Again, the paper that said that transformers only copypasted input samples was highly misleading.
It seems clear to me that Codex has true understanding.
(Yes, I know that people have gotten secrets to appear in the output by prompting it in clever ways. That this happens doesn't prove that Codex doesn't understand what it's doing, it just shows that Codex doesn't understand everything.)