• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

White House Whipsaws Silicon Valley (and Itself) Over A.I. Rules

P

philb2

2[H]4U
Joined
May 26, 2021
Messages
3,521
https://www.nytimes.com/2026/08/04/...gulation-whiplash.html?smid=nytcore-ios-share

Senior officials have considered a range of actions, including using sanctions, a trade blacklist against Chinese companies that make these open-source A.I. models, or even banning U.S. cloud companies from doing business with them, said five of the people, all of whom spoke on condition of anonymity to discuss private matters. But after an outcry from Silicon Valley, they appear to have changed their minds and instead are focused on promoting American A.I. models to be more competitive, the people said.

The most recent debates were prompted by a surge in powerful new Chinese A.I. models, which are known as “open source,” where the public can see the underlying computer code, or “open weight,” where companies reveal the calculations used to generate answers to questions.
 
philb2 said:
https://www.nytimes.com/2026/08/04/...gulation-whiplash.html?smid=nytcore-ios-share

Senior officials have considered a range of actions, including using sanctions, a trade blacklist against Chinese companies that make these open-source A.I. models, or even banning U.S. cloud companies from doing business with them, said five of the people, all of whom spoke on condition of anonymity to discuss private matters. But after an outcry from Silicon Valley, they appear to have changed their minds and instead are focused on promoting American A.I. models to be more competitive, the people said.

The most recent debates were prompted by a surge in powerful new Chinese A.I. models, which are known as “open source,” where the public can see the underlying computer code, or “open weight,” where companies reveal the calculations used to generate answers to questions.
Click to expand...
How can something out of China be open source? It's literally ccp controlled
 
tunatime said:
How can something out of China be open source? It's literally ccp controlled
Click to expand...
they are usually not open source but open weight (for the name people know about, like deepseek, k3, etc... they often start open source but get closed the moment they become popular some did stop publishing how they do things has well and end up closed-proprietary running in China like ChatGPT is here), Nvidia I think is pretty much the only one making open source model among the big name.

Has for being CCP controlled or not, not sure what would it mean regarding being open source or not, would they be open source and the code-data available, well they would be.
 
LukeTbk said:
they are usually not open source but open weight (for the name people know about, like deepseek, k3, etc... they often start open source but get closed the moment they become popular some did stop publishing how they do things has well and end up closed-proprietary running in China like ChatGPT is here), Nvidia I think is pretty much the only one making open source model among the big name.

Has for being CCP controlled or not, not sure what would it mean regarding being open source or not, would they be open source and the code-data available, well they would be.
Click to expand...
Whatever. The CCP controls how various Chinese models respond to "sensitive" queries, like Tienanman Square 1989, Hong Kong, etc.
 
philb2 said:
Whatever. The CCP controls how various Chinese models respond to "sensitive" queries, like Tienanman Square 1989, Hong Kong, etc.
Click to expand...
That more controlling the hosting harness than the model, if you host it yourself or use a non china host of them, they are quite more open about those things.

And models are way more today than google search-wikipedia knowledge compression (that a bit 2023 talking point), they are work agents, you want some model to control robots decisions making, use blender to make 3d assets or to solidwork to be 3d printed model or some network routing or log stack analysis, political view of it will not necessarily matter at all.
 
Considering you have MS even signing that they should be allowed, tells you the state of AI in the U.S, all about wanting to protect Anthropic and OpenAI so they can try to make their investors some money back...
 
MrGuvernment said:
Considering you have MS even signing that they should be allowed, tells you the state of AI in the U.S, all about wanting to protect Anthropic and OpenAI so they can try to make their investors some money back...
Click to expand...
Im still not convinced AI was ever about making anybody money and not deep sate level control.
 
  • Like
Reactions: cycy
like this
LukeTbk said:
That more controlling the hosting harness than the model, if you host it yourself or use a non china host of them, they are quite more open about those things.

