[screcth]

The idea is that the LLM would provide "intuition" to guide the optimizer to find better optimizations, but a formal proof would be necessary to ensure that those optimizations are actually valid.