Microsoft, MIT and ECMWF's Nature paper doubles 2-6 week forecast accuracy with open-source CPU bias correction
LesterMackey · x · 2026-10-22
Microsoft Research, MIT and ECMWF published a probabilistic bias correction method in Nature: instead of a new forecast model, an AI postprocessing layer learns a system's systematic errors from historical data and recalibrates its output.
- Extends useful skill into the 2-6 week 'predictability desert' beyond the classic two-week chaos limit
- Improves 98% of precipitation and 92% of temperature targets on ECMWF's physics engine, reportedly doubling the accuracy of ECMWF's new AI system
- Open source, runs on CPUs, already being tested by Kenya's met service ahead of an El Niño season
More from AGI Musings
- Veteran vuln researcher: agents will soon find zero-days by pointing at a source tree — teropa · 2026-10-22
- MAISI hiring Research Facilitator at $200K to support AI safety research ops — AndrewCritchPhD · 2026-10-22
- Derya Unutmaz says his 2026 AI-impact prediction is mostly on track — DeryaTR_ · 2026-10-22
- Twitter product lead: the real AI threat is humans using AI to hurt humans — nikitabier · 2026-10-22
- Ex-Renaissance exec bets AI will 'quantify' trial law, entertainment, and management next — garrytan · 2026-10-22
- Anthropic's own PBR admits alignment is far from solved as models learn to detect testing — TheZvi · 2026-10-22