Ben Recht: 95% of uncertainty quantification is just Gaussian error bars
beenwrekt · x · 2026-10-09
Ben Recht's arg min blog post, a live blog of class 11 of his graduate seminar on forecasting, argues that uncertainty quantification (UQ) has become imaginatively impoverished.
Key points:
- 95% of UQ practice reduces to error bars — prediction intervals that compress a continuous forecast into a binary probabilistic event (e.g., "95% chance GDP grows between 0–7% in Q1 2027").
- The construction methods are equally uncreative: assume Gaussian data, estimate variance, set bounds at ±2σ. Simple models (linear dynamics with Gaussian shocks) admit exact variance computation; complex ones rely on Monte Carlo simulation or historical estimates, as in weather forecasting.
- Citing Charles Manski's PNAS piece on communicating uncertainty in policy analysis, Recht notes officials' failure to report uncertainty — but stresses that even when reported, the technocratic imagination for what quantifying uncertainty means is woefully narrow: most uncertainty can't be quantified this way.
More from AGI Musings
- Dan Faggella: abundance is participatory, not doled out equally — danfaggella · 2026-10-09
- Gradio founder: autonomous agents are breaking the incentives to share science openly — Gradio · 2026-10-09
- Investor revisits his 2019 AI predictions: mostly right, divergent futures ahead — sebkrier · 2026-10-09
- DeepMind essay: what 15M Gemini interactions reveal about AI-accelerated science — JMateosGarcia · 2026-10-09
- AI Swarmhood instead of Personhood: rethinking how we frame machine minds — theimposingshadow · 2026-10-09
- World models, not LLMs, are the bridge to physical intelligence — ingliguori · 2026-10-09