Most interview loops were designed for a kind of engineering that is going away. They test whether you can produce an algorithm on a whiteboard, or recall how a system you will never touch behaves under load. These were decent proxies once. They are bad proxies now, because the skill that matters most is the one they cannot see.
If you want to find someone who is good with models, watch them work with one. Give them a real problem, a messy one, and let them use whatever they would use on the job. Do not ban the tools. The tools are the point. You are not testing whether they can do it the hard way. You are testing whether they have taste.
Watch for a few things. Do they trust the model too much, or not enough. Do they notice when it is confidently wrong. Do they build a way to check the output, or do they eyeball it and hope. When it goes sideways, do they patch the one prompt, or do they fix the system. Speed is the easiest thing to see and the least interesting signal. Judgment is the thing.
Ask them to show you something they built this way. Not the demo. The part that broke, and what they did about it. People who are native to this have a story about a failure case they still care about. People who are pretending do not.
The reason the usual loop fails is that it filters on what is easy to measure and slow to change. Years of experience. The name of a school. A language on a resume. None of those tell you whether someone can do the new thing, because the new thing is younger than all of them. If you sort by the old signals you will pass on exactly the people you are looking for, and never know you did.
This is most of what we do. We watch people work, and we are patient about it. It is slower than reading resumes. It is the only thing that works.