Calibration is all you need: trustworthy probabilities beat confident predictions
zsakib_ · x · 2026-09-20
The author argues that calibration matters more than raw accuracy: saying "I don't know" with well-calibrated probabilities beats being confidently wrong. If you predict something happens 50% of the time, it should happen exactly 50% of the time in the long run — whether your individual prediction is right or wrong is irrelevant. What counts is that your probabilities are trustworthy over time.
More from AGI Musings
- "If AI automates half of work but we all still work full-time, where did the savings go?" — VraserX · 2026-09-20
- Alignment is day-one; AI systems need management like corporations do, argues exec — mattturck · 2026-09-20
- Gary Marcus tells David Sacks: cut agent web access before the tech is controllable — GaryMarcus · 2026-09-20
- Beyond utilitarianism: the good may extend past sentience — danfaggella · 2026-09-20
- AI alignment blogger revisits his Evangelion-as-alignment essay: three years feel like a decade — arpitingle · 2026-09-20
- Dwarkesh interviews Noam Brown on multi-agent AI, math progress and alignment — soumitrashukla9 · 2026-09-20