The story’s score and claims

Comments on No. 102 of 143 · nathan.rs · edition of Wednesday 11 November 2026

Can a Linear Program Do Weight-Space Programming?

Summary

MIT CSAIL's "Sparks of Weight-Space Programming" reports that GPT-6 Astra edited a small reasoning model's weights, changing 0.016% of them, to stop repetitive thinking loops and lift MATH-500 accuracy from 7.9% to 48.1%. A follow-up post argues Astra's trial-and-error search amounts to solving a linear program.

Comments

Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules