I'm seeing this as well. I don't quite understand how it's doing that in the context of LLMs to date being a "next token predictor". It is writing code, then adding more code in the middle.