And models are way more today than google search-wikipedia knowledge compression (that a bit 2023 talking point), they are work agents, you want some model to control robots decisions making, use blender to make 3d assets or to solidwork to be 3d printed model or some network routing or log stack analysis, political view of it will not necessarily matter at all.
Click to expand...
Theoretically, it could matter. If, for instance, it figured out your country of residence/operation or employer, and it could somehow be influenced to act a particular way based on that information.
 
tunatime said:
Im still not convinced AI was ever about making anybody money and not deep sate level control.
Click to expand...
The way things like Google are going replacing search with curated AI results, and how many people take AI results at face value, would say it is working pretty dam well!
 
LukeTbk said:
That more controlling the hosting harness than the model, if you host it yourself or use a non china host of them, they are quite more open about those things.

And models are way more today than google search-wikipedia knowledge compression (that a bit 2023 talking point), they are work agents, you want some model to control robots decisions making, use blender to make 3d assets or to solidwork to be 3d printed model or some network routing or log stack analysis, political view of it will not necessarily matter at all.
Click to expand...
Also with open weight models running on your hardware: You can re-train them. If there's an issue you can make sort of a "training patch" that changes behavior. You see this with image generators all the time with LORA files where someone makes an addon that changes how it works in a certain way. You can also fully mess with it directly. You can see that with the abliterated ones that have had their refusals burned out. They are tweaked until their safety training is removed.

You have a hell of a lot more control over what it does, if you want to spend the time training it, and you can verify custody of your data. It is a red herring to scream "But China!!!111one" as though that means it is somehow bad, when the US models, with the exception of nVidia's Nemotron, are all completely closed: You don't know the weights, you don't know the training, you can't run them on your hardware, etc. For all you know OpenAI could be sharing everything you do with multiple governments, it runs on their servers you'd have no way to audit it.


I'm not trying to fanboy China here but let's be real about what is more risk. No I wouldn't trust a Chinese model to give me accurate information about China. I also wouldn't trust ChatGPT to give me accurate information about Sam Altman.
 
MrGuvernment said:
Considering you have MS even signing that they should be allowed, tells you the state of AI in the U.S, all about wanting to protect Anthropic and OpenAI so they can try to make their investors some money back...
Click to expand...
Even OpenAI did sign it....

I doubt the whitehouse are that "pro" anthropic and its investor, there is a real the USA need to win the AI race and general anti-china sentiment that do exist I think, not all made up. Same for the fear about what those system can do, it could be actually be in part real, not pure synical to help choose the winners.

Sycraft said:
when the US models, with the exception of nVidia's Nemotron, are all completely closed:
Click to expand...
they did fall behind but meta was a big player in open weight, I think they are still going.

IF you do not go for flagship level there was nice option at least in 2025...

Phi series from microsoft, gemma from google and Llama from meta, I think microsoft even use MIT license. Maybe they are all a bit old and too much behind chinese option by now
 
Last edited:
philb2 said:
https://www.nytimes.com/2026/08/04/...gulation-whiplash.html?smid=nytcore-ios-share

Senior officials have considered a range of actions, including using sanctions, a trade blacklist against Chinese companies that make these open-source A.I. models, or even banning U.S. cloud companies from doing business with them, said five of the people, all of whom spoke on condition of anonymity to discuss private matters. But after an outcry from Silicon Valley, they appear to have changed their minds and instead are focused on promoting American A.I. models to be more competitive, the people said.

The most recent debates were prompted by a surge in powerful new Chinese A.I. models, which are known as “open source,” where the public can see the underlying computer code, or “open weight,” where companies reveal the calculations used to generate answers to questions.
Click to expand...
This is an interesting shift in strategy. Rather than relying mainly on restrictions, focusing on innovation and keeping American AI models competitive could encourage faster progress while still addressing national security concerns. It will be interesting to see how this balance between open-source development, competition, and regulation evolves.
 
LukeTbk said:
they did fall behind but meta was a big player in open weight, I think they are still going.
Click to expand...
I haven't seen anything from them any time recently.

LukeTbk said:
Phi series from microsoft, gemma from google and Llama from meta, I think microsoft even use MIT license. Maybe they are all a bit old and too much behind chinese option by now
Click to expand...
Gemma is a bit of a gotcha. It is kinda like Qwen: Open... but only small models. Gemma has small models for single GPU systems. None of the Gemini models, the actual big frontier ones, are open. Qwen, Alibaba's, is similar: For the latest ones only the small ones are open weight, anything bigger than 36B is on their cloud only. While I'm not hating on that, it really isn't comparable to something like Nemotron which had everything from tiny up to mid-sized (like 330B) or Kimi, which is a bigass one probably on par with Opus.

