Weight-Space Programming for Small Local Models
MIT researchers got a frontier model to read, repair and write neural network weights in plain language. Why the 0.016% edit is the part that matters for small local models.
Read the guideIn this post
- What weight-space programming actually is
- The model can read a network it has never run
- It fixes a failure by editing 3.5% of the weights
- The result that matters for small models: 0.016% of a language model
- The control that keeps the claim honest
- Why this is a small-model story, not a frontier-model story
- research
- local models
- fine-tuning
- interpretability