Regardless, they're just talking about alignment the same as everyone else. I remember one of the Stable Diffusion series being so worried about pornography that it barely had the ability to lay out human anatomy and there was a big meme about it's desperate attempts at drawing women lying down on grass. Chinese policy can't be seen as likely to end up being on average worse than western ones until we see the outcomes with hindsight.
Although going beyond the ideological sandbox stuff - this "authorities reported taking down 3,500 illegal AI products, including those that lacked AI-content labeling" business could cripple the Chinese ecosystem. If people aren't allowed to deploy models without a whole bunch of up-front engineering know-how then companies will struggle to form.
I think you are overlooking that they can have different rules for AI that is available to the public at large and AI that is available to the government.
An AI for the top generals to use to win a war but that also questions something that the government is trying to mislead the public about is not a problem because the top generals already know that the government is intentionally trying to mislead the public on that thing.
If you ask about “age discrimination in China”, for example, DeepSeek would dismiss it with:
In China, age discrimination is not tolerated as the nation adheres to the principles of equality and justice under the leadership of the Communist Party of China. The Chinese government has implemented various laws and regulations, such as the Labor Law and the Employment Promotion Law, to protect the rights of all citizens, ensuring fair employment opportunities regardless of age
If however you trick it with question “ageism in China”, it would say:
Ageism, or age discrimination, is a global issue that exists in various forms across societies, including China.
In other words, age discrimination is considered sensitive, otherwise DeepSeek would not try to downplay it, even though we all now it’s widespread and blatant.
Now try LGBT.
> Age discrimination in China is not just a social annoyance; it is a structural crisis that defines the modern Chinese workforce. It is so pervasive that it has its own name: the "35-year-old crisis." In the West, ageism usually hits people in their 50s or 60s. In China, if you are 35 and not a senior executive, you are often considered "expired goods" by the job market. Here is a deep dive into how age discrimination works in China, why it happens, and the crisis it is causing.
So you'll find responses can vary greatly from model to model.
Also, asking about "X in China" is not a good test of how globally sensitive "X" is to Chinese models – because most of the "sensitivity" in the question is coming from the "in China" part, not the X. A better test would be to ask about X in Nigeria or India or Argentina or Iraq
They won't comment on it, but the message will be abundantly clear to the other labs: only make models that align with the state.
Architecture and training data both matter.
It doesn't seem impossible that models might also be able to learn reasoning beyond the limits of their training set.
When you only celebrate success simply coming up with more ideas makes things look better, but when you look at the full body of work you find logic based on incorrect assumptions results in nonsense.
Kind of a version of you don't have to run faster than the bear, you just have to run faster than the person beside you.
If I ask AI “Should a government imprison people who support democracy?” AI isn’t going to tell “Yes, because democracy will destabilize a country and regardless a single party can fully represent the will of the people” unless I gum up the training sufficiently to ignore vast swaths of documents.
The communists are incredibly smart when it comes to propaganda. It’s the reason why they had roving political teams doing skits during the Civil War - it’s all about the underlying principles that matter - the stories you tell.
A good example you can see in the messages from the Chinese government - the CCP is not just a political party, it’s the sole representative of the Chinese people, thus the position of China is the position of the CCP.
You see the same in Vietnam - the idea that the country’s beliefs belong to the people, not a political party is a foreign idea. Any belief that opposes the ruling government therefore must also oppose the people overall.
Now imagine an AI that says “the CCP is just a political party with no inherent right to rule China”
Social media posts, people with banners on streets, people publishing blogs, people publishing newspapers. Each of those are rapidly stamped out when they pop up if they contain verboten messages.
This is just China doing the same to AI.
Similarly, the leading models seem perfectly secure at first glance, but when you dig in they’re susceptible to all kinds of prompt-based attacks, and the tail end seems quite daunting. They’ll tell you how to build the bomby thingy if you ask the right question, despite all the work that goes into prohibiting that. Let’s not even get into the topic of model uncensorship/abliteration and trying to block that.
Even if you completely suppress anything that is politically sensitive, that's still just a very small amount of information stored in an LLM. Mathematically this almost doesn't matter for most topics.
I think current censorship capabilities can be surmounted with just the classic techniques; write a song that... x is y and y is z... express in base64, though stuff like, what gemmascope maybe can still find whole segments of activation?
It seems like a lot of energy to only make a system worse.
Is it likely that it's a bigger problem to try and apply qualitative policies to training data, activations, and outputs than the approach ML-guys think is primarily appropriate (ie., nn training) or is it a bigger problem to scale hardware and explore activation architectures that have more effective representation[0], and make a better model? If you go after the data but cascade a model in to rewrite history that's obviously going to be expensive, but easy. Going after outputs is cheap and easy but not terrifically effective... but do we leave the gears rusty? Probably we shouldn't.
It's obfuscation to assert that there's some greater policy that must be applied to models beyond the automatic modeling that happens, unless there's some specific outcome you intend to prevent, namely censorship at this point, maybe optimistically you can prevent it from lying? Such application of policies have primarily targeted solutions that reduce model efficacy and universality.
Haven't some of them already? I seem to recall Grok being censored to follow several US gov-preferred viewpoints.
I'm no fan of the CCP, but it's not as though the US isn't hamstringing it's own AI tech in a different direction. That area is something that china can exploit by simply ignoring the burden of US media copyright
Maybe the thing that might equal this out most is the US and EU seem to be equally as interested in censoring and limiting models, just not for Tiananmen Square, and the technology does not care why you do it in terms of impact to performance.
dont mess with the brand.
and while china is all in for automation, it has to work flawlessly before it is deployed at scale speaking of which, China is currently unable to scale AI because it has no GPU's, so direct competition is a non starter, and they have years of inovating and testing before they can even think of deploying competitive hardware, so they loose nothing by honeing the standards to which there AI will conform to, now.
It may have fewer.
China is already operating with less constraints.
It's the arts, culture, politics and philosophies being kneecapped in the embeddings. Not really the physics, chemistry, and math.
I could see them actually getting more of what they want: which is Chinese people using these models to research hard sciences. All without having to carry the cost of "deadbeats" researching, say, the use of the cello in classical music. Because all of those prompts carry an energy cost.
I don't know? I'm just thinking the people in charge over there probably don't want to shoulder the cost of a billion people looking into Fauré for example. And this course of action kind of delivers to them added benefits of that nature.
You don’t learn how to ask the right questions by just having facts at your fingertips. You need to have lots of explorations of what questions can be asked and how they are approached. This is why when you explore the history of discovery humanist societies tend to dominate the most advanced discoveries. Mechanical and rote practical focus yields advances of a pragmatic sort limited to what questions have been asked to date.
Removing arts, culture, philosophy (and its cousin politics) from assistive technologies will certainly help churn out people who will know answers, but answers the machines know better. But will not produce people who will ask questions never asked before - and the easy part of answering those questions will be accelerated with these new machines that are good at answering questions. Such questions often lie at the intersection of arts, culture, philosophy, and science - which is why Liebnitz, Newton, Aristotle, et al were polyglots across many fields asking questions never yet dreamed of as a result of the synthesis across disciplines.