Comment by Jnr
3 hours ago
All good points, but I am not sure who I trust less. Chinese companies or random people on Hugging Face. :) It is still a black box. And I have a feeling I could not abliterate those models on my own hardware.
I am not saying I would not use any of them at all, but I would avoid doing certain tasks with them. And with the testing I have done with abliterated models in the past, it is not 100% successful and it can still hit the same walls as original model.
I agree. I think the future for privacy conscious people is going to be an adversarial open weights model, where competing models (ideally from opposing superpowers and labs) critique each other's work to find consensus. Or at least alert for suspicious behaviour.