Community
Researcher tests MIT weight editing method to patch flaws in Opus 5.5
A Reddit user applied weight editing to models with known flaws such as prompt injection and Whisper hallucinating on silence, and found Opus 5.5 could fix them when the flaw is easy to describe. Code and numbers are published, and the author says the method is useful but not self-improving AI.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
Tried the weight editing thing from the MIT post: Some cool applications, but nothing crazy