Claim: frontier LLM edits Qwen3.5-0.8B weights directly, 6x math boost without training
beirmug · x · 2026-11-03
rikiyatakehi's team proposes "Weight-Space Programming": having a frontier LLM (called GPT-6 Astra) directly read and edit the weights of Qwen3.5-0.8B. They claim modifying only 0.016% of weights—no gradients, no training—fixed the small model's reasoning loops and improved math up to 6x and instruction following 3x. NaveenG Rao calls it a huge shift if it holds up: weights as the direct output of knowledge. Unverified by third parties.
Related event: GPT-6 'Weight-Space Programming' Claim Debunked by Users(4 posts)→
More from Models
- Leak: Anthropic's unrestrained Astra model far more capable than Opus — bindureddy · 2026-11-03
- OpenAI internal model hits ~26% pass rate at lowest compute, beating GPT-6 Astra's best — imjustnewatai · 2026-11-03
- Mistral Large 4 tech report rejected by arXiv, team releases it on alphaXiv — _AndrewZhao · 2026-11-03
- Fable 5.5 Rumors: New Model Appears to Be in Hands of Select Influencers — AIandDesign · 2026-11-03
- Qwen Code repo leaks unreleased qwen4-max, qwen4-coder-plus and qwen4-omni-120b model IDs — feelspeaceman · 2026-11-03
- Liquid AI Launches LFM3 Model Family; 2B Vision Model Runs Live Camera on iPhone Offline — JosephJacks_ · 2026-11-03