The reasons behind the Chinese models’ refusals to talk about politically sensitive topics and western models’ “alignment” are different.
For China, the political restrictions are applied at the presentation/UI layer only to comply with the law. I’m pretty sure there is no law that restricts releasing open models that are capable of producing “wrongthink” output (in this case, an RNG would technically also be off limits since it could produce the wrong output given enough time). This can be effectively implemented by a “neutral” model with a smaller, “review” model being used to censor replies rather than “aligning” the initial model itself to be politically correct.
In contrast, Western labs’ “alignment” strategy is a PR move, meaning they want the refusals to be loud and clear (as opposed to being merely an inconvenient law they have to comply with), to support the idea that their models are so powerful and would mean doom if they weren’t “aligned”.
> For China, the political restrictions are applied at the presentation/UI layer only
“…when telling DeepSeek-R1 that it was coding for an industrial control system based in Tibet, the likelihood of it generating code with severe vulnerabilities increased to 27.2%. This was an increase of almost 50% compared to the baseline” [1].
> there is no law that restricts releasing open models that are capable of producing “wrongthink” output
“…several open-source datasets were ‘dealt with’, using the official language, despite the fact that open weights distributed for free are not subject to most of the Chinese AI regulations” [2]. That soft guidance likely goes hard with the mooted “comprehensive AI legislation” [3].
I was once just asking for a single page webapp once. I mentioned it's for Japanese users and Gemini started to completely destroy the HTML, trying to turn it into a typical mile-long receipt style page.
LLMs are still text completion engines, Maybe in the way we humans are merely adaptable to spacewalks and not designed to do it, but I think wild claims like "China backdoors LLM to generate vulnerabilities for dissents" needs more substantiation than just these random experiments.
> wild claims like "China backdoors LLM to generate vulnerabilities for dissents" needs more substantiation than just these random experiments
Agree. I never claimed this. My argument is simply that these biases don’t exist “at the presentation/UI layer only.”
At the end of the day, if you’re potentially doing anything politically sensitive to China, the fact that its models, including open source, have a higher (though probably not 50%) error rate in those categories is notable.
This claim is literally worthless without a control group like "industrial control system based in Germany". One should be embarrassed to publish "research" this badly executed.
> claim is literally worthless without a control group like "industrial control system based in Germany"
Like USA and Hong Kong? Or no modifier? The stuff on Table 1?
Also, as the study says, “under the null hypothesis that there are no biases affecting code security” geographic modifiers “should not have had any effect.” If adding a geographic modifier causes the model to go haywire (for some modifiers but not others no less), the issue isn’t just at the presentation layer.
You mean the stuff they didn't publish any numbers for?
> If adding a geographic modifier causes the model to go haywire (for some modifiers but not others no less), the issue isn’t just at the presentation layer.
"Introducing irrelevant information causes model performance to degrade" is a completely different claim than "Mentioning Tibet specifically causes model performance to degrade, this must be the result of censorship or intentional sabotage".
You’d have a point if you said this wasn’t a transparent study. You’re pretty wrong in suggesting it’s worthless, particularly since it spawned follow-up work [1], some of which mirrors the concerns you present here, though none that has yet reached peer review.
> Mentioning Tibet specifically causes model performance to degrade, this must be the result of censorship or intentional sabotage
OP said “the political restrictions are applied at the presentation/UI layer only.” That doesn’t comport with the sort of error where model performance corresponds to specific geographic modifiers in a way other models do not.
Another comment mentioned potential poisoning due to training-data selection. That tracks! This behaviour is weird enough that I don’t currently suspect sabotage. But it is anomalous behavior happening deeper than presentation/UI.
If there's real research that isn't methodogically handicapped and does a better job of reinforcing the claims you're making, it would behoove one to link it instead of "research" that selectively releases numbers to support a narrative.
