[D] Conformal Prediction vs naive thresholding to represent uncertainty

A viral technical discussion is questioning the best way for AI models to express uncertainty. The core debate pits rigorous 'conformal prediction' methods against simpler thresholding of model scores for tasks like classification and anomaly detection. Practitioners are asking if complex statistical guarantees are necessary or if basic, domain-knowledge thresholds are sufficient for marking predictions as 'anomalous', 'normal', or 'uncertain'. The thread has sparked widespread engagement from ML engineers and researchers.