GPT-6 edits Qwen3.5-0.8B weights directly: 0.016% changes boost math 6x, no training
teortaxesTex · x · 2026-11-03
- The quoted core: a team had GPT-6 Astra directly edit Qwen3.5-0.8B weights to improve math. It changed only 0.016% of weights — no gradients or training — fixing the 0.8B model's reasoning loops, boosting math up to 6x and instruction following 3x. They call this "Weight-Space Programming": frontier LLMs can natively read/write neural network weights.
- The attached screenshot is a hiring post from Fleet researcher Michael Sklar (ex-Anthropic interpretability resident): Fleet is launching a "mechanize for mechinterp" project, paying competitive rates for researchers to do model-aided interpretability research and turn agent failures into training tasks. Fleet currently sells RL gyms/tasks to every frontier AI lab. The reposter concludes labs are clearly already doing RL on mechanistic interpretability.
Related event: MIT's Weight-Space Programming: GPT-6 Rewrites Weights to 6x Math Skills(31 posts)→
More from Models
- Kimi K3.1 Pro open weights expected to drop this week — thursdai_pod · 2026-11-03
- Model swap cut costs 59% and latency 75% — and he still blocked the release — Fun_Employment6042 · 2026-11-03
- Perplexity distills 9B embedding teacher into 0.6B model sharing the same embedding space — antoine_chaffin · 2026-11-03
- Gemini keeps wiping users' chat histories, sparking reliability complaints — matrixmayhem · 2026-11-03
- Astra's Qwen weight edit was a published interpretability technique, not RSI — markjeffrey · 2026-11-03
- Yoav Goldberg: Astra's weight-editing feats are mostly known interpretability tricks — yoavgo · 2026-11-03