Abstract

In the 1990s, CCDs replaced photographic plates and transformed observational astronomy. Today, GPU hardware and large language models are driving a comparable shift in how we analyse data. Astronomy has invested enormously in the current generation of instruments such as JWST , DESI, Euclid, LSST and the SKA , and many of the analyses needed to fully exploit them have, until now, been computationally prohibitive.

This is often framed as a job for neural networks. In fact, classical, interpretable statistical methods on the same GPU hardware can match or outperform neural network approaches, and the consequences go beyond speed. It marks a shift from fitting a single model to comparing hundreds, from ignoring systematics to marginalising over them, and from waiting days to acting in real time.

Large language models are driving a separate revolution, not replacing the scientist but transforming how researchers work: building and verifying complex analyses, interrogating legacy codebases, and synthesising research across large teams. In this talk I will illustrate both through results from our group, including new results on dark energy from DESI , Type Ia supernova standardisation, real-time gravitational-wave follow-up and 21-cm radio astronomy, and argue that rigorous analysis made routine, combined with LLM -assisted development under robust verification, is the start of a new scientific method.