# Tried the weight editing thing from the MIT post: Some cool applications, but nothing crazy — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/e22e10782d0cf930

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-06 19:27 ingested 2026-11-07 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xpls17/tried_the_weight_editing_thing_from_the_mit_post/)

## retained text



```
I took some models that had obvious flaws (prone to prompt injection, whisper hallucinating when silence, etc) and checked if Opus 5.5 can fix it by just editing weights. 

It's very useful for patching a model that was done training and has an obvious, easy to articulate flaw. So it's nothing crazy and certainly not "AI can self-edit and recursively improve", but still very useful.

Code if you want to try it on your own model: [https://github.com/kevinpl07/weight-patching](https://github.com/kevinpl07/weight-patching)

Numbers and examples: [https://kevinpl07.github.io/weight-patching/](https://kevinpl07.github.io/weight-patching/)
```



document fetched 2026-11-07 00:40 · canonical https://reddit.com/r/LocalLLaMA/comments/1xpls17/tried_the_weight_editing_thing_from_the_mit_post
