Uncensor any LLM with abliteration

There was a recent paper about a way to censor LLMs by just deleting the connections to any bad outputs, rather than training it to refuse them. I think this technique wouldn't work.

Obviously you could train any bad outputs back into them if you have the model weights.

interesting, there's going to be an arms race over censoring and uncensoring future powerful llms a lot like the getting a cracked version of photoshop back in the day