NGI Proposal: Need to Develop Earth Models Based on Direct Observations Beyond Reanalysis Data Dependence

NGI (Natural General Intelligence) proposes the development of nature models trained on actual Earth observation data rather than human-generated outputs. Existing models such as NVIDIA Earth-2, Microsoft Aurora, and Google WeatherNext primarily learn from reanalysis data generated by upstream physical simulations, failing to preserve the balance between observation and inference and unable to reproduce the dynamics of small perturbations like the butterfly effect. While these models are faster and cheaper than physics-based baselines and generally perform better, they suffer from the problem of not learning the fundamental limits of atmospheric predictability. NGI argues that, following the 'Bitter Lesson', it is necessary to leverage more data and compute to learn a single Earth state estimate encompassing the atmosphere, ocean, and land. The goal is to establish an empirical feedback loop where AI capabilities lead to real-world Earth management.