This is really exciting! They're laying out an architecture that may mean even small players with cheap GPUs can compete with the majors. The idea implies that eventually crowd-sourcing an open AI is probably technically feasible and we've got the Chinese actively researching how to do it to a high standard that competes with the monolithic models.
I was sceptical of the US sanctions but this seems like a real win if this can be taken all the way to its logical conclusions.
mensetmanusman · 4h ago
Yeah the sanctions will (not sarcastically) actually improve the world on a number of fronts. Increasing diversity of compute, forcing decentralization of manufacturing, etc. etc.
seydor · 1h ago
also increase smuggling, theft, espionage, crime, sabotage.
China spends the normalized equivalent of America’s defense spending on suppressing their own citizens.
From a western standpoint, this is amazingly helpful because it’s a form of Chinese self destruction and waste.
avn2109 · 16m ago
>> ... spends the normalized equivalent of America’s defense spending...
I'd be interested in seeing the numbers for that claim broken down if you can cite them. From napkin math it seems hard to make the budgets line up, unless we're doing a very large purchasing power parity adjustment?
brabel · 59m ago
Why would the Chinese self destructing be “amazingly helpful “ to the West? This sounds like spiteful vitriol.
No comments yet
zeld4 · 24m ago
I'm struggling to see which is worse, using AI to police their own people, or using AI to genocide in Middle East.
Der_Einzige · 4h ago
The sanctions will (not sarcastically) massively harm the world because Nvidia may no longer be a free money cheat code. I like having an easy economic strategy for investing...
mensetmanusman · 4h ago
The world doesn’t have to optimize policy to increase the profits of a single American company.
am17an · 5h ago
Deepseek-R1 is at the level of GPT 4.1 already, it's open-weight, open-source and they even open-sourced their inference code.
jjordan · 4h ago
I don't know why everyone keeps echoing this, my experience with Deepseek-R1, from a coding perspective at least, has been underwhelming at best. Much better experience with GPT 4.1 (and even better with Claude, but that's a different price category).
am17an · 4h ago
I'm not arguing which model is better for your use-case. I'm saying in general as it's "powerful" as GPT 4.1 in a lot of benchmarks, and you can peak underneath the hood, even make it better for your said use-case
seunosewa · 4h ago
Do you mean V3? V3 is 4.1 level or above.
Zambyte · 1h ago
In my experience, all reasoning models feel (vibely) worse at structured output like code versus comparable non-reasoning models, but far better at knowledge-based answering.
jorvi · 2h ago
This is everyone with every model.
People sang praise from the roof for Google's Gemini 2.5 models, but in many things for me they can't even beat Deepseek V3.
CamperBob2 · 2h ago
What would be an example of 2.5 Pro failing against R1 (which is what you'd actually want to compare it to)?
jorvi · 2h ago
R1 sometimes fails against V3 for me too, so its not a specific dig against Gemini.
In terms of code and science, Gemini is way, way too verbose in its output, and because of that it ends up getting confused by itself and hurting the quality of longer windows.
R1 does this too, but it poisons itself in the reasoning loop. You can see it during the streaming, literally criss-crossing its thoughts and thinking itself into loops before it finally arrives at an answer.
On top of that, both R1 and Gemini Pro / Flash are mediocre at anything creative. I can accept that from R1, since it's mainly meant as more of a "hard sciences" model, but Gemini is meant to be an all-purpose model.
If you pit Gemini, Deepseek R1 and Deepseek V3 against each other in a writing contest, V3 will blow both of them out of the water.
CamperBob2 · 2h ago
Agreed on the last point, V3 is terrifyingly good at narrative writing. And yes, R1 talks itself out of correct answers almost as often as it talks itself into them.
But in general 2.5 Pro is an extremely strong model. It may lose out in some respects to o3-pro, but o3-pro is so much slower that its utility tends to be limited by my own attention span. I don't think either would have much to fear from V3, though, except possibly in the area of short fiction composition.
iJohnDoe · 4h ago
I got the impression that 03-mini or 03-mini-high were meant for coding? GPT 4.1 was meant for creative writing, not coding?
conradev · 4h ago
It’s good at a lot of things:
GPT‑4.1 scores 54.6% on SWE-bench Verified, improving by 21.4%abs over GPT‑4o and 26.6%abs over GPT‑4.5—making it a leading model for coding.
It depends on the model, probably, but there are multiple layers of censorship, one of which is the post-facto nuking these models will do online, and that goes away "for free" when you download the open weight model.
I don't have a powerful enough system to run DeepSeek, but I've tried this with some of the Qwen3 models. They'll write answers that discuss Xi Jinping (which results in an auto-nuke of the reply from Chinese-hosted models, at least DeepSeek) or other very mildly/nominally sensitive topics.
(This is probably a coarse measure to easily ensure compliance with a recent national security law that requires commercial providers of web services address sensitive topics "appropriately" or something like that, and LLMs run non-deterministically. That's why this layer of censorship often comes across as laughably extreme— it's an extreme compliance strategy that exceeds the demands of the law for the sake of guaranteeing legal safety from an unpredictable software system.)
But the same models will altogether refuse to discuss the Tiananmen Square Massacre, even locally.
Some "decensored" versions of the Qwen3 models will discuss the Tiananmen Square Massacre, but in a very concise, formulaic, "official" way. After some chatting about it, it fell into an infinite repetition of one of its short formulaic answers (a behavior I didn't see with the original Qwen3 models with the same settings).
SubiculumCode · 48m ago
I suppose its exciting, but whether that is a good thing depends entirely on how much you think AI technologies pose existential threats to human survival. This may sound hyperbolic, but serious people are seriously thinking about this and are seriously afraid.
reactordev · 3h ago
SETI@Home style peer2peer open GPU training network is something I’m looking into as well.
coolspot · 1h ago
Possible and has been done, but super-slow and inefficient resulting in long training times for small models.
To keep compute occupied you need to pass gradients very fast.
reactordev · 48s ago
Yes but could you break it up into chunks of sets of gradients to compute? I know that compute needs the full chunk to compute a set. Again, things I’m exploring but ultimately no different than just having the full dataset on disk and just scaling out compute nodes in ro mode.
Since the license ban the use and installation in EU, I would ask: It is possible to formulate a license that claims: "The restriction A is motivated to protect our ass but we will not directly or indirectly enforce it against you"?, Such kind of phrasing in the license could be categorized or called "isolating clause" but I don't know if judges could consider it a circumvention of the law.
Edited several times, I should add: IANAL, but this sounds similar to meta releasing llama weights. I think that the spirit of the European law is to control concrete uses of AI and not a broad distribution of weights and architecture. So my question is: Does the EU AI act ban this distribution?, I think it provides more competition and options for Europeans.
Edited: Thinking a little more, installing open weights could allow backdoors (in the form of a way to manipulate intelligent agents via specials prompts designated to control the system), so perhaps from a national security point of view some care should be taken (but I personally hate that). So another question: Is there a way to control if open weights can create back doors (via prompt injection)?, I recall a paper in which prompt by symbols like 0?,#2! could put the system in a state in which one can read information that the LLM is asked to hide (that is a well known attack available to those that know the weights).
Another question: Is fine tuning or Lora a way to eliminate o amilliorate such prompt attacks?, is there any python library to defend against such attacks. Download - install - modify by fine tune or lora - now you are protected.
seydor · 9h ago
It's not up to Huawei to tell EU citizens what to do. In fact they did not need to add this restriction to their license at all. As EU citizens we shoud know the laws of the land and protect ourselves by avoiding using these models like the plague.
Fluorescence · 9h ago
IANAL but the EU legislation is very broad about what it covers e.g.
"AI systems should fall within the scope of this Regulation even when they are neither placed on the market, nor put into service, nor used in the Union."
I don't really understand the limits of it's scope e.g. the difference between making a system available vs. controlling how it's used is not clear to me. I don't think you can escape the regulation of high-risk uses by offering a "general purpose" AI with no controls on how it's used.
In terms of the open-source nature - I can see it being treated like giving away any other regulated product e.g. medication, cars, safety equipment etc. The lack of cost won't transfer the liability from the supplier to the consumer.
> for example of an operator established in the Union that contracts certain services to an operator established outside the Union in relation to an activity to be performed by an AI system that would qualify as high-risk and whose effects impact natural persons located in the Union.
> this Regulation should also apply to providers and users of AI systems that are established in a third country, to the extent the output produced by those systems is used in the Union
Otherwise it seems to reach way beyond what it actually is.
Explicitly prohibiting EU usage in the license is probably a move to reduce liability under the eyes of those “used in the Union” clauses.
Fluorescence · 5h ago
It would be a misreading to think that example implies a more limited scope. The passage as a whole is pretty clear why they are they so broad: in order to avoid circumvention. I can understand why - it seems to be both a necessary yet unacceptable way to write laws!
The passage continues:
"To prevent the circumvention of this Regulation and to ensure an effective protection of natural persons located in the Union, this Regulation should also apply to providers and users of AI systems that are established in a third country, to the extent the output produced by those systems is used in the Union."
