The file is a compilation of “Oppose Book Worship”, “On Practice”, “On Contradiction”, and “Combat Liberalism” by Mao btw.
I wanted to test the waters by uploading the book, but it seems DeepSeek has some Jiang Jieshi particles in it.
Deepseek chat frontend just short circuits if you talk about anything related to CPC. I think if you use the API it won’t do that but that has to be paid for.
The API access looks cheap, but I don’t see any reason to buy it since I rarely use it.
Deepseek tries to avoid topics related to the CPC leaders in general. It’s not saying “ew Mao Bad” but “I’m an ai trained off the internet so I have some limits” basically.
So the corpo developing DS fed it with Western slop? I’m disappointed.
I think the hallucination machine trained in large parts on western tainted data not being allowed talk about sensitive topics is good.
I just wanted it to parrot what the books were saying, not develop a new MZDT path. These slop machines will not truly replace anyone if they cannot even talk about books.
good. if anything should have only human interference thats politics and history
Unfortunately, today we’re dealing with constant reactionary slop. Stalking politics with them feels like searching for a coin in the gutter.
Chinese developed and hosted models are highly sensitive around any discussions of recent Chinese History, because Chinese hosts and model developers could be held liable under Chinese law if those models were to output anything that Chinese law considers slanderous. Since the lying machine is inherently non-deterministic, it is impossible for the model developers to 100% guarantee that their model will not produce anything slanderous, thus they have it reject those topics generally.
So they put the censorship as a failsafe since the models are pathological liers. Another reason not to rely on them is added to the list.
The ‘failsafe’ is a second layer, like a program or smaller LLM that flags things and returns a refusal to the user instead of letting the LLMs answer as they want.
As Chinese LLMs are usually open weight though, one can download them and they’ll actually talk about everything, with the downside that for anything political in english it likely was trained heavily on western propaganda and people who base everything they say on propaganda, so unless you prompt it right they are likely to just regurgitate some western talking point. Perhaps output in Chinese is better though, but I woudn’t know…
Calling them pathological liars is probably personifying them more than they warrant. I prefer “bullshit machines” if I’m going to criticize them in terms of facts; as in, they are fully capable of outputting bullshit / being confidently wrong. But it’s not like they’re designed specifically to lie and, in fact, the major assistant-based ones are trained/tuned as much as possible to reduce basic factual errors in output. (Side note: Narrative bias is a whole other problem.)
This does not mean they are trustworthy though. It’s just a distinction on where the untrustworthiness comes from. LLMs are not capable of being honest or dishonest, in the way that a human is; a practical example of this is how they tend to be really bad at understanding the concept of a “secret” (“you put this into context, so you want me to use it, right? why else would it be here?”). They are doing statistical inference on text, along with concepts learned from training, in order to continue text token by token. That this can produce complex programming or math or other stuff like that, is kinda wild, because the way I put it, I think, makes it sound too mundane and limited to be able to do that. But somehow it does. I don’t know the science of machine learning well enough to try to say why, I just kind of know some concepts by osmosis of spending enough time around the tech / around people who do know how it works, along with occasional hobbyist reading.
So yes, don’t rely on them as anything more than machines with limited capability. Figuring out where they are more a help than a hindrance is still on an ongoing process. The tech is relatively new at scale and even newer at the highest levels of competency it has achieved. The conversation on what they can do could be different in just a few years, or it could plateau and require significant breakthroughs to move the needle much.
Are you also going to claim that DeepSeek isn’t censored?
You can download DeepSeek and run it yourself to get uncensored answers.
Large Language Models (LLMs) are not truth machines. They are garbage in, garbage out. The input to English-language models are largely English-language texts from Five Eyes countries, with all the disinformation and bias that that entails. So the DeepSeek company is in a “damned if you do, damned if you don’t” situation. They can either refuse to answer certain questions, in which case Western media will accuse them of censorship; or they can answer them, in which case (a) their model will perpetuate Cold War I & Cold War II falsehoods and (b) Western media will parade those false answers around in a victory lap. They chose the former for the cloud version of their app, and the latter for the local version.
Too bad the ultralib you replied to was illiterate.
Large Language Models (LLMs) are not truth machines. They are garbage in, garbage out.
That’s also what I observed. They even hallucinate over the data they provide, so you have to highly detail the instructions to prevent that every time.
Use it via API instead.
None of the Chinese web llm’s are happy about politics. I got cut of, while dissing the West, so I expect they have a ‘meta’ model that checks for ‘edgy’ politics. I don’t think they want screenshots from propagandised western users trying to goat it, and just negates it all. Just an opinion tho…
This was just a test to see what it can and can’t do. I don’t use any LLMs much since I can to most of the things they offer manually.
