where the feedback is slow or expensive. A model is not about to reliably invent a wonderful medicine for the same reason that a large and extremely competent pharma company cannot: the evaluation process is extremely slow and it’s so expensive that the kind of utterly enormous corpus that is driving the current progress in coding is simply not available. Running RL on m iterations of n medication-development trajectories each is going to cost n*m times $10-100 million and take m years if it’s even possible at all.)
[0] The US advantage in this space will likely decline, since the brain drain from the rest of the world via the US university system to US labs is drying up.
Perhaps. RL env companies based in the U.S. sell to Chinese labs quite a bit too though (though on a discount, once they're no longer on the frontier)! And it would make sense that a lot of these problems which are based on work in the U.S. enterprise economy would be coming from the U.S.
Literally no one cares. There are "full" open certified GMO free grass fed training data blah blah models. Apertus, Olmo, etc. No one cares. For all intents and purposes people use the term to describe a model that you can run locally and are allowed to modify and re-release. The rest is useless semantics. No one can "rEpRoDuCe" a model anyway.
open source vs source-available. Companies taking an extremely cautious approach to AI can't use source data that is potentially a violation of copyright (pending worldwide court decisions and/or regulation on said topic). Although that cat is already out of the bag for basically every stock-traded company using LLMs trained on non-licensed data, so I don't see there being much actual risk in using them.
I wouldn't say that no one cares, but obviously many fewer people care when the cost of "recompiling" a model from its open source training pipeline is so high. Also, if you only have the weights, you can still use it to generate training data for a new model (i.e. distillation) so it's inherently less locked down then closed source binaries were.
If the Chinese government is as involved in LLM development strategy as many people claim, wouldn't you expect them to immediately cease releasing open weight models and restrict access as soon as they start producing the frontier models? I am assuming this is what the USG thinks and is why they are trying to cut off the flow to foreign nationals ASAP.
LLMs are an undeniably valuable tool, and governments like to control those.
I talked about this before but China would be in much better position if LLMs turn into a commodidty. Where they can dominate is in hardware, as fast and cheap inference is probably going to be the moat.
I believe the open model party will eventually end. Perhaps because companies realize it’s too much of a commercial advantage, countries don’t want to give other countries commercial or military help, or maybe even an outright ban after someone uses an open model to guide them through how to make a bomb.
Frankly it does not matter if there is gap because for most practical use-cases the end user can barely perceive the difference in intelligence.
On paper frontier models will be ahead of the curve but I don't think hardly anyone will be able to tell if a piece of work, say a landing page, is created with Fable or GLM and that is the point. The perceptible intelligence will reach a point beyond which it is no longer considered, except for some narrow use-case.
Now let’s look at the economics of buying versus renting. I’ve seen a lot of attention given to hardware capital costs. But a comment the other day got me thinking about power costs, too—at what performance differential do these factors intersect to make on-prem economically competitive with datacenters for businesses?
Achilles and the tortoise [0] is usually a fallacy. If the tortoise has a head start, then Achilles will never catch it because in the time it takes Achilles to reach the tortoise's location the tortoise has moved some degree further, ad infinitum. Obviously not real because Achilles will pass the tortoise -- I think a fallacy because the framing creates a fake asymptote (they will both pass the point where they're approaching a tie).
In this case it may actually apply though, no? Open models get better from closed model distillation?
Am I the only one flagging inconsistencies in the different evaluations on the 18 benchmarks ?
Why is sometimes the closed frontier model grok ? And then opus 4.8 ? Compared to GLM 5.2 once or sometimes Kimi 2.6 ?
Are you talking about hosted vs the ones you can easily run locally? Because there are open models that require hundreds of gb of vram which are apparently pretty close.
IMHO, the biggest problem with the future of open weights models is that currently, open weights models are the result of philanthropy by some private org. (e.g. DeepSeek).
The spigot can be turned off at any time.
Until there's some sort of "community owned hardware", open weights models are always at risk of being discontinued.
Yeah, but the biggest plus for open models is that they can never be taken away. In other words, whatever capabilities they reach (even if there will never be another model), those stay forever. That can't be said for API-based models where a provider can sunset models whenever they feel like (i.e. gpt5-mini will soon be gone, and replaced by a more expensive 5.4-mini, same for goog, etc).