An AI would come under this regulation even it's just the outputs that are used in the EU. Interesting to think about what that could lead to.
foursilly · 8h ago
Thanks for all that information, I agree with you that the EU legislation is very broad. In my opinion, this justifies or motivates the inclusion of the ban in the EU.
hammock · 6h ago
What happens if you ignore overly broad EU regulations? Does your home country observe the EU’s violation of its sovereignty and throw you in a home country jail? Does Brussels throw you in an EU jail? What country hosts the EU jail?
Not to sound too snarky (just a little snarky), I’m just curious how it all works.
coliveira · 5h ago
As always, EU and USA courts will act as if they had jurisdiction over the rest of the world and motivate bans and similar measures against other countries.
bee_rider · 5h ago
Realistically lots of multi-national companies have an EU presence, so concerns about “violating sovereignty” are sorta moot. Huawei probably wants to do business in the EU which requires following EU law.
As an aside: in general a sovereign country can do whatever they want in their own territory, this includes the right of the country to bind itself to treaties. So in your hypothetical,
> Does your home country observe the EU’s violation of its sovereignty and throw you in a home country jail?
This doesn’t look like a violation of sovereignty to me; the non-EU has decided to enforce an EU law. Why? I don’t know, maybe it makes business easier for the multinational companies of the non-EU country.
Countries can also do things like apply secondary sanctions to an entity. So, again hypothetically, the EU doesn’t need to be able to enforce a ruling against you. They can make you toxic to anyone who wants to do business in the EU.
paganel · 4h ago
> protect ourselves
"Protect" ourselves against whom? I'm a EU citizen (unfortunately), and I'm fully on board with China against Brussels. Which is to say, don't try to speak for everyone in this God-forsaken so-called union.
chaosharmonic · 5h ago
> Since the license ban the use and installation in EU, I would ask: It is possible to formulate a license that claims: "The restriction A is motivated to protect our ass but we will not directly or indirectly enforce it against you"?, Such kind of phrasing in the license could be categorized or called "isolating clause" but I don't know if judges could consider it a circumvention of the law.
Maybe not the exact thing you're talking about, but that description reminds me of the Alliance for Open Media -- their codec licenses are royalty-free, but the same terms revoke your usage rights if you sue anyone for the use of these formats.
HPsquared · 9h ago
For security, I'd always treat ANY LLM generated code as untrusted until reviewed.
jandrese · 6h ago
100%, which makes me nervous when I hear stories about AIs that write and execute their own code. It's just asking for trouble.
tartoran · 4h ago
They'll blame it on AI from now on. This could have serious implications and further erosion of any responsibility tech companies have will probably accelerate.
coliveira · 5h ago
That's correct, independent of the source of the LLM.
foursilly · 8h ago
The problem is with AI agents when you give them some control.
Just a warning, the license [1] specifically blocks EU use:
> 3. Conditions for License Grant. You represent and warrant that You will not, access, download, install, run, deploy, integrate, modify, or otherwise use the Model, directly or indirectly, within the European Union.
Most likely EU AI act regulations they don't see any value in bothering with.
Keyframe · 4h ago
Even so, why would the licensor put it in and force it through a license. It's on the licensee to check the laws and regulations they themselves operate in.
dr_dshiv · 6h ago
Hence, the lack of European innovation in AI
p2detar · 6h ago
I don’t agree. Tools like DeepL were and still are better than Google Translate long before chat bots became a thing. The French-made Mistral AI is pretty decent as well.
seanw444 · 5h ago
Saw some benchmarks recently that put Mistral well behind basically every other competitor. Don't have them on hand unfortunately.
p2detar · 5h ago
FWIW, I refactored 500+ Junit 4 to Junit 5 tests with locally running Mistral 8B on an M3 MBP. It worked flawlessly, but surely I cannot attest for other use cases.
And who thinks that, for even a second, that an European (in this case) will not download, install, and try to run this just because the LICENSE says you can't?
FYI, this is not intended to be offensive to Europeans, I am European myself. That is not the point. The point is, who gives a damn about the LICENSE in reality, on their PERSONAL computer? Serious question.
viraptor · 10h ago
The licence is not there for enforcement from their side. It's a legal protection for Huawei. Essentially "We told you it's not for the EU. If you get sued don't try to put it on us."
Also any company of a serious size will have lawyers interested in licences of everything you're running.
johnisgood · 10h ago
I am not talking about companies. I edited my comment. Emphasis on "[their] PERSONAL" computer.
I know that companies would probably not. But individuals?
coldtea · 10h ago
It's probably the inverse:
they might license it to companies in the US, but don't want to have to deal with the changes and bureucracy needed to support individuals.
The statement's purpose is to say the equivalent "if you're a European and do run it, it's on you, this is not a product we release or support for the European market, don't expert support, liability, etc".
johnisgood · 10h ago
I get that. What I do not get it some other commenters "scaring" Europeans attempting (or thinking) to run this product.
I mean, this other commenter literally said:
> You'll be both breaking their licence and potentially your local European data laws.
benreesman · 4h ago
I'm really torn on the whole thing. I consider myself a patriotic American and would never do anything to undermine the security of my country or its allies (using the same definition of national security that the serious sworn oaths use, "all enemies foreign and domestic", which makes NSA backdoors that compromise American devices squarely a "domestic enemy").
But loyalties don't change facts and China is where serious hackers are rising on merit, doing a lot with limited resourves, giving zero fucks about empty slick talk.
If we wanted to hobble the PRC's technical rise we should have subsidized wasteful NVIDIA use and had Altman/YC be in charge: they'd still be gladhanding about how to pump their portfolio companies sticker price and avoid "systemic shocks" to the stock market anchored on NVDA.
johnisgood · 3h ago
Just for the record, I would never run this product, but it has nothing to do with the LICENSE itself.
coldtea · 2h ago
Well, people say such things even for watching pirated shows, which to be truthful, almost everybody does...
Some just are narc types.
viraptor · 10h ago
> The point is, who gives a damn about (doing an illegal thing) in reality, on their (private property where nobody is likely to see that)?
I'm not sure which part of that you find confusing. Some people will estimate benefit>risk and won't care.
johnisgood · 10h ago
What? I do not find anything confusing. You live in a Marvel world if you think a LICENSE is going to stop people from using a product. But like you said, it is not intended to be for enforcement purposes, but Huawei is trying to save its own ass.
So what is your answer? Mostly companies only? That is a fair answer, but you are the one who said this:
> You'll be both breaking their licence and potentially your local European data laws.
Again, who cares, dude? Companies might, but individuals probably give a rat's ass. So why leave that comment?
And just for the record, if you quote someone, quote them verbatim, otherwise it is not a quote.
linotype · 6h ago
Breath.
johnisgood · 5h ago
Been there, done that.
That said, I agree that it is my fault that self-contradicting virtue signaling hypocrites always find a way to irk me.
And I think it is good for the world to know that the LICENSE often means jack shit, unless when companies of significant size are involved.
Again, we all agree that they put it there to cover their own asses, not that Europeans cannot download, install, and run their product, right?
linotype · 3h ago
Yes we’re all aware of the unlimited rights of Europeans, including subjecting the rest of the world to annoying cookie notifications.
Quarrel · 10h ago
For those that would not remember, this was a real thing in the late 80s and 90s relating the cryptography.
There were serious laws limiting the export a "modern" cryptography software from the USA.
Some of us had to face up to the serious challenge of connecting to an FTP server and downloading PGP and risking violating US law to download a software package.
A few years later we had to decide "Do you want the secure Netscape, or the insecure Netscape?".
I'm sure we all chose the ethical choice.
johnisgood · 5h ago
You should elaborate on this for the unacquainted.
coldtea · 10h ago
Legalese and licenses aren't to make sure no X will download/install/or run something.
It's to make it a matter of legal record that you stated they should abstain.
Copyright warnings on music and DVDs never stopped people pirating them either.
bit1993 · 8h ago
Try selling pirated copies and see what the warnings are really about.
coldtea · 2h ago
When CDs and DVDs were a thing people wanted, there were people selling pirated copies on every corner, so...
johnisgood · 10h ago
I know. That is why I do not get what is the big fuss about.
randomNumber7 · 7h ago
Most companies abide the law. So no self hosted LLM for europeans.
johnisgood · 7h ago
... what? Self-hosted LLMs are precisely for individuals.
coliveira · 5h ago
I wonder if you'd say the same if the license were coming from Microsoft of Apple...
johnisgood · 5h ago
Why would I not? Of course I would.
randomNumber7 · 7h ago
A lot of companies and research institutes in the EU would like to be able to use a locally hosted LLM for their employees so they don't have to worry what data they give away.
They will certainly not violate EU laws and also probably not the licence.
Arnt · 10h ago
It's plausible deniability. Someone at Huawei presumably thinks there's a chance that exporting this to Europe might be a legal problem at some point in the future. So they added a restriction, enough for plausible deniability.
coldtea · 10h ago
It's not exactly "plausible deniability" in the common sense of the term.
