GPT 6.2 Astra shows surprisingly strong zero-shot crystal generation, a year after the task looked hopeless

xennygrimmato_ · x · 2026-11-19

Inspired by recent posts on frontier models generating protein sequences, AdeeshKolluru re-tested zero-shot generation of MSUN crystal structures and found GPT 6.2 Astra appears trained to be good at it — a task that looked mostly hopeless when he tried it over a year ago. He notes autoregressive models work well once trained on stable structures in formats like cifs/json, and that Fable generates many more duplicates, hurting uniqueness. An empirical AI-for-Science capability observation.

Original post →

More from Models

Models channel →