Writing

Polymathic's Walrus is a 1.3B-parameter foundation model trained across 19 physics scenarios including MHD. Zero-shot on plasma turbulence: one-step prediction is excellent, but autoregressive rollout exhibits a structural basin instability in the perpendicular magnetic plane. FP64-confirmed structural, not numerical. The published headline doesn't tell you any of this.

A small workshop paper on neural MHD surrogates: 35% gain from in-domain pretraining, 14% penalty from wrong-domain, and a long-horizon instability that fine-tuning introduces from nowhere. ~15 hours over a weekend, <$20 cloud.