Go big or go home, baby! Scaling laws for AI weather models
Apple Podcasts
|
Spotify
008 | 2026-06-09
It's been a while! But we're making the wait worth it with a deep dive into scaling laws for AI weather models. As Alden put it, "this feels like a huge growing-up moment for the field."
We unpack how scaling laws — first crystallized in the LLM world — are evolving for weather prediction: why the laws from LLMs don't transfer directly, what scaling actually means (more data, more compute, bigger models), which dimensions to scale for the best performance gains, narrow versus wide architectures, and what responsible scaling looks like amid real-world constraints. Today's word of the day is wide.
Paper
- Scaling Laws of Global Weather Models, Yu et al
Chapters
- 00:00:36 - LLM context
- 00:06:36 - Scaling discussion & big model talk
- 00:19:33 - Paper time (finally)
- 00:48:58 - Closing thoughts
- 00:57:34 - Books!
Recommended reading
- What Are People For? by Wendell Berry — specifically the essay "Why I Am Not Going To Buy A Computer"
- The Broken Earth trilogy by N.K. Jemisin
- Anything by Irene Haas: The Maggie B, Summertime Song, The Little Moon Theater, etc.