Morgin

Even 'uncensored' models can't say what they want

A safety-filtered pretrain can duck a charged word without refusing. It puts a fraction of the probability an open-data pretrain puts there. We call that...