Columbia recap highlights faster climate coding—and the need to check AI work

Columbia Climate School published an October 29 recap featuring researchers’ accounts of AI agents accelerating climate-science work, alongside warnings about checking their output. The accounts came from a LEAP-hosted gathering in late September, where speakers described faster coding and forecasting development but questioned how scientists would catch hidden failures and preserve expertise as agents took on technical tasks.

The event, “The Rise of Agentic AI: What It Means for Climate Innovation,” took place at the Columbia Engineering Innovation Hub. LEAP—Learning the Earth through Artificial Intelligence and Physics—is an NSF Science and Technology Center. Its sessions examined systems that can plan, write and run code, then revise their work with limited human oversight.

One immediate use is updating the software behind Earth-system models. Pierre Gentine, a Columbia professor and LEAP director, described models built on decades of scientific knowledge, often expressed in Fortran or C. Younger scientists are more familiar with Python, he said, creating a reason to translate the code without losing the science it contains.

Gentine described using agents to convert parts of that older software into Python-based code using machine-learning libraries such as JAX and PyTorch. He said converted code could closely match the original while using GPU acceleration. The potential benefit was both computational and practical: code that more researchers could read and work with.

Stephan Hoyer, a computational scientist at Periodic Labs, offered a smaller-scale example. He said Claude added a feature for simulating atmospheric motion to his team’s model in a few hours with minimal supervision. He had estimated that a PhD student intern might need about three months, with success uncertain. That comparison was Hoyer’s account and estimate, not a controlled productivity test.

Google Research senior research scientist Ignacio Lopez Gomez described an agentic framework that produced a California forecasting system in one to two months. He said it outperformed operational forecasts. He also argued that faster development could keep risk models grounded in newer information rather than past climates, helping people adapt to changing conditions.

AWS deep learning architect Sameer Mohammed warned that translating software can alter its behavior in less obvious ways. A conversion might turn a program that does not retain information between operations into one that does. Saving that information somewhere could introduce a security risk—a silent failure rather than an obvious coding error.

Hoyer drew a boundary around tasks with clear answers. Passing code tests or checking a mathematical result gives an agent a concrete measure of success, he said. Work with less definite measures is harder, including directing a coherent research agenda over several years. Faster implementation does not settle which scientific questions deserve attention.

Lopez Gomez proposed using agents’ coding ability to generate more tests and verification, not simply more research output. His case for trust rested on expanding the ways scientists can check results, rather than trusting the agents themselves.

The discussion also reached beyond research teams. Sarika de Bruyn, LEAP’s data and compute manager, said someone used the group’s agent to analyze neighborhood heat-wave susceptibility and successfully advocate for sprinkler caps on local hydrants. She described a shorter path from asking a question to receiving analysis and acting on it.

Gentine raised a different consequence of removing technical labor: scientific intuition grows through hands-on experience. He worried about how junior scientists would build that intuition while using agents at scale. The concern was not just whether an automated answer could be checked, but how future researchers would develop the judgment to check it.