When LLMs edit neural weights directly, how do we version-control them? Announcing wpgit & Weight-Space Software Engineering (Paper + Code)
Links: Project Page: https://v-code-895.github.io/wsse/ Paper (Zenodo DOI 10.5281/zenodo.23956114): https://doi.org/10.5281/zenodo.23956114 Code, Checkpoints & Git Bundles: https://github.com/v-code-895/wsse TL;DR Earlier this week, MIT CSAIL published Sparks of Weight-Space Programming, demonstrating that frontier LLMs…
By Aisha Rahman
Links: Project Page: https://v-code-895.github.io/wsse/ Paper (Zenodo DOI 10.5281/zenodo.23956114): https://doi.org/10.5281/zenodo.23956114 Code, Checkpoints & Git Bundles: https://github.com/v-code-895/wsse TL;DR Earlier this week, MIT CSAIL published Sparks of Weight-Space Programming, demonstrating that frontier LLMs can inspect raw neural network parameters and write direct, gradient-free weight edits to patch models or steer behavior. Direct weight editing by an LLM is like hex-editing a binary in memory. It works as a one-shot demo, but what happens when an autonomous agent (or multiple agents and engineers) needs to steer different behaviors, add features, and compose dozens of edits over time? Without software engineering discipline, weight surgery causes silent regressions, breaks when branches merge, and collapses when the base model updates. We built Weight-Space Softwar…
It is part of a fast-moving week in AI, as researchers, builders, and enterprises keep adjusting to what the latest capabilities mean for them.
