'Weight-space programming': GPT-6 Astra edited a small model's weights and lifted its maths score from 7.9% to 48.1%

Researchers at MIT CSAIL (Rikiya Takehi, Omar Khattab and colleagues) had GPT-6 Astra act as a 'programmer' that edits another model's weights directly, with no gradient training. On Qwen3.5-0.8B it changed 131,072 weights (0.016%) in two layers using just 50 test runs, and the maths score rose from 7.9% to 48.1% (GSM8K from 13.8% to 67.1%). The authors' own analysis says the gain came mostly from stopping the model looping until it hit its length limit (93% of answers before, 33% after), not from new ability, and it barely helped Gemma 3 270M, which rarely loops. @kimmonismus called it a 'giant step toward recursive self-improvement'; that's his framing, not the authors'. Fact or hype: partly fact. The results and logs are public and match the claim, but it's an unreviewed blog post on one tiny model, and it fixes a flaw rather than making the model smarter.

Why it mattersNothing to buy. The practical lesson: many small-model failures are loops and runaway answers. If you run small or local models, cap answer length and watch for repetition before deciding a model is too weak for the job.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.