And there will always be incentivised parties that release models. Nvda for one has every incentive to keep the nemotron line going, as they're directly profiting from people running this. And the models aren't really far from open SotA anyway.
Goog will probably continue to release the small models, since they'll use them for browser stuff anyway, and know that they'll leak. So for them it's a win-win to release the small models and gain some dev market share.
And the chinese labs also have incentives to keep releasing models, and will likely continue to get gov support to do so (yay commercial wars between nations).
> they can never be taken away
Your right to 3d print whatever you want is about to be taken away (in California).
What software you can run on your computer can already be restricted.
Absolutely everything can be taken away. The simplest way to remove open models is probably to declare them a tool that terrorists could use. Crazy? Yes, the world is totally crazy these days.
That only affects people in California. Whereas Fable being shut down affects people all over the world.
There's also, importantly, a distinction between what are told we can no longer use, and what can actually be taken away.
Open source and open hardware can be called illegal by a government, but, if we collectively invest our energy into open alternatives, they can't be taken away in the same sense. I can build a RepRap printer and I can use a local AI model. It's on all of us to make sure that the open alternatives are viable, maybe in the current global political reality now more than ever.
Making something illegal isn't a disincentive for everyone. When they start banning books, some of us start assembling printing presses.
Just like declaring piracy illegal stopped piracy and removed pirated materials from everyone's computers.
Everything cannot, in fact, be taken away. Don't propagandize yourself. Some things, like information, are free. Not even China can prevent all its citizens from accessing Western internet. USGov simply does not have the resources to find and audit every hard drive and USB stick in the country for illegal files. The internet cannot be censored 100% without literally cutting every cable and confiscating every radio.
The software that runs on my computer cannot, in fact, be restricted. It can be declared illegal, but there literally is no mechanism by which it can be enforced other than a government goon standing over my shoulder 24/7.
Some freedoms really cannot be removed without utterly implausible amounts of effort. Arguing otherwise is helping to erode freedom. So stop it.
Remote attestation?
On PCs, the best you could really do is restrict access to certain websites on certain boxes with TPMs the users can't disable. Remote attestation can lock people out of your stuff, but not out of their own stuff. For that you need control of the device. Of course, most mobile phones aren't easy for the user to have control of, but most PCs still are, so long as you scrub the rootkits (e.g. windows) off 'em
> What software you can run on your computer can already be restricted.
Are laws that are inherently unenforceable even laws?
True, but the capabilities and knowledge of that model are also frozen in time, so the value of that model declines over time.
A model that writes code without knowledge of any language or library changes for half a decade is less useful. A 2021 era chatgpt would be quite quaint in 2026.
Right now the Chinese labs might have incentives to release their models for free, and maybe Google is happy to release open weights today, but I'm sure there are already bean counters at Google salivating at the idea of having Gemini in Chrome as part of a Google AI monthly subscription just like YouTube premium and other Google subscriptions.
Fine tuning and updating is far cheaper than training from scratch.
> Nvda for one has every incentive to keep the nemotron line going
They're releases so far have been kind of lackluster compared to Qwen and other Chinese models. My suspicion is that Nvidia won't be releasing models that appear to compete with frontier models because that would upset their big customers.
I don't think that's the case, it's not philanthropy, they are getting something out of it. The labs are learning from one another from the shared models.
Plus I am certain it makes financial sense. I am guessing here but fully utilizing a subscriptions limits probably costs the operator more money than the subscription revenue, that is why anthropic is making such a big stink about the chinese data harvesting. By releasing the weights, you are relieving yourself from that burden because the competition does not need to hammer your subscription service they can just download your model and analyze it and run it all day.
Also for the largest models it makes no sense to run it yourself unless you are a major player. Renting the hardware is ludicrously more expensive than their subscription tens of thousands of dollars. And buying the hardware to run them is in the hundreds of thousands of dollars.
