Building a global weather forecasting system is a litmus test for AI coding ability. We did it — in a restricted environment with no web search: 🧠 Designed a ViT-backbone neural network with 400M+ parameters, trained over 45,000 steps to model 69 weather variables evolving over time ⚡ Predicts a full week of global weather in under 1 minute 📊 Surpasses NVIDIA's classic FourCastNet on several metrics — with less training data A research agent turning a research question into a runnable, verifiable experiment. And this is just the beginning. Coming soon. 🚀