felipepenha/pi-weight-space-programming is an extension that lets language models read and edit neural network weights without training inside the Pi coding agent. The repository's latest release is v0.0.4, but the provided materials do not list what changed in this version. The README describes checkpoint analysis, the ability to compose algorithms into weight matrices, and gradient-free editing tools.
This approach focuses on inspecting and modifying the weights themselves instead of retraining the neural network. In pure NumPy it analyzes the singular value distribution and effective rank of safetensors files, along with attention circuits, with no forward computation. It compiles symbolically expressed algorithms directly into weight matrices and also includes tools for adjusting specific weights via directional subspace projection. It links paper experiment setups with run records, and supports reproducing experiments with a single flag as well as downloading Hugging Face checkpoints.
A separate Python sandbox is CPU-only, with networking turned off and backpropagation and optimizer use blocked. It therefore provides a workflow for working with weights, but its scope stays within what the sandbox can run. The provided materials contain no performance figures for judging editing quality and no comparison with other tools. This repository can be seen as a tool for experimenting with that approach, but no evidence backing the effect of the editing results is presented.