We need a SETI@Home but for model training
I think model training is pretty hard to do efficiently on a vastly distributed network. If the model cant fit into the VRAM of the node your performance becomes so bad its useless, so a distributed model could only be properly trained if the size of the model doesnt exceed the majority of the nodes VRAM sizes. Maybe there is a different way of doing training but this would be the only way I can see. And it would still be much worse than just using a big datacenter where everything is fully interconnected. BOINC projects work great because its usually just a lot of small compute and memory required so every old desktop and laptop can contribute. Training a model which can compete and is not tiny requires neither low compute or low memory amount. BOINC tasks take minutes usually or sometimes hours but not weeks or months like training a model from scratch. But something like 7B or lower could maybe be trained like this. Im not sure but I think someone is already working on something like this but I dont remember the name of the project.
My understanding is that in addition to your comment and the development of a method to separate the training data for distributed learning, the latency/bandwidth of systems connected on the internet is a challenge, too. Information has to be sent around before and after any hypothetical number crunching.
This has been a (noble) goal of lots of different projects in the community for a long time. Federated learning projects like Flower have been chipping away at it for a long time. There are many many hurdles to be cleared before anything in this area is super feasible as an alternative, but I applaud everyone who works on it.
Consumer hardware over the internet is not really suitable for this, AFAIK.
There's some really early days work on making training loops robust to failure but they all have trade-offs right now.
I remain hopeful that we'll be able to democratize the entire tech stack for this tech.
Here's a project trying that - https://nousresearch.com/nous-psyche
Have been thinking about this a lot lately.
It's just a smart business decision that allows their models to compete and gain market-share against much pricier private models. No philanthropy there.
> Until there's some sort of "community owned hardware"
Or until some bright people figure out drastically more efficient means of training.
> The spigot can be turned off at any time.
True. And it's possible that this has already happened at Alibaba Qwen - at least for the smaller models that people had a chance of running at home (122B and smaller).
We'll see. The qwen team has always released a few close to sota but proprietary models in between tgeir open releases. We did get 3.6 35B and 27B so its not all set in stone yet.
Its higley unlikely we get another open llama model though after the llama4 flop, even if their muse spark seems pretty good.
Training these models is not a "hardware" problem.
I think that simplifies it a bit. You can't train without hardware, which is why the Chinese companies are illegally importing Nvidia cards [1].
[1] https://www.theinformation.com/articles/deepseek-using-banne...
This seems backwards. Access to Fable can be removed. I don't see how an open weight model can ever be put back into the bag though.
The model itself, sure; the comment is about the production of more advanced models (to keep open weights near the frontier).
The proprietary spigots can be turned off at any time also. To me, that seems more likely.
It's not pure philanthropy: https://gwern.net/complement
How is this a complaint? Once you have the model, you have the model. Download DeepSeek-R1 671B and you have it. You might not get improvements in the future, just like you may not ever get a future release of an open source project. Is that an indictment of open source?
But consider the alternative. OpenAI and Anthropic can shut off your account or API key at any time for any reason. How is this better? You have way more security when you're running your own model.
The Chinese models will not overtake the frontier US ones given the current way things are going. The US models derive their lead from incredible efforts to source more and higher quality (mostly synthetic data) via great feats (eg generating with humongous teacher models that could never feasibly serve interactive traffic). The Chinese models advance via heroic efforts to optimize models and great feats to secure more and higher quality training data from the US frontier models.
For an (Chinese) open weight model to surpass the (US lab) frontier models, this equation must flip and the Chinese labs must entirely retool from harvesting frontier model data to producing the data systems and efforts to produce novel data; as well as procuring latest generation hardware en masse for this. This does not happen easily. Also training a frontier scale model is actually not such an unimaginable feat: doing all the inference with the teacher models is where the hardware goes.
Unless you are working at one of these companies you don't know what they are doing.
You don't know what's happening in z.ai nor alibaba. And you don't know what's happening in anthropic and open ai.
I don't know what they are all doing, but I find it extremely unlikely that they are not all collecting data from one another. I am confident anthropic has a team going over GML 5.2 weights even if it's just to see where the competition is.
Just because some labs are getting data from Anthropic does not mean they are not also doing their own research.
They were focused on optimization because they could not get the best hardware.The only reason their top labs are behind may be because they did not have h200s and MI350s. And now they do.
