In the end we discovered this had nothing to do with safeguards — OpenAI simply bribed them to switch. https://news.ycombinator.com/item?id=47197505