> Restricting the transfer to non-trusted actors of the model weights for advanced closed-weight models. The rule does not in any way inhibit the publication of model weights for open-weight models.
> Setting security standards to protect the weights of advanced closed-weight AI models, permitting them to be stored and used securely around the world while helping prevent illicit adversary access.
Controlling the weights is becoming critical for the US…
I think it's interesting that they're actually using the term "open-weight" here instead of what AI corps would want you to use which would be "open-source", trying to conflate sources and weights. Perhaps a good sign, for a week.
Note Oracle's statement is already a week old (Jan 5th), it's not in response to this Whitehouse order. Expect we'll see a new statement from them soon.
Singapore’s Copyright Act already had the best Text and Data Mining exception. If scraped yourself, anything on the Internet is probably legal to train with.
Now, both Europe and the U.S. are trying to regulate both training and distribution of A.I. models. That might make Singapore an even better place for training A.I. models. The only question is if they have any regulations on training or distribution outside of the Copyright Act.
The linked executive order restricts exports of large numbers of GPUs to Singapore, which might make training models there hard, at least until Chinese GPU technology catches up enough.
There’s companies hosting GPU’s in Singapore. There’s also probably places without export restrictions to Singapore that have GPU’s to sell them. They might also buy them in smaller quantities from many suppliers online. Finally, they could try distributed training on the various, low-cost clouds.
A higher-risk option would be to build or buy FPGA accelerators. They’ve had lots of FPGA’s over there. They might be able to do A.I. on them. It would also be easier for their engineers to port published designs to FPGA’s than to make ASIC’s. Then, maybe make their own FPGA like some academic teams did.
You and I have to compete in this "competitive" job market thanks to AI, but we can't have US megacorps compete with China who might develop better tools used to swindle us.
In seriousness, this make Meta's decision to push open weight models incredibly valuable. Now there's no restrictions on their geographic use or "Security Standards" unlike their competitors.
The US' usual pressure tactic is to create a bluff, first raise sanctions, and use this as a bargaining chip in negotiations to gain benefits.
If you observe closely enough, you will find that every time before the U.S. government negotiates with the Chinese government, it announces new sanctions or other pressure tactics.
Biden, however, played the card that Trump was going to play prematurely.
Trump can not go 'harder', because Biden has compressed Trump's tactical space.
So, when Trump enters the White House, he will find that he has few cards left to play - Chips, Oil, Ukrain War, interest rates, debt and deficit
, etc
Instead, there'll be mines left for him by Biden admin.
As far as I can see, if Trump follows the rhetoric of "Anti-China" for security policy and "America First" trade policy, I believe they will continue with more enforcement.
>As the First Trump Administration demonstrated, America wins through innovation, competition and by sharing our technologies with the world — not by retreating behind a wall of government overreach.
Trump's tariffs plan is all about wielding power over companies. Now companies need to suck up to him to avoid tariffs as much as possible. Suck up could mean public endorsements, pressure on other companies, campaign "donations" (bribes), "investments" (bribes) etc.
It is political power for money in a direct way and it is going to get a lot worse and less subtle.
Is there a Viktor Bout of GPU smuggling? Has any journalist written about it? Whenever I see discussions of GPU restrictions I just think about what would happen if I boarded a US to China flight with a stack of H100s in my carry-on.
> Restricting the transfer to non-trusted actors of the model weights for advanced closed-weight models. The rule does not in any way inhibit the publication of model weights for open-weight models.
So the rule pre-supposes that close-weight models will just always be better than their open-weight counterparts. How do they know this will hold up?
>So the rule pre-supposes that close-weight models will just always be better than their open-weight counterparts
It doesn't necessarily presuppose that. Trying to limit the export of open weight models would likely fail on first amendment grounds, the same way the supreme court struck down attempts to limit the export of "military grade" cryptography.
The 18 countries with no restrictions are: UK, Canada, Australia, Japan, France, Germany, Belgium, Denmark, Finland, Ireland, Italy, the Netherlands, New Zealand, Norway, Republic of Korea, Spain, Sweden and Taiwan. https://www.wired.com/story/new-us-rule-aims-to-block-chinas...
Israel probably due to NPT violations (same with India) and probably Singapore due to existing compliance issues with export controlled products being exported to China or Russia (same with UAE and India).
> Chip orders with collective computation power up to roughly 1,700 advanced GPUs do not require a license
Structuring, but for gpus. Seems like the sort of loophole you could drive a truck through.
I don’t know what you’re talking about officer! I always ship just 1,699 GPUs, otherwise the box gets too full!
Officer leans over and finds another gpu right underneath the bumper... got em.
2 replies →
> Restricting the transfer to non-trusted actors of the model weights for advanced closed-weight models. The rule does not in any way inhibit the publication of model weights for open-weight models.
> Setting security standards to protect the weights of advanced closed-weight AI models, permitting them to be stored and used securely around the world while helping prevent illicit adversary access.
Controlling the weights is becoming critical for the US…
I think it's interesting that they're actually using the term "open-weight" here instead of what AI corps would want you to use which would be "open-source", trying to conflate sources and weights. Perhaps a good sign, for a week.
Related:
Oracle: Export Control Diffusion Confusion
https://news.ycombinator.com/item?id=42690300)
Note Oracle's statement is already a week old (Jan 5th), it's not in response to this Whitehouse order. Expect we'll see a new statement from them soon.
Singapore’s Copyright Act already had the best Text and Data Mining exception. If scraped yourself, anything on the Internet is probably legal to train with.
Now, both Europe and the U.S. are trying to regulate both training and distribution of A.I. models. That might make Singapore an even better place for training A.I. models. The only question is if they have any regulations on training or distribution outside of the Copyright Act.