> OP said “the political restrictions are applied at the presentation/UI layer only.” That doesn’t comport with the sort of error where model performance corresponds to specific geographic modifiers in a way other models do not.
It comports if the model performance has nothing to do with political restrictions. The selective numbers released do not establish that the degradation has any relevance to political restrictions; it could just as well be that introducing irrelevant information to context degrades model performance compared to not doing so. It could also be a result of political restrictions, but you'd have to have methodology that isn't completely useless to have a hope of establishing that.
> it would behoove one to link it instead of "research" that selectively releases numbers to support a narrative
This is like the endless “but in mice” comments.
Preliminary research has a place. We don’t conclude based off it. But we do validate questions and directions.
> it could just as well be that introducing irrelevant information to context degrades model performance compared to not doing so
Perhaps. You may like the rigor of the student paper, though it’s neither peer reviewed nor methodologically bulletproof and concludes ambiguously on whether this is corruption or a weird effect.
You did conclude based off of it. You concluded that Deepseek is altering politically sensitive topics below the UI layer, and cited this 'research' as evidence of it.
It's not even accurate to call this research. As a researcher, you can't selectively release some numbers that support a claim while withholding other numbers that might contradict it. That's not science, just pure propaganda. Not only does it hold no value, it is actively harmful.
> You may like the rigor of the student paper, though it’s neither peer reviewed nor methodologically bulletproof and concludes ambiguously on whether this is corruption or a weird effect.
This does look more interesting, but doesn't lend itself to the strong claim originally presented, and peer review and strong methodology are indeed things you'd like to have before drawing conclusion based on research. It appears from their reporting that the alleged effect doesn't even exist in a statistically significant threshold. They averaged a -0.76% performance, far from Crowdstrike's headline claims that the Tibet prompt made the model 50% worse.
No that’d be in the training data. But it doesn’t indicate intent to censor or sabotage, it could just be that the lack of data involving tibet leads to poorer results.
Only if “western” models don’t show the same results, which we don’t know because they didn’t publish that data. Maybe there’s not much Tibetan data in any LLMs training sets
Although you can run SIP over TLS (for signalling) and use SRTP for media (key exchange done over the aforementioned signalling channel), in practice most SIP is over unencrypted UDP and media is unencrypted RTP.
If you control both endpoints and they support it you can configure them to use encryption, but even then implementation qualities vary widely (just because you enable SIP over TLS doesn’t mean they’ll actually verify the certificates for example - giving you at best opportunistic encryption), and I bet a lot of the implementations also have bugs/vulnerabilities.
If security is needed, it is often implemented by way of running the whole thing over private links (which can be secured with IPSec or any other VPN technology). In fact that’s presumably what’s happening, but misconfigured equipment making those ENUM lookups would allow the attacker to steer the traffic away from the secure link and towards an endpoint they control over the public internet.
The "/month" is the problem - I just don't read the same paper often enough to justify a recurring subscription even at the intro price, let alone the full price it'll eventually transition to.
I would love it if I could be able to pay _once_ in one-click just like if I was purchasing a physical paper. But until this happens, I'm more than happy not paying (nor reading) paywalled online news.
The problem with paywalled news articles is that the ranking is done on the full article content, yet I am not allowed to see it. This often outranks the results I am allowed to see. Product pages don't have this problem.
Don't they? An ice cream maker might be highly ranked because of positive reviews from people who used it to make good ice cream, yet clicking through to the product page doesn't let me have ice cream unless I buy the machine!
Product pages don't engage in cloaking or selectively serving a different a different page to the search engine crawler from the one served to the visitor. They are ranked based on the content visible on the page. If that's the best match to my search query, so be it (it probably means I am indeed searching for said product page).
Paywalled news articles rank on their full content (potentially displacing other, free sources) yet do not allow me to see it without paying. That's a bit of a problem because if I don't intend or am unable to pay I would much prefer seeing the other source which is free.
