I find it interesting that OpenAI's safety worked best, where the others didn't work at all. I had different impressions before