Plus you are discounting other risks, Anthropic is currently sitting on "the best" models in the world because they got in a pissing match with the US administration.
btw: This could be the case in china as well, their administration has been surprisingly open on AI exports and open weight models, that we know of. There is a very small but not trivial chance they are hogging a better version of glm 5.2 for example, but no one is allowed to talk about it. Now I am not saying that is the case, I am saying the two cases (chinese labs are 6 months behind, they are forced to suppress their best models) are indistinguishable.
> Chinese labs must entirely retool from harvesting frontier model data to producing the data systems and efforts to produce novel data
Even if your characterization is accurate, they could do this tomorrow and are not so myopic that they wouldn’t have thought about it. I don’t see this as a barrier, and I see a lot of the same underestimation of Asia that’s been happening for 50 years. There’s not some innate American advantage to building LLMs, and personally I think whatever head start the US has is going to be squandered on delays from the export control “to dangerous for release” LARPing we’re seeing.
Exactly. If they wanted to they could produce the same amount of data. Companies like Scale, Mercor, Surge exists for a reason, a reason that doesn't need to exist in China if they mandate Chinese enterprises to provide all their real world data (or have them work inside RL environments) to the model companies for post training. There is no real advantage that US companies have except a head start, and as Jensen said, a ton of the research advantage is skewed since a lot of the best researchers in the US are Chinese nationals. I do think the model is just one piece of the pie (not to echo Jensen too much), and hopefully we will always be able to serve these bigger frontier models in a much more efficient way as well as building out the application layer faster which actually makes them useful and/or more dangerous/powerful.
Why would those have any impact on R&D speed? Most are funded and close to cash flow positive
The amount of data Anthropic has claimed was extracted for distillation is tiny in comparison to the entire internet, which is right there for the taking and holds most of the knowledge people expect models to have.
Distilling even with small amounts of data from a better model is still helpful, but not in the sense of transferring capabilities the raw internet-trained model doesn't have at all, but for identifying those capabilities that are compatible with the servile assistant persona and suppressing others that are undesirable (e.g. trolling). A primitive version of this were instruction-tuning datasets generated with ChatGPT, as used e.g. for Alpaca.
Without a clear target to emulate, competitors might have to rely more on human raters, but there are plenty of data labeling companies in China, so that's hardly a hurdle.
“China can only copy the US” is a very short sighted and uninformed opinion. there is more coming out of china than just new ways to distill models
Coding a case where it's possible to programmatically generate large amounts of data relatively cheaply. China could realistically surpass the US in coding while still being behind in many other areas.
How so? You'll soon have your choice of a very old OAI model or a new Chinese model, because the USG has no interest in letting you access the newest models without explicit permission.
Their point is that the Chinese models will also me limited to the very old OAI models, unless things flip. as they said.
The use of US models for Chinese model training is part of the motivation of all of this.
Apologies - I was too quick in my response. I was speaking from a "how the users will perceive it" point of view. China's pretty good at the internet reputation thing.
Chinese frontier models don't need to catch up in every category. They just need to win in coding and that's exactly where they are going. The gap went from 12+ months to 1-2 months with the latest release of GLM 5.2 and coding is a task that you don't need heroic efforts to find rare and long-tail training data, you can just outsmart your competitor by optimizing algorithms and training recipes. This is something they can do at scale with the money and talent pool.
> They just need to win in coding and that's exactly where they are going.
They don't even need to 'win' in the sense of maxing the benchmark. They can be 20% worse/50% cheaper and many of us (and our managers who approve our token budgets) will be in.
Deepseek is 30x cheaper for input/75x cheaper for output than sonnet on openrouter, and it's not a whole lot worse for many things.