It's not supposed to make them appear as plausibly denying that some European can download and use this.
It's role is to signal that if someone does, it's on him, not them, and he wont have any support, liability claims, etc as if they could if it was a product intended for their use.
_joel · 10h ago
Quite a few, actually.
No comments yet
pfortuny · 11h ago
Wow this is a huge caveat: a guarantee that they are using data and not complying with GDPR.
sigmoid10 · 11h ago
GDPR is not the issue here, the new AI act is. Since this is an open-weight release it is not bound by the training data disclosure rules, but it probably didn't go through the evaluation that is required above a certain number of FLOPs. That's why many recent big player model releases had a staggered release in the EU.
waffleiron · 10h ago
There is not a single AI model that fully complies with GDPR. How can you inform everyone, even those not named by actual name but otherwise identifiable, that their data is being processed and give them the ability to object when the data they train on isn’t public.
Literally the same for all other open weights, this is just legal ass covering where most others don’t even do that.
imiric · 11h ago
Shocking. At least they acknowledge it.
trhway · 10h ago
It isn't acknowledging. It is just a legalese to wash their hands away from following whatever EU restrictions and requirements may be applicable here otherwise.
2Gkashmiri · 11h ago
If you download to your PC and run locally, what will happen?
nord73 · 10h ago
Picture your PC as a cheery little planet in the EU’s cosmic backwater, sipping a digital Pan-Galactic Gargle Blaster. You download Pangu Pro MoE, hit “run,” and expect to chat with an AI wiser than Deep Thought. Instead, you’ve hailed a Vogon Demolition Fleet. Your machine starts moaning like Marvin with a hangover, your screen spews gibberish that could pass for Vogon poetry, and your poor rig might implode faster than Earth making way for a hyperspace bypass.
The fallout? This AI’s sneakier than a two-headed president—it could snitch to its creators quicker than you can say “Don’t Panic.” If they spot your EU coordinates, you’re in for a galactic stink-eye, with your setup potentially bricked or your data hitchhiking to a dodgy server at the edge of the galaxy. Worse, if the code’s got a nasty streak, your PC could end up a smoking crater, reciting bad poetry in binary.
shit_game · 10h ago
To translate for those not familiar with the writings of Douglas Adams:
nord is suggesting it's possible that the physical computer running this model could be used as a "hub" for potential spyware, or be overloaded with workloads that are not related to the actual task of running the model (and instead may be some form of malware performing other computational tasks). It could potentially perform data exfiltration, or act discriminatorily based on your percieved location (such as if you're located within the EU). At worst, data loss or firmware corruption/infection may be of concern in case of license violation.
I'm not sure I would outright disagree that this as possible, but with some caveats. I would think the reason that the license stipulates that usage within the EU is forbidden due to the EU AI Act (here is a resource to read through it: https://artificialintelligenceact.eu/ai-act-explorer/).
2Gkashmiri · 6h ago
how will the "open weights" know that the pc is running within EU?
again, you are not talking about software that actually runs in your pc but the file that the software reads and loads into memory for its own use.
tliltocatl · 9h ago
Sorry, sounds like total bullshit. The weights aren't going to do anything. And if you are worried about the code, with current deployment practices of curl | sudo bash there are much more low-hanging fruits out there. That's not even mentioning the possibility of running the model on a PC without internet access (no matter how good the new Chinese AI is, it's still not good enough yet to convince you to let it out of the box).
arccy · 6h ago
with mcp, and the right tools, it's effectively already out of the box
tliltocatl · 6h ago
Don't give it mcp then (and I struggle to understand why would anyone give a stochastic model such access even if it is trained on very American NSA-certified hardware approved by Sam Altman himself).
2Gkashmiri · 6h ago
you can use existing apps that take random huggingface files, do you expect weights to somehow coax the software to do exfiltration?
same. i call bull on this.
remember how they convinced huawei was public enemy without evidence because nokia and others were unable to compete with them?
jaggs · 5h ago
No it's actually worse. Approximately three seconds after you install the model in offline mode on your computer, a small detector van will come and park outside your door with an antenna on the roof, and relay your position to a Chinese ICBM for immediate targeting.
2Gkashmiri · 6h ago
>If they spot your EU coordinates
how.
can anyone give a technical answer how will weights get to know this fact?
viraptor · 11h ago
You'll be both breaking their licence and potentially your local European data laws.
lukan · 10h ago
" potentially your local European data laws."
If run locally, why?
johnisgood · 10h ago
I called him out in another thread. It makes absolutely no sense. He is talking against himself, judging by his comments.
To answer your question, he modified my comment (see the parentheses):
"> The point is, who gives a damn about (doing an illegal thing) in reality, on their (private property where nobody is likely to see that)?"
So... at best what he said is purely theoretical. He admitted it himself: "nobody is likely to see that". Though I am not sure I agree with it, but then again, in reality, no one gives a fuck, at least not in Europe.
viraptor · 9h ago
There are likely multiple potential issues here, but one specific example: Processing and storage of PII without consent/authorisation is not allowed, regardless of whether you do it yourself or for others. And you can't guarantee that this model does not contain private information hoovered up by accident.
2Gkashmiri · 6h ago
so what. will they send cops after you.
breaking license will do what? whats up with licenses and violations?
you and me are random people on internet
Barrin92 · 6h ago
>breaking license will do what?
The same thing breaking any license does. If you do it in your basement, nothing by definition. If you incorporate it in a service or distribute it as part of a project, well then you're on the hook. (and that is what license holders tend to care about)
seydor · 11h ago
energy consumption
HSO · 10h ago
a siren will go off and in 10 secs your computer will explode
rusk · 11h ago
The same thing that usually happens when you violate TOC …
Probably safe enough on your own computer, but could have consequences if it’s a work computer.
2Gkashmiri · 6h ago
>could have consequences if it’s a work computer.
consequences for employer who "might" get a license audit done on their machines.
does it really happen so often that a random employer in the eu would have to be concerned?
seydor · 11h ago
does anyone comply with gdpr & Ai act? Even for mistral i m not sure, the best we can say is "we don't know"
yard2010 · 11h ago
There's something nefarious about this.
knowitnone · 2h ago
I doubt the Chinese ever care about licensing so I would not care about following their license
JKCalhoun · 5h ago
I know people get upset when open source is used when open weight is more correct (happily here open weight is specifically being applied).
My question: is open weight even interesting? What does that really offer? Does it allow one to peer into the biases (or lack thereof) of a model? Does it allow one to train a competing model?
Would open source be something different and preferable — or are "weights the new source" in this LLM world we are finding ourselves in?
I'm trying to educate myself.
HPsquared · 3h ago
I really don't get why there's any confusion. These models are LITERALLY compiled binary data. Weights are definitely not source. Source is "the source from which the thing is generated" i.e. the training data (or a script to assemble it) and all scripts, procedures etc required to make the binary blob.
zeld4 · 16m ago
The silicon valley should thank Huawei and Deepseek, as the only two reasons AI exists as an industry in US are (a) reduce labor cost (b) win over China.
(a) alone sounds super negative to ordinary people, but with (b), justified by the existence and achievement of Huawei & Deepseek, the AI cause now sounds a legit jihad silicon valley is carrying.
If current LLMs hit a scaling wall and the game becomes about efficiency, I wonder if there's going to be space in the market for small models focussed on specific use cases.
I use Gemini to extract structured data from images and the flash model is great at this. I wonder how much effort it would be to create a smaller model that would run on something like a NUC with an AMD APU that is good enough for that one use case.
Or perhaps you end up with mini external GPU sticks that run use case specific models on them. Might not be much of a market for that, but could be pretty cool.
that's already the case, and it's called model distillation. You use LLMs to generate labels but then you use a dedicated smaller model (usually NN) to run at 1000x cheaper cost of inference.
crowcroft · 5h ago
I think beyond the technical aspect it's a product and packaging problem.
All the effort is in productizing foundational models and apps built on top of them, but as that plateaus distilled models and new approaches will probably get more time in the sun. I'm hopeful that if this is the case we will see more weird stuff come available.
bjord · 6h ago
throwback to that brief period where people would mine bitcoin (ineffectively) using ASICs in their USB ports
crowcroft · 5h ago
Yes, and people buying random GPUs for ether etc. I'm not a huge fan of what crypto has become but there was something exciting about hacking stuff together at home for it which is currently missing in AI IMO.
Maybe it's not really missing and the APIs for LLMs are just too good and cheap to make homebrew stuff exciting.
bjord · 5h ago
no, I think you're right—there's definitely something missing right now
but more likely it's going on and we're just not seeing it
in general, though, I think once a certain amount of money is involved, people just start to get rabid and everything becomes a lot less fun
seydor · 11h ago
Sic transit gloria nvidii
HPsquared · 11h ago
Linguistic deep lore: "invidia" is Latin for "envy".
fakedang · 10h ago
That was the reasoning behind the NVidia name choice btw - the cofounders wanted the competition to be envious of their company's capabilities.
