ML driven compilers are different than LLM assembly generation, and former is more sound than the latter one. I believe we both have enough experience to understand that.

However, starting with LLMs directly generating machine code and moving to ML directed optimization passes, you're moving the whole stadium with a crawler-transporter, not only the goalposts.

Helping compiler in a well defined way, and working on IR or driving it repeatedly to optimize the output is one thing, generating machine code from a training set is other. Let's not mix them.