I can’t speak for Tom, but my experience with using LLMs in the domain of “primitives” and “odd API” (such as windows and powershell) is underwhelming. But I only have used them for search. E.g. they don’t understand powershell very well and getting things like process invocation right is something that requires meticulous research or extensive trial and error. LLMs often converge to common workarounds rather than best possible solution. And that’s probably not a place you want to be in when you maintain a core library.
They also tend to take away from that feeling of being insecure about decisions… e.g. when you have an intuition that you don’t understand all the implications. Getting superficial confidence from an LLM prompt in that case can have negative consequences.