Large Language Models (LLMs) are rapidly growing in size and capability and are transforming modern knowledge work. AI agents, which are essentially multi-step pipelines of LLM inference, have been demonstrated to successfully carry out complex knowledge work tasks in minutes, and which would otherwise require hours from a human. This rapid development thus warrants asking whether LLM-based AI has a place in Earth science research or forecasting and what that place could be. In this talk, I will review the state of the art of agentic AI and demonstrate its usefulness in a downstream atmosphere-ocean science application: operational tropical cyclone forecasting. A requirement of many operational needs is deterministic output of tools, and we will explore ways in which this can be accomplished for LLMs. We will then look at an unstructured text archive of 20-years worth of operational tropical cyclone discussions and forecasts, and see how deterministic AI agents can aid and accelerate human forecaster tasks such as search, interpretation, and prediction.
Copyright. All Rights Reserved.