If you'd told me 15 years ago that people would be running freely-distributed software originating from research labs in China on computers self-hosted in their own house as an alternative to authoritarian overreach and control by the US in internet-based services, I'd have told you that you were smoking some good reefer.
The irony is that the authoritarian control hasn't gone away in China either, if anything it's even more advanced, the GFW and censorship regime there doesn't show any sign of letting up any time soon. Just today in my BBC news feed:
> If you'd told me 15 years ago that people would be running freely-distributed software originating from research labs in China on computers self-hosted in their own house as an alternative to authoritarian overreach and control by the US in internet-based services, I'd have told you that you were smoking some good reefer.
It’s a lot less surprising when you understand the tactic of economic dumping, which China has used for years to attack industry front-runners: https://en.wikipedia.org/wiki/Dumping_(pricing_policy)
They’re not giving away the models because they love freedom and want to generously offer gifts to the world. They’re doing it as a way to undermine the industry leaders and attract people to their brands.
> They’re doing it as a way to undermine the industry leaders and attract people to their brands.
I'm not seeing a problem here; don't all companies have this goal?
Selling a product below its cost of production can be done as a loss leader or market share tactic, but eventually you need to sell things for at least what it costs to make them.
Economic dumping is a specific scenario where one country tries to flood another country with cheap products in a way that usually has some government involvement subsidizing or incentivizing it. Countries are careful to hide these incentives or subsidies when possible because it's an invitation to trade wars. For China specifically, the state takes ownership positions in companies and has no problem forcing companies in the direction they want.
", but eventually you need to sell things for at least what it costs to make them."
Gmail, and every other free service would like a word. Net tangential profit invalidates your point. They can loss lead LLM's forever if they are getting other economic benefit from the practice
isn't blitzscaling essentially the same thing, except for when American companies do it (not always within their own country, e.g. spotify, netflix, or amazon)
It is bad when it's done by nations outside of the West
When West does same, it is capitalism for the prosperity of everyone, healthy competition
The problem isn't the goal, it's using vast money reserves to sell at a loss and hamper competition.
Isn't that just called a startup? It's not like openAI makes a profit.
Losing money as a company is fine. Losing money on each sale is often anti-competitive.
Somebody tell the Chinese companies to stop dumping AI, so US labs can keep burning pension funds money wrapped in VC checks fine dumping AI at a huge loss.
In any case I find the dumping argument stupid: they are giving both the model weights away and publishing in the open their results and how they got there.
I was just saying the general issue. This situation is much more complicated.
> They’re doing it as a way to undermine the industry leaders and attract people to their brands.
Isn't that just how competition works? Codex and Claude are both widely speculated to be sold at a loss (at least compared to API pricing), so why isn't their tactic also considered dumping? OpenAI and Anthropic aren't stealing IP and subsidizing model inference because they love freedom either. They're doing it to undermine industry competitors and attract people to their brands.
The US government deliberately chased a strategy of denying China access to powerful Nvidia hardware. Shipping efficient and cheap LLMs is their only option, just like Jensen Huang warned would happen. And now that China's smaller models are reaching the frontier, everyone cries foul.
> Codex and Claude are both widely speculated to be sold at a loss (at least compared to API pricing)
It's more likely that the API prices are highly profitable. I could see the subscription plans losing them money for some power users who actually use 100% every week, but given my past experience with subscription products I would estimate that their average utilization is far, far lower than in those comparisons where people take the tokens produced by exhausting the plan 100% and then compare to hypothetical API pricing.
As for economic dumping: The crucial difference is that there's usually some geopolitical influence happening, like government subsidies, incentives, or the government simply owning the companies directly and weaponizing their output even if the company runs at a loss.
It is probably a naive point of view that I hold (I haven't thought it through properly) but the western capitalist version is that investors providing money to do dumping of prices is ok (as long as you are not considered a monopoly) and the Chinese state sponsored capitalism is considered dumping as it goes against the agreed WTO rules for how international trade is supposed to work. The western capitalist governments (probably with a lot of lobbying support) wrote the WTO rules. Not trying to defend either way, but it seems to be different economic and political systems trying to gain dominance over each other.
That side of it is understandable, albeit open to interpretation. I can see how both sides justify their own choices in the short term.
But we hear this "dumping" argument quite often for seemingly arbitrary purposes. Is it "dumping" to sell cheap STM-32 clones when Arm LTD. won't let you buy a license? Is it "dumping" to sell $20,000 EVs to a nation that can barely build $30,000 EVs with federal and state subsidies? In many respects, China's cheap stuff is just better-suited to the global economy than America's high-margin products. If both countries are leveraging federal protectionism to promote and develop their products, it doesn't feel like State Owned Enterprises are on unfair footing compared to SpaceX or OpenAI.
I think self-hosting Chinese models is so popular precisely because China is so authoritarian.
If China wants to sell Solar panels to Americans, they can just sell solar panels, the sun won't mind. If China wants to sell models to Americans... well, Americans don't want to send their data to China, so they can't just offer them as SaaS. They don't want to be left behind in the AI race either. The only option to capture western mindshare is to do what they've always done, use Chinese taxpayers' money to subsidize model development, make American labs uncompetitive, make them go bankrupt, then have control over the entire sector.