Sanctions are at best a stopgap measure. Ideally they would buy enough time to shore up domestic capabilities.
Instead, cutting research funding and discouraging foreign students/researchers from coming to the US means that there will be depleted US capability just when China finds its groove.
randomNumber7 · 11h ago
So the Huawei Ascend 920 is produced by SMIC on a 6 nm process.
I always thought sceptical of the US sanctions, but that they backfire so fast is insane.
Out of China's perspective it might make sense to take out the wests AI capabilities soon.
bayindirh · 10h ago
Sanctions are generally a stopgap measure. They can't create any meaningful gap for a very long time.
I live in a country which has experienced some hard and soft embargoes over the years, and let's look what has it done.
- We wanted to buy drones, and denied. Now we are one of the biggest drone manufacturers in the world.
- We denied air defense systems. We are developing a whole arsenal of missiles and rockets now, incl. standoff/cruise missiles.
- We denied planes. Our 4.5th generation fighter program got a great speed boost.
- We denied advanced naval technology. We built stealth ships, fast coastguard boats and all navigational systems which goes inside them.
- We denied optical pods for drones and aerial vehicles. We built our own in 6 months.
etc. etc...
Sanctions and embargoes are the biggest catalyst for a country to advance their tech at tremendous velocity.
seydor · 9h ago
Turkey chose to develop a large defense industry. There were restrictions on them acquiring some US weapons (even though they are NATO member) but that wasn't the motivating factor, it was a political choice to develop large defense sector.
Other countries like Iran however, do develop their own drones because of sanctions
littlestymaar · 8h ago
The strict tensions imposed after Turkey's intervention in/invasion of northern Cyprus in 1974 played a key role in the development of the Turkish defense industry.
Y_Y · 10h ago
I presume you're talking about Turkey/Türkiye.
If so, do you think it makes a difference that Turkey is a NATO member, and on (relatively) good terms with the Western powers?
For all the ideological differences and geopolitcal nervousness I don't think the US or EU see themselves as potentially fighting against Turkey, and so they don't feel the need to go to the trouble of strict sanctions or sabotaging local tech.
littlestymaar · 8h ago
> and on (relatively) good terms with the Western powers?
In recent years the relationship between Turkey and Western countries has been OK-ish (though far from stellar, see the S-400-related tensions or the French-Trukish tension in the Mediterranean).
But if you look at it on a longer perspective, the relationship used to be very tense, first there was the Cyprus crisis leading to pretty harsh western sanctions on military equipment, and then the cold war between Turkey and Greece in the Aegean see, with occasional real fire air combat and casualties.
gabrielgio · 10h ago
> They can't create any meaningful gap for a very long time.
That highly depends on the size and natural resource available to the country.
heavyset_go · 10h ago
They work for countries that kneecap themselves by being at the whims of capital and the market. Countries that don't constrain themselves like that can laugh at sanctions because they don't matter in that context.
sigmoid10 · 10h ago
Who cares about 35% inflation when you got your own drones to bomb Kurds, am I right? Which btw. was the original reason why Obama didn't want to give Turkey combat capable drones. He wasn't sassy or tried to curb competition. The US presumed (correctly with the benefit of hindsight) that they would be used to go after supposed PKK positions inside and outside Turkey. Sanctions do work, but when an opponent is hell-bent on their objective, they will always be willing to sacrifice in order to attain their goals nonetheless. Even at the cost of their own population. You can see the same thing in Russia: They are literally bleeding their middle class into poverty under all those sanctions just to keep going in Ukraine.
coldtea · 10h ago
>Who cares about 35% inflation when you got your own drones to bomb Kurds, am I right?
Orthogonal. You can have 35% inflation AND no drones or good arms in general.
Which is worse, especially if this lack makes it easier to get invaded or "regime changed" into a failed style.
>Which btw. was the original reason why Obama didn't want to give Turkey combat capable drones.
The welfare of the Kurds, or using them as a proxy force against Syria, Iran, and abandoning them whenever convenient?
sigmoid10 · 10h ago
>You can have 35% inflation AND no drones or good arms in general.
I guess you can always be worse off as a country, no argument there.
>if this lack makes it easier to get invaded or "regime changed" into a failed style.
I don't understand what you mean. Are you saying it would be better if a country gets US weapons while being under threat of being "regime changed" into an enemy? Because that somehow seems even worse. And that's literally what the US was afraid of in Ukraine.
>The welfare of the Kurds
I suppose they don't care if the bombs are dropped by Turkish or US drones. But at least the sanctions delayed it by ~half a decade, so there's that.
coldtea · 2h ago
You said "who cares about 35% inflation when you got your own drones"
Based on the parent you were replying to, I take it to imply: "don't make drones to avoid sanctions (even if it means not having arms), lest you get inflation", the latter of which you paint as the worse outcome.
So, the point I made above is that if not having drones/arms etc makes it easier to be regime changed/turned into failed state, then, yes, 35% inflation would still be a small cost to pay to avoid it.
roenxi · 10h ago
> They work for countries that kneecap themselves by being at the whims of capital and the market.
Typically it actually looks like the opposite. When I look at, eg, North Korea or Iran - if I were going to try and make them wealthy it is mostly internal policies that are the problem and not external ones.
If North Korea set itself up with single digit % company and income tax combined with a strong rule of law, local education programs and a liberal economy it would barely matter what sanctions were imposed on them. A tide of money would flow in and they'd eventually be wealthy under their own power anyway if not. Although it isn't obvious why anyone would sanction a small well run country; there is a correlation between sanctions and incompetent governance.
kingkawn · 9h ago
This is the most profoundly arrogant thing I’ve read in a while, kudos! It’s hard to stand out on the internet these days
anticodon · 9h ago
Most of the sanctions are whac-a-mole game played by the West (mostly by USA): as soon as one of the peripheral countries starts developing its economy, gains more wealth, starts to produce something the competes with products of Western companies, or attempts to avoid use of USD (because it allows immense enrichment of USA simply by printing it in whatever quantities), soon there starts:
- firstly, media attack ("they're dictatorship!", "they're genociding someone", etc), preparing population for stricter measures
- secondly, color revolution, which, if successful, puts puppet malleable government in power, and makes the country ultra-poor (most of the countries, where color revolutions staged by US/West succeeded, became significantly poorer)
- if color revolution didn't work, there's always an option to just bomb the country, because it's always ignored if all the international treaties and laws are ignored if USA or Israel bomb any other nation.
littlestymaar · 5h ago
The US has a strong record of interfering with foreign countries in way that advantage their interests, but at the same time the fact that you bring the “color revolution” phrase is a strong marker that you are reading way too much Russian (imperialist) propaganda disguised as anti-imperialism, which really is no better than feeding on the neoconservative narrative.
sillystu04 · 10h ago
To play devils advocate, a lack of sanctions also allows countries to advance technology at a tremendous velocity.
The Ford Motor Company's dealings in Nazi Germany and the Soviet Union significantly helped the adoption of mass production in those countries.
firesteelrain · 9h ago
What country? Your website is out of Germany. Germany is in NATO.
perks_12 · 8h ago
Hetzner is still a thing.
akarlsten · 9h ago
Judging by his name, Turkey
firesteelrain · 9h ago
Okay, in that case it’s not a good example because Turkey is a NATO country.
littlestymaar · 8h ago
Which didn't stop the US and other western countries to embargo them after their invasion of northern Cyprus (yes it was in 1974, but it's when the Turkish domestic defense sector really started so it's not irrelevant even if it's 50 years old).
firesteelrain · 1h ago
Survivor bias
Turkey’s defense advances took decades and came with major setbacks
Not to mention many NATO incompatibilities
Just look at all the other sanctioned countries
postexitus · 9h ago
inferred from username and stated facts, Turkey.
firesteelrain · 9h ago
Sorry I didn’t pick up on it. Not familiar with Turkish names.
corimaith · 9h ago
If this was true then countries could just impose import restrictions if they wanted to "advance their tech at tremendous velocity".
randomNumber7 · 8h ago
It also depends if the country has a realistic capability to catch up with their own tech. So for Congo it would probably be a bad move. China is a different story imho.
zorked · 9h ago
This is called "import substitution" and was incredibly popular in the age before neoliberalism. If your country is rich today you probably did a lot of that in the past.
blackoil · 10h ago
10s (maybe 100s) of billions that could have gone to Nvidia are going to Huawei, so not surprising they are able to make the progress and pulling along SMIC with them. Most of the sanctions against Huawei were because they were a credible threat to US companies, so again not surprising.
seydor · 10h ago
What do we know about the production capacity of those NPUs ? asking for a friend
wordofx · 8h ago
And the spying they have been doing.
yorwba · 10h ago
You cannot just point to some amount of a sanctioned good being manufactured to claim a sanctions backfire, you need to look at the difference between with and without sanctions, and also consider the cost.
Huawei was going to work on GPUs anyway, SMIC was going to fab chips anyway. How much of the total GPU compute is the result of increased investment after sanctions, and how much was already planned? And how does it compare to the alternative of importing Nvidia GPUs for the same amount of money?
