If the LLM is given the code for its training and is able to improve that, does that count? Because it seems like a safe bet that we're already there, the only problem is latency of training runs.