3 ms·
Are they going to "lobotomize" ML models to make csam unthinkable for it, or are they going to teach it that csam is bad?
by akomtu 3y ago
Are they going to "lobotomize" ML models to make csam unthinkable for it, or are they going to teach it that csam is bad?
- GaggiX 3y agoThey are going to report the content to the CDNs and try to remove the entry URL from the LAION dataset and that's mostly it.
- tedivm 3y agoThe two paragraph abstract states that the goal is to identify CSAM in the training data itself and remove it. They ran a study and found "many hundreds of instances of known CSAM in the [LAION-5B] training set".