Unless Huawei is getting better performance per dollar than Nvidia, this is them implementing a costly workaround, which is the point of sanctions: increasing cost.
like_any_other · 10h ago
Increasing short-term cost, at the price of losing leverage in the medium/long-term. Or even ending up on the wrong side of that leverage, if SMIC (who was just handed control of China's domestic market) ends up outcompeting TSMC/Intel/Samsung.
phkahler · 6h ago
>> Huawei was going to work on GPUs anyway, SMIC was going to fab chips anyway.
Sure, but China has started dumping a lot of money into competing with ASML as a result of the sanctions. And no, they are not going the super complicated route of firing lasers at drops of tin to get EUV. They are trying to sidestep that costly complexity, and if they are successful they'll be providing equipment to the top chip producers (around the world or just their own).
randomNumber7 · 9h ago
Without sanctions noone would buy the Huawei chips.
With sanctions they get free money to develop better chips.
== sanctions backfire
yorwba · 8h ago
I see now, the problem is that you're confusing expenses and profit. If Huawei had spent a quadrillion dollars to make the same number of GPUs, that wouldn't be an incredible sanctions backfire (look at how much money they got!) it would be an incredible sanctions success (look at how much money they spent to get what would've been much cheaper otherwise!)
joshuaissac · 6h ago
Their customers who would have bought Nvidia chips are instead giving money to Huawei even though they prefer Nvidia. With more money, they can deliver better tech faster. It is costing these customers more, but in return, they get domestic capability with better supply chain security, and a possibility that it can eventually become cheaper than imported chips. Meanwhile, Nvidia gets less money from those lost customers.
randomNumber7 · 8h ago
I was looking at it under the assumption that the goal of the sanctions is to keep China away from beeing able to compete in AI not how much they have to pay for it.
Also the money the spend on top is mostly domestic. They can build some more powerplants to run less efficient chips. They can pay their workers and their developers to build it.
zmgsabst · 7h ago
“Compete” is an economics term.
Making it cost more is keeping them from being as competitive — which is the point of sanctions.
MangoCoffee · 5h ago
China's Made in China 2025 already lays out which key industries they will go after. Semis are always a key industry for China.
Huawei already designed and developed its own chips before sanctions, and SMIC is developing their next node before sanctions.
Whether sanctions work or not is another matter
corimaith · 9h ago
Why not just put in import restrictions then?
quantum_state · 9h ago
Sanction motivates … import restriction discourages and may not work in reality …
corimaith · 8h ago
Sounds like a weak argument to me, businesses are profit making entities, they certainly would be motivated to keep their profits irrespective of the cause.
polytely · 10h ago
Adding more sanctions seems to be the default US move it seems, to the point that there are so many sanctioned entities that they've formed a paralel world economy.
even if their chips weren't as good, what stops a company from training a large model for a very long time in less capable hardware? Is there a way to overcome memory limitations somehow?
umgefahren · 10h ago
In general yes, you can (and do) shard the model over multiple GPUs. If you want to do that yourself look at DeepSpeed or FSDP . There is a communication overhead though and the speed at which the GPUs can communicate is key. Thats where NVLink comes in btw.
So yes, it’s actually what you can and do do. However this limits your ability to iterate on the models quickly and from what I‘ve read a lot of times the foundational labs throw out their models because by the time they are done training they are already outdated.
rfoo · 5h ago
> the speed at which the GPUs can communicate is key
Guess what a telco equipment company is good at :p
checker659 · 10h ago
NVLink isn’t magic
tonyhart7 · 7h ago
china literally "steal" talent from taiwan TSMC, of course they "progress" fast, it literally the same people
codedokode · 6h ago
You mean people are ready to exchange a democracy and freedom for a tiny increase in income?
em500 · 2h ago
They're not just doing this for money... They're doing it for a shitload of money!
It's funny how it's impossible to find English coverage of this book :')
starfallg · 10h ago
SMIC is running into real problems without EUV. Just because they are able to produce something at 6/7nm, it doesn't mean that it is efficient or competitive. Right now, they do it because of strategic considerations.
randomNumber7 · 8h ago
What is so hard about EUV that they cant develop/reverse engineer themself?
dist-epoch · 8h ago
China would have invested in SMIC anyway, with or without sanctions. They consider semiconductors critical technology.
corimaith · 10h ago
Did they do it with EUV or just more multipatterning DUV? If it's the latter then it's not anything particular to worry about or unexpected.
bayindirh · 10h ago
If the chip is working, has a comparable performance and slows the gap or reverses the trend, does it really matter?
trq01758 · 10h ago
"China has EUV" would be news of the century. No doubt they are trying.
randomNumber7 · 9h ago
Why is EUV so hard to build? If the europeans deveoloped it, why cant the chinese (with or without reverse engineering)?
corimaith · 9h ago
EUV was an international effort by multiple (of the richest) countries specializing in different areas. There is no one country that does it all. When someone says China can do it, it's tantamount to saying China can outcompete the world. Maybe they can, but its not going to be an easy effort.
randomNumber7 · 8h ago
Why is the technology then owned mostly by a single dutch company (if I understand it correctly)?
HPsquared · 2h ago
It's not, at all. The EUV light source itself comes from an American company, for instance. The mirrors come from Zeiss. ASML integrate the parts into a system.
botro · 7h ago
The International Space Station was an international effort by multiple (of the richest) countries specializing in different areas. There is no one country that does it all.
China built their own.
bilbo0s · 6h ago
You would think people like HN User corimaith would have learned from the ISS fiasco.
I've gotten to the point where I no longer even listen to people who underestimate the Chinese. Whatever points they're making can be safely ignored. Indeed, strategic sense demands we ignore them. The time for underestimating the Chinese is long past. That's not the reality we live in any longer.
corimaith · 4h ago
On the contrary, the use of sanctions precisely recognizes the Chinese ability to innovate and the need to not give them our technology on a silver plate while they are it. Tens, if not hundreds of billions are already being funded into the chip industry's r&d and bringing manufacturing back home.
saubeidl · 10h ago
The only ones that have EUV are the Europeans.
Despite all the anti-EU propaganda in these parts, the entire AI stack is powered by us.
roenxi · 9h ago
If you want the anti-EU points on that subject; the EU once (the 90s) represented around 40% of the world's semiconductor fabrication. Then the Asians decided that was a bit silly and the situation should be rationalised.
Things like ASML are the tattered remnants of the EU being a powerful force in this market. China might not be able to replicate ASML, we'll see. But if the EU can compete meaningfully with the Asians at semiconductor manufacturing that'd be a shocking development.
saubeidl · 9h ago
I have a different interpretation of this. As our economy and expertise grows, we move further and further into foundational parts of the stack while leaving the implementation details to less developed countries.
All modern semiconductors need ASML machines.
All modern AI is based on research by Sepp Hochreitner.
The www is based on work done by CERN.
Most factories are run on Siemens automation systems.
We've moved on from lower-level concerns like building chips to a higher level, directionally controlling the way economies and societies develop.
patapong · 8h ago
Unfortunately, apparently these lower-level concerns also seem to include building an industry around and benefitting from these innovations...
veber-alex · 8h ago
Calling TSMC "implementation details" is delusional.
If the only thing you needed to make cutting edge chips were some machines from ASML every country in the world would be making their own chips right now.
ReptileMan · 8h ago
And it fails spectacularly when you see EU economic growth.
saubeidl · 8h ago
It only fails if you see economic growth as the metric to optimize for, which I would argue is another symptom of societies still focused on those lower levels of the stack.
Economic growth is an enabling factor for a higher level metric, quality of live. It succeeds spectacularly there.
ReptileMan · 4h ago
EU countries have enough money to either pay for overly generous social services or defense, not for both. A stronger economy would have been enough to pay for both.
saubeidl · 4h ago
Defense will pay for itself through economic stimuli.
ReptileMan · 2h ago
... and other fantastic stories
saubeidl · 8m ago
It's basic keynesianism.
delfinom · 7h ago
And the Europeans bend the knee to the US to follow sanctions on technology that the US doesn't even produce.
imtringued · 10h ago
Intel struggled with 10nm and you're shrugging off 6nm as if it was nothing?
nickysielicki · 9h ago
Process node naming is completely made up. It’s been this way for a very long time.
In fact, process node naming has always been made up.
nickysielicki · 5h ago
Not really? The physical gate length matched the node name all throughout the 2000s.
linotype · 3h ago
even that concept was in fact made up.
ReptileMan · 8h ago
Do you remember the speed with which Chinese phones went from absolute crap to legitimate flagships? Their automobiles? From having no AI as to speak of to Deepseek? China advancing at breakneck speed when they really want is something to be worried about if you want to be technologically ahead of them.
randomNumber7 · 8h ago
Their government reasons with cause and effect while the western world argues about ideologies of different colours.
