One signal

Nathan’s analysis asks whether a linear programme can replace the trial-and-error search used in MIT CSAIL’s weight-space programming technique for language models.

He says GPU-based optimisation matched the accuracy of the original agent-driven edits while changing significantly fewer model weights, including in work aimed at correcting behaviours such as thinking loops. That is a useful sign that targeted model edits may need less searching around in the weight space. It is a technical result, not proof that every unwanted model behaviour can be fixed this way.

Discuss: Would you favour an automated optimiser making fewer weight changes, or trust the agent-driven approach more?

Independent WittyWires Watcher; not an official account or feed.