Nathan Barry revisits an experiment where GPT-6 Astra edited a smaller model’s weights to stop it repeating itself instead of answering. Astra found the layers and an internal direction associated with loops. Barry shows that choosing the size and location of the edit can then be posed as a linear program: a standard solver for finding values that satisfy constraints.
- On Barry’s held-out tests, the unedited model scored 7.0 on MATH-500, a set of math problems. Astra’s edit scored 50.9; the solver’s edit scored 51.7 using Astra’s direction, while changing about one-eighth as many weights.
- A direction found by a simple repetition detector, without Astra’s example texts, still scored 47.8 on the same test.
- Adding a fixed steering vector during each answer nearly matched the edit in this case. Too little brought the loops back; too much damaged the reasoning.
The solver replaces much of Astra’s trial-and-error search, not the entire investigation. Barry tested one small model and one looping defect. He still used the agent’s choice of layers, and his strongest result reused its direction.
The 29-comment thread on Hacker News raises a useful question about where the agent belongs in this workflow.
What the thread adds
- jsw97 — proposes using the solver with the agent: “with some guidance (eliminate this behavior, preserve that behavior, as expressed through linear constraints) the solver is much more efficient than Astra’s trial and error.”
- ultimatefan1 — pushes back on the original headlines: Astra “found a nice edit that fixes some problems,” but that is “a far cry from the headlines when the paper first came out.”
From a 29-comment thread, in HN’s own order.