Yeah, this is, to be perfectly blunt, cope, for several reasons:
1. It's unclear if there is a law of diminishing returns with ever-larger models. They're more expensive to run and for many applications, you'll probably find smaller models are sufficient;
2. There's an inbuilt market for local LLMs. This is an effective limit on how large models can get. Case law hasn't been established yet on, for example, if a law firm using ChatGPT breaks privilege. Specifically, chat logs may be discoverable. Medical applications have this issue too and I think you'll find that financial firms are going to be leery about this as well;
3. Better, larger models will bleed into smaller, open source models. The chat logs themselves are training data. There's a whole market in China for Claude tokens around this;
4. China has a national security interest in not being beholden to US tech giants when it comes to AI. China has a history of being able to commit to large-scale long-term projects and Anthropic just won't be able to compete with a national project by one of the world's superpowers, if it comes down to it;
5. Winning doesn't necessarily mean being the best. Often it's just being good enough;
6. As an example of a national project, China is busy replicating EUV because of the US ban on ASML and NVidia exporting their best stuff. I don't think many in the West are prepared for how rapid this will be. I'm reminded of the policy debate in 1945 when many in American policy and militarey circles thought the USSR would never catch up with atomic bomb or, if they did, it would take 20+ years. It took 4 years. For the hydrogen bomb, it took 1. The US hardware advantage is a lot more tenuous than many realize.
I wonder if a lot of the companies and governments that seem to think it is essential to be on the forefront of applying leading edge LLMs to the point of starting to become dependent on them are going to find themselves in a situation like that from the Arthur C. Clarke short story "Superiority"? [1] [2].
[1] The story: https://nob.cs.ucdavis.edu/classes/ecs153-2019-04/readings/s...
[2] Wikipedia: https://en.wikipedia.org/wiki/Superiority_(short_story)
Interesting to consider this inline with recent us export bans, could the US be squandering its lead by giving the open source, largely Chinese labs catch up (in terms of model quality available to masses), will US labs be able to maintain the lead without users being able to use their latest models?
It would be interesting to know how much of a boost the closed models companies are giving the open models.
If the closed models stop improving will the progress of open models slow?
Why are we assuming only American labs can innovate? DeepSeek already innovated a lot in efficiency, for example.
It's really unclear how much innovation DeepSeek has actually done, vs training on frontier model conversations.
For now, "training on frontier model conversations" are just allegations for which no evidence has been provided, while their research publications are certain evidence about their innovations.
The Americans should wake up to reality because their fantasies that are repeated continuously in all Internet media, that supposedly the Chinese copy the US technology so they will not be able to surpass it, were true many years ago, but there are already many years since this theory has become false and now there are many domains where USA would have to copy the Chinese technology if they do not want to remain behind.
Among other "sanctions", USA has forbidden the export to China of high-performance computing devices, but this has backfired as China has just demonstrated a supercomputer that is faster than any US supercomputer and which uses custom CPUs designed in China, apparently by Huawei, the company that was the main target of the US efforts to sabotage the Chinese competitors.
The US "sanctions" have hurt China for a few years, but they have convinced them that they must allocate resources to become able to make themselves everything that they previously bought from USA. The result is that now China has become stronger and USA weaker.
USA should have never sold technology to China a quarter of century ago and then the power relationship between the 2 countries would have been very different. But even 5 years ago it was already too late for any US "sanctions" to have lasting effects. Nowadays any hopes that US "sanctions" will keep China in the dark ages are pathetic.
With the kind of policies that are promoted by the US government, the chances that USA will keep its leading position in AI are minimal.
Wym it's unclear? They publish their research...
> It would be interesting to know how much of the "distillation" boost is helping the open weight models keep up.
Some people in China surely know.
> Like if the closed models stop improving will all the closed models also stop improving?
Seems extremely unlikely, unless the models all hit some kind of wall soon. The Chinese companies may be behind the US in compute capacity, but they have excellent researchers [0] who are probably approximately as good as their US counterparts at the kind of problem generation and RL that is currently working so well.
I would be very surprised, though, if the models cannot continue to be improved rapidly in any area that allows a tight feedback loop like programming, at least up to the point where we puny humans lose the ability to define objective functions.
(And, conversely, I don’t expect magic in fields where the feedback is slow or expensive. A model is not about to reliably invent a wonderful medicine for the same reason that a large and extremely competent pharma company cannot: the evaluation process is extremely slow and it’s so expensive that the kind of utterly enormous corpus that is driving the current progress in coding is simply not available. Running RL on m iterations of n medication-development trajectories each is going to cost n*m times $10-100 million and take m years if it’s even possible at all.)
