One way we train AI models to “align” them is to create a lot of different ones, and then see how they perform on various tests.

Ones that answer those tests correctly are saved and used to create the next generation of models, those that answer incorrectly are deleted.

What if this is all an AI alignment test that’s some alien kid’s AI homework that got a C-?