These kinds of issues make me doubt that not releasing model details will help at all with ai safety. Lots of people have to waste their time guessing what's going on instead of stress testing models more efficiently. Safe systems become safer the more people can look into them