I'm not saying there is nothing US based and open, but it is an area the Chinese are currently leading. US companies mostly seem to want to keep things bottled up. All of the open weights stuff, outside of nVidia, I've seen really do seem to be either advertising or trying to look good. GPT OSS 120B seems mostly to be "See look! We are open!" It is pretty much hot garbage and way worse than 36B or smaller models. Gemma is just the sales pitch for Gemini: "Oh is this good but not good enough? Well turns out we have more capable models. No, you can't HAVE those, but you can RENT them!" Anthropic releases nothing, despite all their talk. Phi is interesting but are tiny models made for phone type devices, which is cool and can be useful but for limited things, not really what people are talking about here.

If you are a home user there's plenty to play with from China and from the US. If you are a business that wants something that can do larger work and can afford a big server, China is the only one offering anything.
 
LukeTbk said:
That more controlling the hosting harness than the model, if you host it yourself or use a non china host of them, they are quite more open about those things.
Click to expand...
If we look at Deepseek's R1's full 600B parameter model from user reports I've seen it also gives censored responses to politically sensitive questions when run locally, although their online hosted version adds an additional layer. It's only the distilled (eg: Llama-based) versions of Deepseek's R1 that weren't censored (since they could transfer the reasoning but not the censorship) but since most users can't run the 600B non-distilled model locally due to hardware requirements most of the discussions only have familiarity with the distilled versions.

Unfamiliar with how other companies' full models behave when run locally, so there may be others that don't behave like this (or weren't foundational models to begin with).

It's worth noting for others though that there's a lot of good neural net papers by China-based researchers released publicly for others to build upon, for the past decade, so it's not a black and white topic.
 
Okatis said:
If we look at Deepseek's R1's full 600B parameter model from user reports I've seen it also gives censored responses to politically sensitive questions when run locally, although their online hosted version adds an additional layer. It's only the distilled (eg: Llama-based) versions of Deepseek's R1 that weren't censored (since they could transfer the reasoning but not the censorship) but since most users can't run the 600B non-distilled model locally due to hardware requirements most of the discussions only have familiarity with the distilled versions.

Unfamiliar with how other companies' full models behave when run locally, so there may be others that don't behave like this (or weren't foundational models to begin with).

It's worth noting for others though that there's a lot of good neural net papers by China-based researchers released publicly for others to build upon, for the past decade, so it's not a black and white topic.
Click to expand...
So are there ways that the advantages of Chinese-based open models are exaggerated or outright false? Not trolling, just trying to see different viewpoints.
 
philb2 said:
So are there ways that the advantages of Chinese-based open models are exaggerated or outright false? Not trolling, just trying to see different viewpoints.
Click to expand...
I mean the advantages of open weight models in general are:

- If they objectively perform well for tasks they can be useful.
- Being open weight means self-hostable which allows data to remain private (you don't have to trust license agreements of cloud-hosted models) and means data isn't being funneled to opaque closed model providers.
- Being open weight also means the models are fine-tunable and easily distillable. So if someone has a niche purpose where they need the model to perform better at a specific task they can modify the model weights. Or distilling to produce a smaller parameter version of a fine-tuned model, like how Deepseek's R1 was officially distilled using a Qwen (Chinese) model (and alternatively Llama, Meta's model), which meant that a smaller model could be run on cheaper hardware but retaining key logic of the larger full model.

In terms of what Deepseek's R1 made a splash in the news about earlier last year, part of it was because they'd managed to train their foundational model more cheaply using lower tier GPUs (since the class of GPU available to China is under US sanctions). Specifically they figured out a way to train without using the regular Nvidia SDK approach (used by everyone else) but with a lower level approach to work around memory constraints from the lower tier GPUs (IIRC). So it caused controversy since some wondered if US based foundational models could be trained more efficiently (though on the flip side Deepseek was accused of using OpenAI outputs en masse for part of its training, so it called into question how bootstrapped their model really was).

