I noticed that most of the errors the AI is making in its code are because it can't anticipate the _behaviour_ of the program.  The syntax is fine.  Then I realised "oh yeah, that's the Turing halting problem".  Undecidable.  No AI or human will ever solve it in general.