GPT-6 edits Qwen3.5 weights directly: 0.016% changes, 6x math gains, no training
lateinteraction · x · 2026-11-05
- Weight-Space Programming: GPT-6 Astra directly edited Qwen3.5-0.8B weights to improve math—no gradients, no training.
- Changing only 0.016% of weights fixed the small model's reasoning loops, boosting math up to 6x and instruction following 3x.
- Takeaway: frontier LLMs can now natively read and write neural network weights.
- Follow-up discussion: can coding agents distill a trained neural net into readable code? Not yet, but likely coming.
Related event: MIT's Weight-Space Programming Lets GPT-6 Directly Edit Qwen Weights(5 posts)→
More from coding & agent
- POKI open-sourced: a no-jailbreak iOS agent that can fully control your iPhone — alexcovo_eth · 2026-11-05
- Foresight-SWE benchmark: coding agents solve 60.7% of issues but only 14.8% without introducing bugs — OfirPress · 2026-11-05
- User: Every Client Deck Automatically Improves HQForWork's Skill File — jacob_posel · 2026-11-05
- 10 Grok Bots to Build Your First $1K: Idea Stress Tests, Competitor Stalking & More — ChrisUniverse · 2026-11-05
- Developer says Claude and Codex nearly bankrupted him by making procrastination comfortable — oran_ge · 2026-11-05
- "We're gardening now": devs rebrand their role as agent babysitters — heyneighbor · 2026-11-05