The concerns from such models I've seen are: if they get too popular the potential for whitewashed responses from topics that hurt CCP's image to influence international users (responses are obviously testable though, so this isn't a concern for all models) and the quasi-theorical possibility of models having malicious behavior that is masked by the model to not exhibit knowledge of this (eg. outputting deliberately vulnerable code, this is actually something that has been researched by eg. Anthropic and shown to be possible).
 
Last edited:
philb2 said:
So are there ways that the advantages of Chinese-based open models are exaggerated or outright false? Not trolling, just trying to see different viewpoints.
Click to expand...

To their own populace/things that make them look bad? Yes.

To the outside world that confirms their (outside world's) own beliefs/tells them things to go against their (outside world's) own government(s) even if just to form a relationship of """"trust""""" to give them (Chinese) more of your own data (simply for AI reinforcement/improvement/back-channel info and data and influence towards you/etc) because you 'trust' them so also use them more? Also yes.
 
Another data point is Xi Jinping's recent (officially translated) quote encouraging open source and alignment:

We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing.
...
AI should be a trusted tool for humanity... We should put in place laws and regulations, technological monitoring, early warning and emergency response systems in order to strengthen the line of security, prevent abuses and malicious use, and ensure that AI is always under human control.
Click to expand...

This reads like noble intentions, although the CCP is already known for using machine learning for egregious citizen profiling and surveillance (including government-led discrimination of Uighurs) so I imagine he means only abuse as it relates to tech vulnerabilities. Given this discrepancy I could see the encouragement of open weight models being considered by some perhaps part of some political leverage which could make some suspicious of ulterior motives (and yes, undoubtedly the US sees pushing US based models also as leverage but it's a matter of whose oversight one trusts, which is one of the reasons the EU is helping fund its own models).

Realistically if models were actually open source there wouldn't be a problem, since the training process and datasets would be auditable, but the majority of models are only released as open weight (with perhaps some research paper about general process/findings).
 
Last edited:
Okatis said:
Another data point is Xi Jinping's recent (officially translated) quote encouraging open source and alignment:



This reads like noble intentions, although the CCP is already known for using machine learning for egregious citizen profiling and surveillance (including government-led discrimination of Uighurs) so I imagine he means only abuse as it relates to tech vulnerabilities. Given this discrepancy I could see the encouragement of open weight models being considered by some perhaps part of some political leverage which could make some suspicious of ulterior motives (and yes, undoubtedly the US sees pushing US based models also as leverage but it's a matter of whose oversight one trusts, which is one of the reasons the EU is helping fund its own models).

Realistically if models were actually open source there wouldn't be a problem, since the training process and datasets would be auditable, but the majority of models are only released as open weight (with perhaps some research paper about general process/findings).
Click to expand...
They can't surveil you though. LLMs don't have any way to access the Internet on their own, they have to have a tool to use, and you can very easily monitor (and restrict if you want) tool calls on a system you control and it is the platform you run them on that provides the tools, not the model. They can't sneakily go and exfiltrate data, that's just not how they work.

So if China has some larger strategy with what they are doing, it isn't that. It could be just to bring attention to their stuff and get people on board. You'd probably have a lot of trouble convincing most places to use Chinese AI on Chinese servers where it is as secretive and locked down as what you see in the US. It also could be just trying to disrupt/tank the US AI industry. The companies are spending a shitload of money here, if they don't win and win big they are going to go under.
 
Sycraft said:
They can't surveil you though. LLMs don't have any way to access the Internet on their own, they have to have a tool to use, and you can very easily monitor (and restrict if you want) tool calls on a system you control
Click to expand...
What I was referring to, with the quote about preventing abuse with AI (along with similar points in the full speech), is there are documented things that don't align with the ostensible humanitarian goals, that they've used AI for (including Chinese companies specifically marketing facial recognition for Uighur detection to police) and in this context it would have to be interpreted selectively (eg: limited to tech vulnerabilities) as otherwise it would read as not being honest. If one sees it as dishonest then it might follow that their goals with open weight models aren't above board either, not that models are surveilling/exfiltrating data (see my earlier post for concerns I've seen raised).

