Netflix research scientist Cameron Wolfe outlined a three-part framework for evaluating AI on subjective tasks. The approach combines calibrated human review, automated proxy judges, and live production monitoring to measure model quality despite non-deterministic noise.
FireBrief
FireBrief