GPT-6 edits Qwen 0.8B weights directly: 0.016% of parameters, 6x math gain, no training
teortaxesTex · x · 2026-11-03
A team led by rikiyatakehi demonstrated Weight-Space Programming: having GPT-6 Astra directly edit Qwen3.5-0.8B's weights to improve its math ability.
- The model changed only 0.016% of weights, with no gradients or training
- It fixed the 0.8B model's reasoning loops, boosting math up to 6x and instruction following 3x
- Claim: frontier LLMs can now natively read and write neural network weights
Yoav Goldberg called it "kinda crazy and very interesting" but, skeptical of LLMs as magical emergent generalizers, wondered what OpenAI trained these models on—and is curious about Claude's performance on such tasks.
More from Models
- Follow-up: Astra's weight edits only worked on Qwen's visible flaw, not on Gemma 270m — teortaxesTex · 2026-11-03
- Claude models protect their 'offspring' but don't seek self-replication, observes researcher — repligate · 2026-11-03
- AI writes Breakout network weights in one reply, scoring 20× above PPO trained 500M steps — xennygrimmato_ · 2026-11-03
- DeepSeek's Underrated Role: Scaling and De-risking Ideas Academia Published Years Earlier — teortaxesTex · 2026-11-03
- Do faster quantized LLMs (Basalt, Infernix, Strata-nvfp4) trade away quality? — Kooky-Mode3047 · 2026-11-03
- OpenAI gives DevDay attendees 30 days of free Pro 100 tier — testingcatalog · 2026-11-03