A preprint posted this week describes an open Qwen model from Alibaba changing its own weights without any training. Its edits raised accuracy on unseen problems by nearly seven points. Smaller models from the same family failed or broke themselves, and the work is not yet peer reviewed.