If I would have to bet my money I would bet it on China.
xbmcuser · 11h ago
The world needs Huawei and China to get competitive on its node size with TSMC and Nvidia.
amelius · 10h ago
That would be great, if you ignore geopolitical concerns. Alas, AI technology is a double-edged sword and any competition in consumer space will likely be mirrored in an arms race which (given their current manufacturing capabilities, cheap labor) China would win.
Anyway, they would need to duplicate ASML first which will probably not happen in the foreseeable future.
xbmcuser · 7h ago
Don't count the Chinese out when ever they have been side lined they have copied then innovated and gotten ahead. Now their economy in real terms is actually bigger than the US economy and more well rounded so sanctions or any kind of restrictions boost their local companies by removing any competition.
And looking at how the recent wars and skirmishes in Ukraine, Israel and Pakistan/India have gone I think western military superiority is no longer real. In a conventional war of today US will lose and most likely nuke the other country so I think its best for the world if China gets so powerful that USA accepts that it no longer is the sole super power and we can avoid a nuclear war. As that is where we are heading either a multi polar world or a nuclear holocaust.
bilbo0s · 6h ago
which will probably not happen in the foreseeable future
Can you elaborate on why this will not happen in the foreseeable future?
Because in my version of the foreseeable future I see it happening quite readily.
EUV is not the magic that everyone believes it to be. It can be replicated by us, (the US), at our convenience. It can even be replicated by the Chinese and Japanese. (Personally, I'd throw the Koreans and the Russians into that pool as well.)
But that ain't even the point. The point is that, in that particular game, there is always more than one way to skin a cat so to speak. It's not at all a foregone conclusion that EUV is the best way to skin a cat, and it'd certainly be a bad bet to assume it is the only way to skin a cat. Those are the questions I would hope that we, (the US), are focusing research efforts on, and we should assume that China is also focusing research efforts in that direction.
PS - Please no one bring up Trump gutting research. Here I'm only speaking of clear strategic research priorities in an ideal, (ie - collaborative), political environment. Obviously, politically erected structural concerns impact the viability of any research strategy we want to implement. I'm just talking about what I think would be ideal.
owebmaster · 8h ago
I have the impression that if the US removed the chips exports control, the government of China would impose a imports control. They have so much more to gain from creating a real contender to Nvidia/TSMC/Apple/Google.
No comments yet
MangoCoffee · 5h ago
People on HN are so delusional it's funny.
The semiconductor industry is always a key industry for China. They laid it out in Made in China 2025.
Huawei designed and developed its own chip before any sanctions, and the Chinese government never stopped throwing money into the semiconductor industry.
In the short term, money can go to Nvidia, but it won't be long before China creates its own "Nvidia" like BYD.
The sad part is America elected Trump, and he's gutting American research and cutting out the CHIPS Act.
One country is going for the long term while another country is short sighted
Havoc · 9h ago
Time to sell nvidia shares?
Havoc · 10h ago
Strange that they'd ban EU via license but not US
123yawaworht456 · 6h ago
EU has a nebulous AI act that prohibits a myriad things.
seydor · 9h ago
They wanted to exclude EU reviewers for when publishing their paper to a journal. EU reviewers are notoriously obnoxious and often chosen to be reviewer #2.
mzl · 8h ago
Yes, as a reviewer of a paper I am assigned a number, and must by nominative determinism fulfill the expectations...
yanhangyhy · 11h ago
huawei's Ascend GPUs is the only choice for many chinese company for now. Huge win for huawei.
FirmwareBurner · 11h ago
Will the US also sanction countries buying Ascend GPUs*?
*) They're NPUs not GPUs, since they can't render graphics
chenzhekl · 11h ago
US Warns That Using Huawei AI Chip ‘Anywhere’ Breaks Its Rules
They may have misunderstood it as free from Huawei.
atlasduo · 10h ago
So, it is "buy made in US products or face sanctions" now? This will end well, I bet.
kesor · 11h ago
> The agency's Bureau of Industry and Security said in a statement Tuesday that it's also planning to warn the public about "the potential consequences of allowing US AI chips to be used for training and inference of Chinese AI models."
The A.I. war is intense, and you can clearly see who is s** their pants here.
This is just beyond stupid. Sanctions as in "not letting them use our advantage" might make some sense, but "not letting us use their advantage" is just another level of retardity.
quantum_state · 9h ago
Very much indeed …
ronsor · 11h ago
Yes.
Huawei GPUs are illegal worldwide now, according to the US government.
falcor84 · 10h ago
Cue the Team America: World Police music
firesteelrain · 9h ago
Yes because Huawei uses US technology. Huawei can’t be competitive without US and EU technology (ASML, Tokyo Electron, Applied Materials, Lam Research, KLA, Cadence, Synopsys, or Mentor, etc)
FirmwareBurner · 8h ago
Here's a thought exercise: If EU and US are the inventors of silicone tech at the basis of all products, why are they struggling to beat China in the free and open AI market and need to resort to kneecapping it?
Like, if you're so good at something why are you scared of the unproven newcomer beating you?
firesteelrain · 8h ago
The issue is not fear but strategic control. The United States and European Union form the backbone of the global semiconductor ecosystem, from design software like Cadence and Synopsys to manufacturing tools from ASML and Applied Materials. Huawei’s chips depend on this infrastructure. When entities use US-origin technology to build AI hardware, they fall under export control regardless of where the chips are used. This is not about competing in a fair market but about controlling the development and use of advanced computing in areas like defense, surveillance, and critical infrastructure.
esafak · 5h ago
You just explained why they are scared.
firesteelrain · 3h ago
Not scared - aware. Strategic assets like compute and semiconductor capability shape global influence. Controlling those assets is about ensuring they are not used to undermine national security interests. That is not fear. It is policy rooted in leverage and long-term risk assessment.
hgomersall · 10h ago
Maybe we could write some code that we can use to use them as gp-npus (graphics-processing NPUs).
I was sceptical of the US sanctions but this seems like a real win if this can be taken all the way to its logical conclusions.
There are much better ways to increase diversity
From a western standpoint, this is amazingly helpful because it’s a form of Chinese self destruction and waste.
I'd be interested in seeing the numbers for that claim broken down if you can cite them. From napkin math it seems hard to make the budgets line up, unless we're doing a very large purchasing power parity adjustment?
No comments yet
People sang praise from the roof for Google's Gemini 2.5 models, but in many things for me they can't even beat Deepseek V3.
In terms of code and science, Gemini is way, way too verbose in its output, and because of that it ends up getting confused by itself and hurting the quality of longer windows.
R1 does this too, but it poisons itself in the reasoning loop. You can see it during the streaming, literally criss-crossing its thoughts and thinking itself into loops before it finally arrives at an answer.
On top of that, both R1 and Gemini Pro / Flash are mediocre at anything creative. I can accept that from R1, since it's mainly meant as more of a "hard sciences" model, but Gemini is meant to be an all-purpose model.
If you pit Gemini, Deepseek R1 and Deepseek V3 against each other in a writing contest, V3 will blow both of them out of the water.
But in general 2.5 Pro is an extremely strong model. It may lose out in some respects to o3-pro, but o3-pro is so much slower that its utility tends to be limited by my own attention span. I don't think either would have much to fear from V3, though, except possibly in the area of short fiction composition.
https://www.socialscience.international/making-deepseek-spea...
I don't have a powerful enough system to run DeepSeek, but I've tried this with some of the Qwen3 models. They'll write answers that discuss Xi Jinping (which results in an auto-nuke of the reply from Chinese-hosted models, at least DeepSeek) or other very mildly/nominally sensitive topics.
(This is probably a coarse measure to easily ensure compliance with a recent national security law that requires commercial providers of web services address sensitive topics "appropriately" or something like that, and LLMs run non-deterministically. That's why this layer of censorship often comes across as laughably extreme— it's an extreme compliance strategy that exceeds the demands of the law for the sake of guaranteeing legal safety from an unpredictable software system.)
But the same models will altogether refuse to discuss the Tiananmen Square Massacre, even locally.
Some "decensored" versions of the Qwen3 models will discuss the Tiananmen Square Massacre, but in a very concise, formulaic, "official" way. After some chatting about it, it fell into an infinite repetition of one of its short formulaic answers (a behavior I didn't see with the original Qwen3 models with the same settings).
https://blog.lambdaclass.com/introducing-demo-decoupled-mome...
It's already technically feasible: https://www.primeintellect.ai/blog/intellect-2
Edited several times, I should add: IANAL, but this sounds similar to meta releasing llama weights. I think that the spirit of the European law is to control concrete uses of AI and not a broad distribution of weights and architecture. So my question is: Does the EU AI act ban this distribution?, I think it provides more competition and options for Europeans.
Edited: Thinking a little more, installing open weights could allow backdoors (in the form of a way to manipulate intelligent agents via specials prompts designated to control the system), so perhaps from a national security point of view some care should be taken (but I personally hate that). So another question: Is there a way to control if open weights can create back doors (via prompt injection)?, I recall a paper in which prompt by symbols like 0?,#2! could put the system in a state in which one can read information that the LLM is asked to hide (that is a well known attack available to those that know the weights).
