Research & Reflections
Essays on astrophysics, artificial intelligence and scientific research.
10 November 2026Eight GPUs, Two Generations: Benchmarking Open LLMs for an Astronomy Institute
We measured open-weight language models on a mixed H200 / RTX PRO 6000 server and built AstroBench-AIP, a benchmark of 216 real astronomy computing tasks. The serving engine matters more than GPU count, and a 27B model on two GPUs serves 64 users at about 1 J per token.
08 November 2026Redefining Research: When AI Does the Science, Who Becomes the Scientist?
AI can accelerate discovery. But if students delegate the work through which expertise is learned, who will recognize when machines are wrong?