It's interesting how well the models handle things like comedy and satire. These aren't simple concepts, and they complicate the idea of AI safety training