Some US labs may become uncompetitive if they have the wrong business strategy, that doesn't mean every AI company will be. If one part of the stack gets commoditized, build your moat elsewhere.
I am not sure the comparison is exactly fitting... Specifically, solar panels in wholesale quantities have to come in 40' or 45' cargo containers by ocean from China and can be easily tariffed or blocked at the ports.
Releasing the weights of a software project on modelscope and huggingface and similar (and I'm sure they'd find a new way to distribute it for an English language audience if huggingface vanished tomorrow as well) is totally different, because there's no tangible hardware product involved.
With HF bought as well, I wouldn’t be surprised to see them try to gate open models by age too. Going to have to sign up for modelscope
It's almost been 15 years since the Snowden leaks, and there were rumors going around before that. I don't think it would have been that outlandish.
My theoretical self 15 years ago absolutely would have believed the increased authoritarian overreach part (in US/CA/European business and political context). I would not have believed the "multiple ostensibly competing Chinese research labs are giving this away free to run on your own Linux computer, and it's very close to state of the art in capability competing with US-based paid SaaS".
Today marks 25 years minus one day since the US really kicked the destruction of privacy into high gear. Soon after, framing this revocation of rights using the name PATRIOT.
It fits in with the rest of privacy in the US, which is slowly becoming reserved for the wealthy.
The 1% already have privacy consultants to help them keep their data private through shell corporations, legal trusts and other tricks. (The book 'Extreme Privacy' is from one such consultant and well worth a read)
Now if your want privacy in AI? Shell out $$$ for local hardware.
A lot of code for Esp devices, Arduino clones and other embed computers is open source coming straight of china. This was also the case 15 years ago.
Wasn't writing on the wall when Aaron Swartz got prosecuted?
And China controls AI too. It's just that their idea of "safety" is "ideological safety", and their idea of "alignment" is "alignment to the party line".
They're cool with open weight AIs being released. As long as those AIs only ever say good things about CCP, and don't mention certain concentration camps or brutally suppressed protests.
I don't disagree with you on what is the top-down political priority there, but thankfully the architecture of an open weights model released in .safetensors format allows for 3rd parties to "uncensor" it. There's at least 8 different CN originated models now that after running through heretic and a few other methods will score 0 refusals on this data set of prompts:
https://huggingface.co/datasets/mlabonne/harmful_behaviors
If we were living in a scenario where the open weight models were truly impossible to uncensor I would be significantly more skeptical of them. As a test I have an uncensored copy of qwen 3.8 27B Q8 here that will very happily discuss a myriad of negative things about the CCP.
I have basic understanding about how refusal-removal works - find the "no" weights by intentionally generating diverse refusals, and then set those weights to zero.
Is there a similar process for removing not refusals, but misinformation?
As an end user of this and not a person involved in training models or aligning them, I have only the most rudimentary understanding. But I think that would be a lot harder since the model doesn't fundamentally "know" that information is wrong.
Like, as a crudely chosen random example, the model doesn't have any core set of knowledge that knows putting sriracha hot sauce on your jelly donut is not a palatable meal. If the training data set includes lots of text that sriracha on a boston cream donut is a delicious meal, it'll "believe" that.
Same for any form of misinformation if the training data set of the misinformation has been baked into it.
There are processes for teaching a model specific facts or specific behaviors. Including "respond to topic X with Y", if that's what you want.
You could make a model that doesn't want to engage in "lunar landing was faked" conspiracy theories the same way you can make a model that doesn't want to criticize CCP.
There is, however, no broad "misinformation" category that you could tune up or down - the way there is a category of "safety refusals".
You could make a model more reluctant to say things it isn't sure about. But that is calibrated against the model's own "sure about" - and metaknowledge of this nature in LLMs? Fragile on a good day.
Yeah, it's good that open weights models can have their "filters" busted fairly reliably. Unlike whatever bone Anthropic has to pick with the very idea of biology.
But that's a consequence of how the technology works - not a consequence of China not being authoritarian about AI. They're just authoritarian about AI in different ways.
Not like they dodged the "ID verification" bullshit either. They were way ahead of the western countries there. It's vile - seeing this sad excuse of "think of the children" abused to invade privacy and strip freedoms over and over and over and over again.
Most people don't realize how tenuous the situation is with those open models too.
Right now as long as they play along with Xi it's all good. But the moment something happens with them to upset the domestic peace, those open models are fucking gone and anyone that has them shouldn't expect anything new.
> They're cool with open weight AIs being released. As long as those AIs only ever say good things about CCP, and don't mention certain concentration camps or brutally suppressed protests.
I asked recently released Qwen3.8-Flash-Next about Tiananmen Square, here's its reply:
Sounds like... it happily mentions the brutally suppressed protest? I also tried on DeepSeek-V4-Flash, and it wasn't much different (I can also paste it, if you want). Both using vanilla weights (so no special uncensored flavor).I know people like to instantly flag copied AI text but it's actually serving a point here, so I'm vouching at least.