On a number of predictably "sensitive" topics, the model first generates and then abruptly deletes already generated text to say "beyond scope" or something.
Interesting: when you ask for illegal drug recipes, deepseek behaves more like other models, it just writes out that it won't help. It's a different mechanism.
has anyone "abliterated" the other good Chinese models yet? This one just came out, so I'd imagine not, but the smaller DSv2.5 ones have been around a while...
Comments
I read some people ran 2.5 locally and it wasn’t as censored as the Deepseek site.
I’m wondering if these GPU sanctions played a role in them deciding to go that direction.