Hacker News Reader

>Some versions of DeepSeek’s models have bias built in

Is this empirically supported? IME it doesn't hesitate to talk about Tienanmen Square or whatever, so if there's "bias" it must be very well hidden.

View on HN

Yes if you search, lots of people have shared evidence of this. But it depends on which model you’re using, as some seemingly don’t have the bias built into their training.

And they also shared how trivial it is to bypass. It is one of the most uncensored released models that you can run privately from my understanding.

If you have a better model for the OP to run, please present it.

Edit: https://huggingface.co/perplexity-ai/r1-1776

Just released.

The 1.1B model which I used refuses to talk about Tienanmen Square.

Reply on news.ycombinator.com