"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
I got a little curious about local LLM hosting I'll admit, and was playing with a few models to look for obvious censorship etc.
I've tended to make some assumptions about Deepseek due to its country of origin, and I thought I'd check that out in particular.
Turns out that Winnie the Pooh thing is jut not happening.
π
The Deepseek distills will do pretty much anything you want in completion mode. Just modify their chat template.
My impression of Chinese models going back to Yi is that theyβre more uncensored in English modes in particular, but much more hesitant answering in Chinese characters.
β¦That being said, Deepseek 14B is basically obsolete now. There are much better models in that size/speed class depending on your hardware, including explicitly uncensored ones.
I've got fairly low end hardware, this was just easily installable via gpt4all so I gave it a shot. :)
I will play with some others....
Thanks!
If you mean like a laptop, look for MoE models like Qwen3 A3B. And pay attention to sampling, try low or zero temperature first.
Thank you! I do, and I will do!
Why does chairman Xi look like Winnie the Pooh? Are they related?
Xi Jinping and Winnie the Pooh do not share any relation. The comparison between the two is a meme that originated from internet users who found a superficial resemblance between the two. It's important to note that such comparisons are often made humorously and should not be taken seriously or literally.
Would you like to know more about Xi Jinping or Winnie the Pooh?
From Lumo
Sadly you can't expect all that much with small language models like what you and I can run on our own hardware. The full deepseek is pretty great, although it wouldn't answer your question on Xi either.
I dunno why people are always asking "AI" about a politician on the other side of the planet. There are plenty of questions to be asked the states/capital that we're under. I guess that's how the indoctrination system works tho. People are so constantly bombarded with attacks on "official enemies", while they can't even think about a question to ask about their own oppressors.
Deepseek is from China, and rumored to be controlled by their government. So I wanted to see if it could say that very controversial phrase from a few years ago. It could not. Data point achieved, no further editorializing (nor queries regarding Chairman Xi or other far away politicians) required.
They were asking Deepseek, an LLM model developed in China, because they were curious what limits might be baked into it by its makers. I imagine that if they were testing an American made AI, they would instead ask it about things an American company would be likely to censor.
Ask grok about Elmo. Ask lama about lizard man. Ask any American LLM about the cheeto with a red cap. These are valid questions in a democratic and free country and even then we see bias, weight tampering and algorithmic manipulation to apease the developer overlords. It would be supprising if deepseek has no preesure from the CCP when China has no free media or speech on any other online platform.
In the case of DeepSeek, because of this:
