Forgive the skepticism: Is this a harness that sits beside a language runtime, looks at traces/logs and uses an LLM to optimize code for given hardware?

It looks interesting but I am really unable to see what is so different from a couple days of prompts and digging into a software to optimize it using LLMs.