When it comes to assessing the risks of DeepSeek, are we asking the wrong questions? Governments, journalists, and coders need to know that it’s a much more sophisticated propaganda tool than we all thought.
Comments
Log in with your Bluesky account to leave a comment
Did you try the uncensored version? You can find on hugging face, bartowski uncensoredLM-DeepSeek-R1-Distill-Qwen-14B, i used a 5-bit quantization, as our office desktop here can't fit much in locally!
As mentioned in the article, the majority of Deepseek training materials are Chinese-censored materials. With that, the model itself will conform with Chinese propaganda with minimal engineering such as hard-coded rules about how to respond to 3T questions. It's the garbage-in garbage-out nature
Comments
Was a little confused when trying to recreate here locally.
Output here is also interesting....
Will check out the uncensored as well👍
(and 32b running here as well)