I think some of this is true for ML interviews as well, e.g., interviewers asking for specific library incantations and the precise mathematical expressions. Most good ML people I know think in terms of the assumptions an algorithm makes, known results that might apply (and therefore inform the practicality of a current approach), latency, time to train, the human workflow impacted, etc. When it comes to discussing the math., they tend to think in terms of proportionalities, e.g., "if I increase k, I would expect recall to increase..." and intended outcomes, e.g., "this step smoothes the loss landscape". If you have a broad knowledge of an area, its is impossible to remember details.