/vt/ - Virtual Youtubers

>>37003157
it was still an issue pre-lob, it was just way less noticeable because sexual shit wasn't automatically triggering incessantly positive loveloops. it's a problem inherent to the model because of the faggots who trained it, see picrel ("PT" stands for pre-training)

>>36729233
And finally at appendix the end they mention how they were able to get people to do label their data for them safety against sensitive and adversarial content. 116 participants. That's enough to affect the response from pic related.
Dammit. I knew it was easy to create a filter, but I didn't know it could be this cheap. All you have to do is hire some cheap slaves to annotate responses like these...