From a 2014 MSR bet to Nature: how deep learning conquered weather forecasting
LesterMackey · x · 2026-10-19
Aditya Grover recalls that in 2014, computer vision was the only AI domain showing promise with deep neural networks. During his Microsoft Research internship with Anil Kapoor and Eric Horvitz, they bet that weather data resembled visual data — pixels swapped for physical variables on a planetary-scale grid. Deep learning is now the leading approach for weather forecasting across broad scales.
Horvitz notes that early KDD paper was the distant ancestor of intensive follow-on work at Microsoft and elsewhere. Their newest Nature-published work uses probabilistic bias correction to learn from errors and improve the challenging 2-6 week subseasonal range. With world models taking off, the authors see a path toward general-purpose AI systems that generalize across physical scales.
More from Research
- Paras Chopra reviews Sensing Qualia: a fresh "external" take on the hard problem — paraschopra · 2026-10-19
- Anthropic gives patient $30,000 in Claude credits to research her own tumor — g3off_record · 2026-10-19
- PaperLeet open-sources a LeetCode for AI researchers: ~100 key papers, graded by tests — udmrzn · 2026-10-19
- Top probability journals shut submissions for a month citing surge and AI impact — michaelchchoi · 2026-10-19
- Bare IPython Harness Matches Production Agents on Qwen3.5-9B; Output Normalization Kills Run-to-Run Variance — HauntingMoment · 2026-10-19
- Bryan Kelly and Dacheng Xiu release 159-page Financial Machine Learning textbook free — mdancho84 · 2026-10-19