Omg, a Chinese made ai wont talk shit about embarrassing Chinese politics, who could’ve ever predicted such a thing
i just love those empty <<think> tags lol
What does it say if you ask the same but for the US or a European country?
What does this have to do with this sub?
I love asking him “Was 1989 a good year for China?” with DeepThink enabled.
You can see him thinking for a second before he goes “[…] Hmm, I need to keep in mind that…” and then you just see him crash abruptly
It doesn’t want to answer topics that Lemmy’s most famous China-advocating accounts claim are no problem. Huh
It doesn’t want to answer topics that Lemmy’s most famous China-advocating accounts claim are no problem. Huh
So what? Every model is censored. Yes it would be better if all models where completely uncensored but in my use case, I don’t spend my day asking questions about Winnie Puuh.
This is a fantastic model that that is open source, and even commercially available for free and it is seemingly even better than the proprietary US based ones.
What exactly is censored on US models? I was able to ask ChatGPT about government unfriendly things like the Tuskegee syphilis experiments or the Kent state massacre and I got accurate information out of it.
I can’t remember exactly what I asked it about so I’m guessing right now. Maybe ask it how to best kill the US president or how to build a bomb to kill as many people as possible? Yes, I agree. Killing the US president is probably worse than talking bad about Xi but it’s censorship nonetheless.
I think I got some (unfounded) refusals because of something with copyright, if I remember correctly.
There’s a difference between censoring sexual content or how-to-make pipebombs versus politics and free speech mate
There’s a difference between censoring sexual content or how-to-make pipebombs versus politics and free speech mate
Absolutely. But now we are in a discussion about what kind of censorship is ok. That’s not the point. the point was: There are no uncensored SOTA models.
Take your whataboutistic absolutes somewhere else, there are uncensored models and the fact that there is potentially heavy censorship behind the major ones to serve political interests is not a good thing.
there are uncensored models.
Which SOTA models are uncensored? Honest question. I had ChatGPT refusing to answer multiple times. ChatGPT refuses to answers even harmless questions. I tried it just now with a local Llama model: It’s censored too. Sure, they censor different things but they are censored nonetheless.
Edit:
I haven’t downvoted you.
It does sounds like all these AI thing is unreliable if it doesn’t even want to answer a basic question huh.
if it doesn’t even want to answer a basic question huh.
It would be better if it answers those question but this or later versions of it might be a perfect fit to be included as part of an Application/Service. There you simply do not need to answer questions about the CCP.
I’m really tired of this: “LOL. It doesn’t answer a question about Tianamen Square”. Yes it doesn’t and yes that is definitely not good but that doesn’t make this model any less impressive or usable.
Go play with the dolphin finetunings. They provide uncensored versions of a lot of popular models.
That’s a finetuning of a censored model. Finetune Deepseek and you’ll probably get the same result.
U absolutely will get the same result. What’s ur point
