Deepseek Tianenmen square controversy gets weirder
Deepseek Tianenmen square controversy gets weirder
So I was just reading this thread about deepseek refusing to answer questions about Tianenmen square.
It seems obvious from screenshots of people trying to jailbreak the webapp that there's some middleware that just drops the connection when the incident is mentioned. However I've already asked the self hosted model multiple controversial China questions and it's answered them all.
The poster of the thread was also running the model locally, the 14b model to be specific, so what's happening? I decide to check for myself and lo and behold, I get the same "I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses."
Is it just that specific model being censored? Is it because it's the qwen model it's distilled from that's censored? But isn't the 7b model also distilled from qwen?
So I check the 7b model again, and this time round that's also censored. I panic for a few seconds. Have the Chinese somehow broken into my local model to cover it up after I downloaded it.
I check the screenshot I have of it answering the first time I asked and ask the exact same question again, and not only does it work, it acknowledges the previous question.
So wtf is going on? It seems that "Tianenmen square" will clumsily shut down any kind of response, but Tiananmen square is completely fine to discuss.
So the local model actually is censored, but the filter is so shit, you might not even notice it.
It'll be interesting to see what happens with the next release. Will the censorship be less thorough, stay the same, or will china again piss away a massive amount of soft power and goodwill over something that everybody knows about anyway?
Worrying about whether or not an LLM has censorship issues is like worrying about the taste of poop.
Fine analogy.
benchmarks show that this LLM has the bouquet of the sunny side of the sewage farm
This is just me being childish, but it would be fun if we could incept the joke that LLM censorship = corn.
The rest of the owl draws itself in the imagination of the listener.
Hmm, the way I’ve chosen to interpret this is to propose the analogy of a person eating corn as a model of LLMs. You can eat a huge variety of foods and your poop looks more or less the same. Sometimes, you eat something like corn, and the result is you can spot kernels of things resembling real food (i.e. corn kernels) in the poop. However, if you were to inspect said kernels, you would quickly realise they were full of shit.
the product produced by the producer operating in a violently censorious state has complied with censorship in the production of the product? omg, stop the presses!
(nevermind the fact that if you spend half a second thinking through where said state exercises its power and control, the reasons for the things you observed should all but smack you in the face)
Local deepseek answers all my questions, but it is definitely biased in favour is the CCP even you ask about Chinese leadership.
Fortunately that doesn’t come up much for me
So I check the 7b model again, and this time round that’s also censored. I panic for a few seconds. Have the Chinese somehow broken into my local model to cover it up after I downloaded it.
what
It's a slightly facetious comment on how the same model had gone from definitely not censored to definitely censored. The tripwire for the filter was obviously already there.
I mean, his username is manicdave. Psychosis is a symptom of mania. It's a pretty wild thought.
Fuck. Just thinking of those bodies run over repeatedly and washed down the sewers...
I absolve you of your perceived need to think about it. You’re cured!
The local models are distilled versions of Qwen or llama or whatever else, not really deepseek's model. So you get refusals based on the base model primarily, plus whatever it learned from the distilling. If it's Qwen or another Chinese model then it's more likely to refuse but a llama model or something else could pick it up to a lesser extent.
You get the exact same cookie cutter response in the llama models, and the qwen models process the question and answer. The filter is deepseek's contribution.
China is an entire economy built on intellectual property thievery and slave labor. The whole thing, that's they're advantage. Theft and slavery. The Chinese CCP can fuck off into the sun. *edit: this post and article is about the ccp censoring it's human rights abuses, and all the downvotes confirm the ccp is here and active on the fediverse itself attempting to do the precise same thing. the irony shouldn't be lost.
China is an entire economy built on intellectual property thievery and slave labor.
so, just like the united states when it suited them, no? do read about united states in nineteenth century once. oh, and: you do realise that slavery in some circumstances is still legal in the u.s.?
The quants who built DeepSeek in their 20% time must really feel the sting of all that communism, I tell you hwat.
an entire economy built on intellectual property thievery
The historic attitude of the US to copyright is interesting (you can still see old English-language books that are labelled not for sale in the USA). In these enlightened days of course there’s a half-trillion dollar plan to shore up the LLM business, which is already a half-trillion dollar crater of debt and is still digging hard.
and slave labor
I always wondered how the manufacture of white goods in the US was competitive with the likes of Mexico, and it turns out that the secret ingredient is incredibly cheap prison labor, where the prisoners face significant negative consequences if they’re not prepared to work for pennies.
Prisoners can be firefighters for a few dollars a day and risk their lives, but are denied jobs when they get out. California, that noted bleeding heart lefty bastion, refused to abolish penal servitude (ie. slavery of prisoners).
The US health insurance industry means that huge swathes of the population may as well be indentured because cannot afford basic healthcare if they quit and changing employers risks rejection of coverage of pre-existing conditions.
I could go on. For quite a while.
all the downvotes confirm the ccp is here and active on the fediverse
By all accounts, the ccp do a pretty poor job of influence operations compared to russia. Personally I suspect that the fediverse is just too small compared to twitter and bluesky and reddit, so why would anyone bother here?
Truth is, both the US and the PRC are capitalist hellholes of differing degrees, and the current team in the White House works hard to reduce those differences. Remember, with the right wing, it is always projection and envy. They hate Iran because they want to be a fundamentalist dictatorship too. They hate china because they want uncontested power and a labor force without human rights.
You are not obliged to carry water for them.
there are lots of complaints one can level at ccp-run china without having to dip right into this strand of sinophobia, what the utter lazy fuck. bullshit like this that helps the world stay bad and intolerance. fuck off with it.
(sincerely: one of your downvotes from sunny ZA)
Oh hey Sam Altman is here, doing his best to bring down deepseek
From a technical perspective, it seems like a pretty clear case of model over-fitting to me. I've not dug too deep into this recent advancement, but this might be something more specific to how this recent model was trained.
Were you running this on local hardware?
All run on local hardware. Llama models do it too. This is 8b
Heh the entire post is about running the self hosted versions.
Right on. Yeah, so that would also point in the direction of extreme over-fitting during training. You might check out this notebook on abliteration if you want to learn more: https://colab.research.google.com/drive/1VYm3hOcvCpbGiqKZb141gJwjdmmCcVpR
My guess is you could confirm this is the case by looking at the high dimensional orthogonality of a given instruction that is refused, and then again some which are not refused. You could probably just try the two different spellings of Tianemen and Tianamen be able to key in on the censored response vector.