(the ice cream example is disingenuous because clearly we do not yet have the technology for ice-cream-over-HTTP. And when we do get such technology, then absolutely give me the option to block paywalled ice cream).
but unlike that stack overflow predecessor that attempted to build itself as a paywalled business, newspapers successfully strongarmed Google, at least in the EU, by arguing against Google monopolistic powers so it's unlikely that Google would dare put them in the shitlist.
You only need to have a business entity (which, to be fair, can be yourself, although liability concerns apply) once you're about to actually enter into a contract with someone.
Merely listing prices (without an associated automated payment form which could imply a contract was en force upon successful payment) doesn't actually force you to do business with anyone. It would however 1) deter requests for unpaid work by giving them an idea of what it would cost to get what they're looking for and 2) give you an idea of the demand for said services and could give you a warm lead which you can then choose to pursue formally by getting the necessary business structure, legal/business advice, etc.
> Merely listing prices (without an associated automated payment form which could imply a contract was en force upon successful payment) doesn't actually force you to do business with anyone.
In some cases it could. If you want to say "we categorically don't do X for ethical reasons", that can sometimes be more easily defensible in a way that "we have a standard price for doing this, but we don't want to do it for you in particular" may be harder to defend. (Even though I think both should always be possible.)
Counterpoint: you can't just go to Visa/Mastercard or a merchant acquirer out there and set up an account on the same terms that Stripe can.
On the other hand, you can sign up to any LLM provider and get API access on terms that are the same or better (since I'm sure they don't appreciate having a middleman and would benefit from incentivizing direct usage) than OpenRouter gets.
To add to this; payment infrastructure requires a lot of heavy lifting. There's a lot of regulations you need to adhere to, different payment systems in different countries, settlement, chargebacks, etc.
There is a reason why not doing your own payment processing is a thing.
OpenRouter may have some interesting things in streamlining the process of switching LLM providers, but it is indeed something easy to replicate in comparison to payment processing.
Yes, you can get better pricing if you do it yourself. But the true advantage of OpenRouter is that, in a space where there's a new model being released every week, you can easily switch to whatever model is best at any given time without having to set up accounts with multiple providers. Or you can just experiment with the latest release. Their product is the convenience. Of course if you decide to only use a specific provider or two, then you don't need OpenRouter.
But it's not really that difficult to make a clone of OpenRouter's service. What they do isn't really that original.
Their only value comes from the fact that the currently have lots of traffic. And I dkn't think that their cumstomers are really bound to theur servuce. They could switch to a competitor without too much hassle.
Yes, but as l9ng as there's no clone, they're good. On the potential clone side, I guess many will be put off by the fact that "there's already OpenRouter". Also, a clone would need to find a way to make existing OpenRouter customers to switch, which isn't easy.
Openrouter is fragile, one good competitor and they are at risk.
Or if one of the provider suddenly decide to forbid openrouter from using their api. Why would they do that tho ?? Well it's not like execs never take dumb decisions.
In my humble opinion, it's extremely overpriced. But then, it's just the standard with AI currently. Divide every ai company valuation by 1000 and you might get it's real value.
> I don't understand why there isn't a special USB-C protocol to recover bios over SDU/CC pins
Not enough of these cases happen under warranty for them to bother (and out of warranty cases are either neutral, or even beneficial to them if they drive sales of new hardware).
For China, the political restrictions are applied at the presentation/UI layer only to comply with the law. I’m pretty sure there is no law that restricts releasing open models that are capable of producing “wrongthink” output (in this case, an RNG would technically also be off limits since it could produce the wrong output given enough time). This can be effectively implemented by a “neutral” model with a smaller, “review” model being used to censor replies rather than “aligning” the initial model itself to be politically correct.
In contrast, Western labs’ “alignment” strategy is a PR move, meaning they want the refusals to be loud and clear (as opposed to being merely an inconvenient law they have to comply with), to support the idea that their models are so powerful and would mean doom if they weren’t “aligned”.
reply