The linked executive order restricts exports of large numbers of GPUs to Singapore, which might make training models there hard, at least until Chinese GPU technology catches up enough.
There’s companies hosting GPU’s in Singapore. There’s also probably places without export restrictions to Singapore that have GPU’s to sell them. They might also buy them in smaller quantities from many suppliers online. Finally, they could try distributed training on the various, low-cost clouds.
A higher-risk option would be to build or buy FPGA accelerators. They’ve had lots of FPGA’s over there. They might be able to do A.I. on them. It would also be easier for their engineers to port published designs to FPGA’s than to make ASIC’s. Then, maybe make their own FPGA like some academic teams did.
Are we trying to slow global competition with bad actors by restricting their ability to run compute on these.
And then we don't want GPT4o closed models going over the fence, but we are ok with Llama3.3?
I am I reading this correctly?
You and I have to compete in this "competitive" job market thanks to AI, but we can't have US megacorps compete with China who might develop better tools used to swindle us.
>Are we trying to slow global competition with bad actors by restricting their ability to run compute on these.
Huawei GPUs
> but we are ok with Llama3.3?
Considering Meta's view about the new president, the new Llama might become closed source at any moment now.
In seriousness, this make Meta's decision to push open weight models incredibly valuable. Now there's no restrictions on their geographic use or "Security Standards" unlike their competitors.
1 reply →
Which will apply for exactly one week.
It seems like Trump would only go harder on this kind of thing "as a negotiating tactic" or something.
He seems unlikely to make it easier for China, at least without something in return.
Quite the opposite
The US' usual pressure tactic is to create a bluff, first raise sanctions, and use this as a bargaining chip in negotiations to gain benefits. If you observe closely enough, you will find that every time before the U.S. government negotiates with the Chinese government, it announces new sanctions or other pressure tactics. Biden, however, played the card that Trump was going to play prematurely.
Trump can not go 'harder', because Biden has compressed Trump's tactical space.
So, when Trump enters the White House, he will find that he has few cards left to play - Chips, Oil, Ukrain War, interest rates, debt and deficit , etc Instead, there'll be mines left for him by Biden admin.
5 replies →
Nvidia responds with a Trump endorsement: https://blogs.nvidia.com/blog/ai-policy/
As far as I can see, if Trump follows the rhetoric of "Anti-China" for security policy and "America First" trade policy, I believe they will continue with more enforcement.
>As the First Trump Administration demonstrated, America wins through innovation, competition and by sharing our technologies with the world — not by retreating behind a wall of government overreach.
that's kind of a strange thing to say
Not long ago like 90% of Hacker News commenters would have agreed with the sentiment that information wants to be free.
Many of you purchased merchandise with DeCSS code written on it.
1 reply →
Maybe they are playing the card of "Trump is with us" and will undo it?
2 replies →
> that's kind of a strange thing to say
Not when your goal is to publicly fellate Donald Trump.
4 replies →
Trump's tariffs plan is all about wielding power over companies. Now companies need to suck up to him to avoid tariffs as much as possible. Suck up could mean public endorsements, pressure on other companies, campaign "donations" (bribes), "investments" (bribes) etc.
It is political power for money in a direct way and it is going to get a lot worse and less subtle.
Is this deepthroating because they are worried about his tariffs raising GPU prices in the US?
This is great news for Huawei: people outside China will finally have a reason to buy their GPUs, because a mediocre GPU is better than no GPU.
China will impose similar regulations and export restrictions.
Maybe, on the US and our 18 friends. Meaning we've accomplished nothing besides giving the rest of the world to China.
I'm reminded of oil regulations, and the days when munitions laws meant encryption couldn't be exported. Won't last long.
This reminds me of the restrictions they had around encryption.
It seems like it may hurt open source efforts.
Is there a Viktor Bout of GPU smuggling? Has any journalist written about it? Whenever I see discussions of GPU restrictions I just think about what would happen if I boarded a US to China flight with a stack of H100s in my carry-on.
GPU smuggling to China has already happened.
https://www.tomshardware.com/tech-industry/artificial-intell...
> Restricting the transfer to non-trusted actors of the model weights for advanced closed-weight models. The rule does not in any way inhibit the publication of model weights for open-weight models.
So the rule pre-supposes that close-weight models will just always be better than their open-weight counterparts. How do they know this will hold up?
>So the rule pre-supposes that close-weight models will just always be better than their open-weight counterparts
It doesn't necessarily presuppose that. Trying to limit the export of open weight models would likely fail on first amendment grounds, the same way the supreme court struck down attempts to limit the export of "military grade" cryptography.
I didn’t read it that way. I read it as trying to control what can be controlled; risk mitigation rather than risk elimination.
Who are the 18 key allies that can get our chips?
The 18 countries with no restrictions are: UK, Canada, Australia, Japan, France, Germany, Belgium, Denmark, Finland, Ireland, Italy, the Netherlands, New Zealand, Norway, Republic of Korea, Spain, Sweden and Taiwan. https://www.wired.com/story/new-us-rule-aims-to-block-chinas...
Singapore and Israel are notably absent.
Israel probably due to NPT violations (same with India) and probably Singapore due to existing compliance issues with export controlled products being exported to China or Russia (same with UAE and India).
4 replies →
Why are you singling Singapore? It's not a particularly close ally. It's like Switzerland, neutral.
14 replies →
Take a look where (US) hyperscalers have presence, they still need to be able to expand.
Only having part of the EU is an interesting decision.
1 reply →
Perhaps Switzerland is too neutral as well.
Or Poland
1 reply →
Related:
Nvidia Statement on the Biden Administration's Misguided 'AI Diffusion' Rule
https://news.ycombinator.com/item?id=42682773