Another question: Is fine tuning or Lora a way to eliminate o amilliorate such prompt attacks?, is there any python library to defend against such attacks. Download - install - modify by fine tune or lora - now you are protected.
"AI systems should fall within the scope of this Regulation even when they are neither placed on the market, nor put into service, nor used in the Union."
I don't really understand the limits of it's scope e.g. the difference between making a system available vs. controlling how it's used is not clear to me. I don't think you can escape the regulation of high-risk uses by offering a "general purpose" AI with no controls on how it's used.
In terms of the open-source nature - I can see it being treated like giving away any other regulated product e.g. medication, cars, safety equipment etc. The lack of cost won't transfer the liability from the supplier to the consumer.
https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52...
> for example of an operator established in the Union that contracts certain services to an operator established outside the Union in relation to an activity to be performed by an AI system that would qualify as high-risk and whose effects impact natural persons located in the Union.
> this Regulation should also apply to providers and users of AI systems that are established in a third country, to the extent the output produced by those systems is used in the Union
Otherwise it seems to reach way beyond what it actually is.
Explicitly prohibiting EU usage in the license is probably a move to reduce liability under the eyes of those “used in the Union” clauses.
The passage continues:
"To prevent the circumvention of this Regulation and to ensure an effective protection of natural persons located in the Union, this Regulation should also apply to providers and users of AI systems that are established in a third country, to the extent the output produced by those systems is used in the Union."
An AI would come under this regulation even it's just the outputs that are used in the EU. Interesting to think about what that could lead to.
Not to sound too snarky (just a little snarky), I’m just curious how it all works.
As an aside: in general a sovereign country can do whatever they want in their own territory, this includes the right of the country to bind itself to treaties. So in your hypothetical,
> Does your home country observe the EU’s violation of its sovereignty and throw you in a home country jail?
This doesn’t look like a violation of sovereignty to me; the non-EU has decided to enforce an EU law. Why? I don’t know, maybe it makes business easier for the multinational companies of the non-EU country.
Countries can also do things like apply secondary sanctions to an entity. So, again hypothetically, the EU doesn’t need to be able to enforce a ruling against you. They can make you toxic to anyone who wants to do business in the EU.
"Protect" ourselves against whom? I'm a EU citizen (unfortunately), and I'm fully on board with China against Brussels. Which is to say, don't try to speak for everyone in this God-forsaken so-called union.
Maybe not the exact thing you're talking about, but that description reminds me of the Alliance for Open Media -- their codec licenses are royalty-free, but the same terms revoke your usage rights if you sue anyone for the use of these formats.
[1]: https://gitcode.com/ascend-tribe/pangu-pro-moe-model
> 3. Conditions for License Grant. You represent and warrant that You will not, access, download, install, run, deploy, integrate, modify, or otherwise use the Model, directly or indirectly, within the European Union.
[1] https://gitcode.com/ascend-tribe/pangu-pro-moe-model/blob/ma...
Edit: it was 8B, not 7B.
[0]: https://dionwiggins.substack.com/p/llama-4-is-banned-in-the-...
FYI, this is not intended to be offensive to Europeans, I am European myself. That is not the point. The point is, who gives a damn about the LICENSE in reality, on their PERSONAL computer? Serious question.
Also any company of a serious size will have lawyers interested in licences of everything you're running.
I know that companies would probably not. But individuals?
they might license it to companies in the US, but don't want to have to deal with the changes and bureucracy needed to support individuals.
The statement's purpose is to say the equivalent "if you're a European and do run it, it's on you, this is not a product we release or support for the European market, don't expert support, liability, etc".
I mean, this other commenter literally said:
> You'll be both breaking their licence and potentially your local European data laws.
But loyalties don't change facts and China is where serious hackers are rising on merit, doing a lot with limited resourves, giving zero fucks about empty slick talk.
If we wanted to hobble the PRC's technical rise we should have subsidized wasteful NVIDIA use and had Altman/YC be in charge: they'd still be gladhanding about how to pump their portfolio companies sticker price and avoid "systemic shocks" to the stock market anchored on NVDA.
Some just are narc types.
I'm not sure which part of that you find confusing. Some people will estimate benefit>risk and won't care.
So what is your answer? Mostly companies only? That is a fair answer, but you are the one who said this:
> You'll be both breaking their licence and potentially your local European data laws.
Again, who cares, dude? Companies might, but individuals probably give a rat's ass. So why leave that comment?
And just for the record, if you quote someone, quote them verbatim, otherwise it is not a quote.
That said, I agree that it is my fault that self-contradicting virtue signaling hypocrites always find a way to irk me.
And I think it is good for the world to know that the LICENSE often means jack shit, unless when companies of significant size are involved.
Again, we all agree that they put it there to cover their own asses, not that Europeans cannot download, install, and run their product, right?
There were serious laws limiting the export a "modern" cryptography software from the USA.
Some of us had to face up to the serious challenge of connecting to an FTP server and downloading PGP and risking violating US law to download a software package.
A few years later we had to decide "Do you want the secure Netscape, or the insecure Netscape?".
I'm sure we all chose the ethical choice.
It's to make it a matter of legal record that you stated they should abstain.
Copyright warnings on music and DVDs never stopped people pirating them either.
They will certainly not violate EU laws and also probably not the licence.
It's not supposed to make them appear as plausibly denying that some European can download and use this.
It's role is to signal that if someone does, it's on him, not them, and he wont have any support, liability claims, etc as if they could if it was a product intended for their use.
No comments yet
Literally the same for all other open weights, this is just legal ass covering where most others don’t even do that.
The fallout? This AI’s sneakier than a two-headed president—it could snitch to its creators quicker than you can say “Don’t Panic.” If they spot your EU coordinates, you’re in for a galactic stink-eye, with your setup potentially bricked or your data hitchhiking to a dodgy server at the edge of the galaxy. Worse, if the code’s got a nasty streak, your PC could end up a smoking crater, reciting bad poetry in binary.
nord is suggesting it's possible that the physical computer running this model could be used as a "hub" for potential spyware, or be overloaded with workloads that are not related to the actual task of running the model (and instead may be some form of malware performing other computational tasks). It could potentially perform data exfiltration, or act discriminatorily based on your percieved location (such as if you're located within the EU). At worst, data loss or firmware corruption/infection may be of concern in case of license violation.
I'm not sure I would outright disagree that this as possible, but with some caveats. I would think the reason that the license stipulates that usage within the EU is forbidden due to the EU AI Act (here is a resource to read through it: https://artificialintelligenceact.eu/ai-act-explorer/).
same. i call bull on this.
remember how they convinced huawei was public enemy without evidence because nokia and others were unable to compete with them?
can anyone give a technical answer how will weights get to know this fact?
If run locally, why?
To answer your question, he modified my comment (see the parentheses):
"> The point is, who gives a damn about (doing an illegal thing) in reality, on their (private property where nobody is likely to see that)?"
So... at best what he said is purely theoretical. He admitted it himself: "nobody is likely to see that". Though I am not sure I agree with it, but then again, in reality, no one gives a fuck, at least not in Europe.
breaking license will do what? whats up with licenses and violations? you and me are random people on internet
The same thing breaking any license does. If you do it in your basement, nothing by definition. If you incorporate it in a service or distribute it as part of a project, well then you're on the hook. (and that is what license holders tend to care about)
Probably safe enough on your own computer, but could have consequences if it’s a work computer.
consequences for employer who "might" get a license audit done on their machines.
does it really happen so often that a random employer in the eu would have to be concerned?
My question: is open weight even interesting? What does that really offer? Does it allow one to peer into the biases (or lack thereof) of a model? Does it allow one to train a competing model?
Would open source be something different and preferable — or are "weights the new source" in this LLM world we are finding ourselves in?
I'm trying to educate myself.
(a) alone sounds super negative to ordinary people, but with (b), justified by the existence and achievement of Huawei & Deepseek, the AI cause now sounds a legit jihad silicon valley is carrying.
Here is the GIT https://github.com/Tencent-Hunyuan/Hunyuan-A13B
I use Gemini to extract structured data from images and the flash model is great at this. I wonder how much effort it would be to create a smaller model that would run on something like a NUC with an AMD APU that is good enough for that one use case.
Or perhaps you end up with mini external GPU sticks that run use case specific models on them. Might not be much of a market for that, but could be pretty cool.
Its only 108 million params.
All the effort is in productizing foundational models and apps built on top of them, but as that plateaus distilled models and new approaches will probably get more time in the sun. I'm hopeful that if this is the case we will see more weird stuff come available.
Maybe it's not really missing and the APIs for LLMs are just too good and cheap to make homebrew stuff exciting.
but more likely it's going on and we're just not seeing it
in general, though, I think once a certain amount of money is involved, people just start to get rabid and everything becomes a lot less fun
https://en.m.wikipedia.org/wiki/Jensen_Huang
Bravo.
Does this mean that Huawei phone which has been hurt badly by sanction will now stand a fighting chance because of homegrown GPU?