[0] The US advantage in this space will likely decline, since the brain drain from the rest of the world via the US university system to US labs is drying up.
Perhaps. RL env companies based in the U.S. sell to Chinese labs quite a bit too though (though on a discount, once they're no longer on the frontier)! And it would make sense that a lot of these problems which are based on work in the U.S. enterprise economy would be coming from the U.S.
Article confuses open source models with open weights models.
Not the same thing.
It’s used right in the articles body, but title is misleading.
Literally no one cares. There are "full" open certified GMO free grass fed training data blah blah models. Apertus, Olmo, etc. No one cares. For all intents and purposes people use the term to describe a model that you can run locally and are allowed to modify and re-release. The rest is useless semantics. No one can "rEpRoDuCe" a model anyway.
open source vs source-available. Companies taking an extremely cautious approach to AI can't use source data that is potentially a violation of copyright (pending worldwide court decisions and/or regulation on said topic). Although that cat is already out of the bag for basically every stock-traded company using LLMs trained on non-licensed data, so I don't see there being much actual risk in using them.
No-one cares to quit social media or stop using Windows, but it’s a goal worthy of discussion all the same.
The name is bad, doesn’t even make any fucking sense and it gives open source a bad rep.
I wouldn't say that no one cares, but obviously many fewer people care when the cost of "recompiling" a model from its open source training pipeline is so high. Also, if you only have the weights, you can still use it to generate training data for a new model (i.e. distillation) so it's inherently less locked down then closed source binaries were.
I was advocating for "available weight" as a value neutral term for a while.
I gave up. No one cares. And no one will ever tell the truth about the training anyways.
Substantial and growing freedom beats zero freedom ever again.
If the Chinese government is as involved in LLM development strategy as many people claim, wouldn't you expect them to immediately cease releasing open weight models and restrict access as soon as they start producing the frontier models? I am assuming this is what the USG thinks and is why they are trying to cut off the flow to foreign nationals ASAP.
LLMs are an undeniably valuable tool, and governments like to control those.
I talked about this before but China would be in much better position if LLMs turn into a commodidty. Where they can dominate is in hardware, as fast and cheap inference is probably going to be the moat.
I believe the open model party will eventually end. Perhaps because companies realize it’s too much of a commercial advantage, countries don’t want to give other countries commercial or military help, or maybe even an outright ban after someone uses an open model to guide them through how to make a bomb.
Frankly it does not matter if there is gap because for most practical use-cases the end user can barely perceive the difference in intelligence.
On paper frontier models will be ahead of the curve but I don't think hardly anyone will be able to tell if a piece of work, say a landing page, is created with Fable or GLM and that is the point. The perceptible intelligence will reach a point beyond which it is no longer considered, except for some narrow use-case.
> except for some narrow use-case.
I think it's entirely the opposite. For narrow use cases, like web pages and crud/GUI, the open source models don't show much of a difference.
Now let’s look at the economics of buying versus renting. I’ve seen a lot of attention given to hardware capital costs. But a comment the other day got me thinking about power costs, too—at what performance differential do these factors intersect to make on-prem economically competitive with datacenters for businesses?
Achilles and the tortoise [0] is usually a fallacy. If the tortoise has a head start, then Achilles will never catch it because in the time it takes Achilles to reach the tortoise's location the tortoise has moved some degree further, ad infinitum. Obviously not real because Achilles will pass the tortoise -- I think a fallacy because the framing creates a fake asymptote (they will both pass the point where they're approaching a tie).
In this case it may actually apply though, no? Open models get better from closed model distillation?
[0] https://en.wikipedia.org/wiki/Zeno%27s_paradoxes
Am I the only one flagging inconsistencies in the different evaluations on the 18 benchmarks ? Why is sometimes the closed frontier model grok ? And then opus 4.8 ? Compared to GLM 5.2 once or sometimes Kimi 2.6 ?
at first glance, these graphs are confusing
Yea these plots are too noisy and dense. Especially that second one, lines all over the place.
Utterly unreadable on mobile
The gap is huge and im tired of reading these articles constantly
Are you talking about hosted vs the ones you can easily run locally? Because there are open models that require hundreds of gb of vram which are apparently pretty close.