Weight-space programming, end to end: an LLM reads and edits a neural network's weights directly, no training. Sandboxed sessions, honest verdicts, shareable patches.
Stars, forks and merged pull requests each day or month, lined up with the repository's rank on Trendshift Trending
Built on "Sparks of Weight-Space Programming" by Takehi, Teoh, Zhang and Khattab: wsplab is MIT-licensed and independent of the authors.
@Ahmed_n1 · x.com
I also tested councils of LLM editors that see each other's work. They lift the weakest member, but never beat the strongest on a real task. Councils of cheap models, and fixed roles, did worse. All…
@Ahmed_n1 · x.com
What didn't work: edits fixed behaviour a network already had the parts for, but couldn't add a missing skill (math in small LMs). On Spurious ImageNet, a two-line scripted recipe matched or beat…
@Ahmed_n1 · x.com
MIT's new paper "weight-space programming" has been revolutionary. I built wsplab, an open-source framework for it. It runs the whole job end to end, and tells you honestly when an edit didn't work.…
@Ahmed_n1 · x.com