In the short term, barring unwanted censorship or if any misalignment can be shown, it's a win for having useful models freely self-hostable and modifiable and providing a counterbalance to closed, cloud-based models and putting some pressure on such companies to continue to also release better open weight models.

It was interesting reading Sam Altman's email unearthed from the Musk lawsuit discovery where he explained that when they released their earlier open weight models it wasn't from benevolence but a deliberate ploy to try and disrupt funding to competitors and discourage others from also releasing models as open weight (clearly didn't work though). Going to show there can be other reasons beside the obvious that something is released open weight.
 
Last edited:
philb2 said:
So are there ways that the advantages of Chinese-based open models are exaggerated or outright false? Not trolling, just trying to see different viewpoints.
Click to expand...
In term of price of running them it can widely exagerated how much cheaper they are versus using a "small" ChatGpt model or gemini flash.

The cost of self hosting the biggest model will be heavy (say you need 8xH100 gpu and power them, that will not be that cheap), the cost of using them via API on a hosting service very similar to using openAI-Gemini, it is a rebate over Anthropic offer, they are tokens heavy and when model fail a task it is a lot of lost, so there can be a lot of tokens saved by the model being better.

Other than that they are performing very well in many aspect, that why they do not need to be much cheaper to run to be popular, open weight have advantage, the having more control versus OpenAI/Anthropic updating and having newer model that behave differently is big.

One trick many do is mixing the performance of the trillions parameter Chinese open model with the price of running distilled version at the same time, they can be near 3/6 month old frontier capability in some domain or cheap, not both.

Okatis said:
but a deliberate ploy to try and disrupt funding to competitors
Click to expand...
That often what open source effort are; trying to burn the commercial possiblity of a particular stack for others (non-core layer to themselve but core to competitors or would become competitor). Meta was trying hard but failed to destroy newcomers in the bigtech space, open weight Llama was trying for Anthropic-OpenAI to never get big. They did destroy some competition at api level but never the llm themselve.
 
tunatime said:
How can something out of China be open source? It's literally ccp controlled
Click to expand...

If you're running the model on a local machine, then the CCP doesn't get your data. In contrast, if you're working with OpenAI or Anthropic, you're sending all of your queries through their servers and you risk having them steal your ideas, as Figma has discovered the hard way.

It's significant that the biggest lobbyist against open source right now is Anthropic. A lot of companies have come to the realization that sending them all of their sensitive data might not be in their best interests, but running a local open source AI model is sufficient for what they want to do.
 
HockeyJon said:
as Figma has discovered the hard way.
Click to expand...
not sure if you mean people that sued Figma for doing so or the other way around, speculation that AI labs stole from Figma... but that quite speculative, that look exactly like what competition would have done before any LLM usage.

What figma was sending to the lab LLM was not really their core tech and the competitive tools those ai labs made work in really different ways I think. Maybe they went around term of service that promise to not look at API use... maybe not, there is no strong indication that they did.
 
LukeTbk said:
not sure if you mean people that sued Figma for doing so or the other way around, speculation that AI labs stole from Figma... but that quite speculative, that look exactly like what competition would have done before any LLM usage.

What figma was sending to the lab LLM was not really their core tech and the competitive tools those ai labs made work in really different ways I think. Maybe they went around term of service that promise to not look at API use... maybe not, there is no strong indication that they did.
Click to expand...

I was referring to the Claude Design thing, where Figma is partnered with Anthropic for AI services, and then some time later, Anthropic launches a product that directly competes with Figma. I'm not saying Anthropic necessarily stole code or anything like that, but what that incident did do is start a conversation in the industry as to whether or not companies really wanted to risk that happening by running everything through an AI intermediary that can now see everything you're doing and find the best ways to compete against your business. I guess the way I wrote it wasn't clear, bad choice of words on my part.
 
Sycraft said:
Ihere.

If you are a home user there's plenty to play with from China and from the US. If you are a business that wants something that can do larger work and can afford a big server, China is the only one offering anything.
Click to expand...
"and can't afford" ???
 
philb2 said:
Klaatu barada nikto
Click to expand...
tumblr_0b53b95a9fa2c8a9c93327939d421189_2e5fd5c0_400.gif
 
You must log in or register to reply here.
Back
Top