“In his microGPT project, Andrej Karpathy found that LLMs struggle significantly with the task of simplifying code, suggesting it is outside their current reinforcement learning optimization circuits.”