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.
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.
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...
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).
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.
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.
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.
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.
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.
"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.
> 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.
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.
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.
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 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.
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.
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.
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.
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.
Man. Huawei is fucking massive, they do far, far more things than just 5G base stations (a giant business in itself) and cell phones. They build even electric cars.
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.
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.
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.
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.
Chinese companies are making great silicon, but they haven't begun exporting them in large numbers. The Chinese internal market is huge, so they will probably continue to grow there before they flood the world with cheap silicon. It is exactly what happened with BYD as well.
Readit News
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
It's already technically feasible: https://www.primeintellect.ai/blog/intellect-2
Dead Comment
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...
"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...
Deleted Comment
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.
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.
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.
just recent discussion on HN: "Small language models are the future of agentic AI"
https://news.ycombinator.com/item?id=44430311
Maybe it's not really missing and the APIs for LLMs are just too good and cheap to make homebrew stuff exciting.
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.
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
Deleted Comment
Anyway, they would need to duplicate ASML first 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.
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.