Q: "How do we reconcile the encouragement from companies to use AI internally, specially to write code, with the requirement of no AI during interviews?"

I think there's no contradiction here. Analogy: if LLMs could do well at the SAT, the SAT would still be useful for accepting students to universities. The thing being tested is not important--it gives *signal* about something else that actually matters.

The point of the leetcode interview is to get signal about the candidate's general problem-solving skills, and how they approach a tough problem they have (ideally) not seen before. The fact that AI can solve leetcode questions doesn't change that leetcode interviews still give you that signal when taken by a human (that is not cheating).

Leetcode interviews are imperfect in many ways, most notoriously that memorizing a ton of questions DOES give you an edge, which sucks for candidates and warps the signal. But Big Tech companies do not currently have a better way.