Summary
Anima Anandkumar — Bren Professor of Mathematics and Computing at Caltech and former head of AI research at NVIDIA — joins the Latent Space podcast to discuss her work at the intersection of machine learning and physical science. The conversation covers neural operators, a class of models designed to approximate solutions to partial differential equations, and how they are reshaping fields from weather forecasting to materials science.
Anandkumar describes how her team’s neural operator approach to weather modeling, begun around 2021, surprised even domain experts: the models matched the accuracy of traditional physics-based forecasting systems while running tens of thousands of times faster — requiring only consumer-grade GPUs instead of supercomputers. The work helped democratize global weather modeling, and prompted DeepMind, Huawei, and others to develop their own AI weather systems in subsequent years. Her team was also the first to open-source a permissive weather model, allowing smaller agencies worldwide — including those in the Global South — to access high-fidelity forecasting that was previously out of reach.
The discussion also explores TorchLean, her more recent work combining large language models with the Lean formal verification language. The core insight is that while LLMs excel at generating scientific hypotheses, verifying that those results hold under rigorous mathematical or physical constraints remains the critical bottleneck for AI-for-science. Anandkumar’s dual foothold in academia and industry gives her a distinctive perspective on what it takes to build AI that is not only capable but also principled and guaranteed to work in the physical world.
📺 Source: Latent Space · Published August 26, 2026
🏷️ Format: Podcast