How good or bad these GPU compares to the SOTA GPU in the west?
And does this mean that Huawei has the ability to crank out the GPU commercially?
oh, man, "stand a fighting chance"? huawei phone sales has already been back and surpassed apple in china.
https://www.counterpointresearch.com/insight/china-smartphon...
I've been using phones without Google Play for years.
My point is that YMMV based on where you are.
https://www.youtube.com/watch?v=NMFdYFmeiBA
Instead, cutting research funding and discouraging foreign students/researchers from coming to the US means that there will be depleted US capability just when China finds its groove.
I always thought sceptical of the US sanctions, but that they backfire so fast is insane.
Out of China's perspective it might make sense to take out the wests AI capabilities soon.
I live in a country which has experienced some hard and soft embargoes over the years, and let's look what has it done.
- We wanted to buy drones, and denied. Now we are one of the biggest drone manufacturers in the world.
- We denied air defense systems. We are developing a whole arsenal of missiles and rockets now, incl. standoff/cruise missiles.
- We denied planes. Our 4.5th generation fighter program got a great speed boost.
- We denied advanced naval technology. We built stealth ships, fast coastguard boats and all navigational systems which goes inside them.
- We denied optical pods for drones and aerial vehicles. We built our own in 6 months.
etc. etc...
Sanctions and embargoes are the biggest catalyst for a country to advance their tech at tremendous velocity.
Other countries like Iran however, do develop their own drones because of sanctions
If so, do you think it makes a difference that Turkey is a NATO member, and on (relatively) good terms with the Western powers?
For all the ideological differences and geopolitcal nervousness I don't think the US or EU see themselves as potentially fighting against Turkey, and so they don't feel the need to go to the trouble of strict sanctions or sabotaging local tech.
In recent years the relationship between Turkey and Western countries has been OK-ish (though far from stellar, see the S-400-related tensions or the French-Trukish tension in the Mediterranean).
But if you look at it on a longer perspective, the relationship used to be very tense, first there was the Cyprus crisis leading to pretty harsh western sanctions on military equipment, and then the cold war between Turkey and Greece in the Aegean see, with occasional real fire air combat and casualties.
That highly depends on the size and natural resource available to the country.
Orthogonal. You can have 35% inflation AND no drones or good arms in general.
Which is worse, especially if this lack makes it easier to get invaded or "regime changed" into a failed style.
>Which btw. was the original reason why Obama didn't want to give Turkey combat capable drones.
The welfare of the Kurds, or using them as a proxy force against Syria, Iran, and abandoning them whenever convenient?
I guess you can always be worse off as a country, no argument there.
>if this lack makes it easier to get invaded or "regime changed" into a failed style.
I don't understand what you mean. Are you saying it would be better if a country gets US weapons while being under threat of being "regime changed" into an enemy? Because that somehow seems even worse. And that's literally what the US was afraid of in Ukraine.
>The welfare of the Kurds
I suppose they don't care if the bombs are dropped by Turkish or US drones. But at least the sanctions delayed it by ~half a decade, so there's that.
Based on the parent you were replying to, I take it to imply: "don't make drones to avoid sanctions (even if it means not having arms), lest you get inflation", the latter of which you paint as the worse outcome.
So, the point I made above is that if not having drones/arms etc makes it easier to be regime changed/turned into failed state, then, yes, 35% inflation would still be a small cost to pay to avoid it.
Typically it actually looks like the opposite. When I look at, eg, North Korea or Iran - if I were going to try and make them wealthy it is mostly internal policies that are the problem and not external ones.
If North Korea set itself up with single digit % company and income tax combined with a strong rule of law, local education programs and a liberal economy it would barely matter what sanctions were imposed on them. A tide of money would flow in and they'd eventually be wealthy under their own power anyway if not. Although it isn't obvious why anyone would sanction a small well run country; there is a correlation between sanctions and incompetent governance.
- firstly, media attack ("they're dictatorship!", "they're genociding someone", etc), preparing population for stricter measures
- secondly, color revolution, which, if successful, puts puppet malleable government in power, and makes the country ultra-poor (most of the countries, where color revolutions staged by US/West succeeded, became significantly poorer)
- if color revolution didn't work, there's always an option to just bomb the country, because it's always ignored if all the international treaties and laws are ignored if USA or Israel bomb any other nation.
The Ford Motor Company's dealings in Nazi Germany and the Soviet Union significantly helped the adoption of mass production in those countries.
Turkey’s defense advances took decades and came with major setbacks
Not to mention many NATO incompatibilities
Just look at all the other sanctioned countries
Huawei was going to work on GPUs anyway, SMIC was going to fab chips anyway. How much of the total GPU compute is the result of increased investment after sanctions, and how much was already planned? And how does it compare to the alternative of importing Nvidia GPUs for the same amount of money?
Unless Huawei is getting better performance per dollar than Nvidia, this is them implementing a costly workaround, which is the point of sanctions: increasing cost.
Sure, but China has started dumping a lot of money into competing with ASML as a result of the sanctions. And no, they are not going the super complicated route of firing lasers at drops of tin to get EUV. They are trying to sidestep that costly complexity, and if they are successful they'll be providing equipment to the top chip producers (around the world or just their own).
With sanctions they get free money to develop better chips.
== sanctions backfire
Also the money the spend on top is mostly domestic. They can build some more powerplants to run less efficient chips. They can pay their workers and their developers to build it.
Making it cost more is keeping them from being as competitive — which is the point of sanctions.
Huawei already designed and developed its own chips before sanctions, and SMIC is developing their next node before sanctions.
Whether sanctions work or not is another matter
https://www.theatlantic.com/international/archive/2018/08/wh...
Guess what a telco equipment company is good at :p
It's funny how it's impossible to find English coverage of this book :')
China built their own.
I've gotten to the point where I no longer even listen to people who underestimate the Chinese. Whatever points they're making can be safely ignored. Indeed, strategic sense demands we ignore them. The time for underestimating the Chinese is long past. That's not the reality we live in any longer.
Despite all the anti-EU propaganda in these parts, the entire AI stack is powered by us.
Things like ASML are the tattered remnants of the EU being a powerful force in this market. China might not be able to replicate ASML, we'll see. But if the EU can compete meaningfully with the Asians at semiconductor manufacturing that'd be a shocking development.
All modern semiconductors need ASML machines.
All modern AI is based on research by Sepp Hochreitner.
The www is based on work done by CERN.
Most factories are run on Siemens automation systems.
We've moved on from lower-level concerns like building chips to a higher level, directionally controlling the way economies and societies develop.
If the only thing you needed to make cutting edge chips were some machines from ASML every country in the world would be making their own chips right now.
Economic growth is an enabling factor for a higher level metric, quality of live. It succeeds spectacularly there.
https://community.cadence.com/cadence_blogs_8/b/breakfast-by...
It has nothing to do with the size of anything.
If I would have to bet my money I would bet it on China.
Anyway, they would need to duplicate ASML first which will probably not happen in the foreseeable future.
And looking at how the recent wars and skirmishes in Ukraine, Israel and Pakistan/India have gone I think western military superiority is no longer real. In a conventional war of today US will lose and most likely nuke the other country so I think its best for the world if China gets so powerful that USA accepts that it no longer is the sole super power and we can avoid a nuclear war. As that is where we are heading either a multi polar world or a nuclear holocaust.
Can you elaborate on why this will not happen in the foreseeable future?
Because in my version of the foreseeable future I see it happening quite readily.
EUV is not the magic that everyone believes it to be. It can be replicated by us, (the US), at our convenience. It can even be replicated by the Chinese and Japanese. (Personally, I'd throw the Koreans and the Russians into that pool as well.)
But that ain't even the point. The point is that, in that particular game, there is always more than one way to skin a cat so to speak. It's not at all a foregone conclusion that EUV is the best way to skin a cat, and it'd certainly be a bad bet to assume it is the only way to skin a cat. Those are the questions I would hope that we, (the US), are focusing research efforts on, and we should assume that China is also focusing research efforts in that direction.
PS - Please no one bring up Trump gutting research. Here I'm only speaking of clear strategic research priorities in an ideal, (ie - collaborative), political environment. Obviously, politically erected structural concerns impact the viability of any research strategy we want to implement. I'm just talking about what I think would be ideal.
No comments yet
The semiconductor industry is always a key industry for China. They laid it out in Made in China 2025.
Huawei designed and developed its own chip before any sanctions, and the Chinese government never stopped throwing money into the semiconductor industry.
In the short term, money can go to Nvidia, but it won't be long before China creates its own "Nvidia" like BYD.
The sad part is America elected Trump, and he's gutting American research and cutting out the CHIPS Act.
One country is going for the long term while another country is short sighted
*) They're NPUs not GPUs, since they can't render graphics
https://www.bloomberg.com/news/articles/2025-05-13/us-warns-...
The A.I. war is intense, and you can clearly see who is s** their pants here.
Huawei GPUs are illegal worldwide now, according to the US government.
Like, if you're so good at something why are you scared of the unproven newcomer beating you?