[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop
Introduced the technique to a practitioner audience, with a concrete Qwen reasoning-loop edit.
5 items · 5 sources · 5 days
Operational story trace
Follow in this browser to see new updates on your Live feed.
Latest change
A Nov 18 arXiv paper tests whether a weight edit can change a model's ethical trade-off, raising the benefit needed to override a formal pledge in a concept-supervised text model.
On Nov 3, Latent Space covered an MIT result showing frontier LLMs can read and write small neural network weights directly, with no training run. Over the next two weeks an open-source instruction set, a game-agent compilation paper, and an independent replication followed.
Arc
Introduced the technique to a practitioner audience, with a concrete Qwen reasoning-loop edit.
First open-source attempt to formalize weight writing as an instruction set.
Tested whether compiled agents generalize to unseen levels, renderings and start states.
Independent replication of the original MIT write-and-read results, with extensions.
Moved the technique from capability edits to value edits, testing a pledge-versus-benefit trade-off.
Introduced the technique to a practitioner audience, with a concrete Qwen reasoning-loop edit.
First open-source attempt to formalize weight writing as an instruction set.
Tested whether compiled agents generalize to unseen levels, renderings and start states.
Independent replication of the original MIT write-and-read results, with extensions.
Moved the technique from capability edits to value edits, testing a pledge-versus-benefit trade-off.
What to watch — open questions
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.