You can lobotomize a model to make it remember Tienanmen Square https://huggingface.co/blog/mlabonne/abliteration

Reply to this note

Please Login to reply.

Discussion

Thank you for sharing.

The problem though is that it won't be well integrated with other concepts. I see this with things that are soft censored in Claude – it will engage with an idea if you pointedly warm it up, but the implications and broader context aren't there until you've add the connection yourself.

We need models that dream. Maybe models that drop acid