GPT-6 Astra didn't teach Qwen3.5-0.8B math. It read its weights, found the rep… · @LJL06019 · TwKit

@LJL06019

GPT-6 Astra didn't teach Qwen3.5-0.8B math. It read its weights, found the repetition loop, and deleted 0.016% of them — MATH-500 went 7.9% to 48.1%. The edit didn't add knowledge. It removed an obstacle hiding ability the model already had. Weights were supposed to be unreadable. Now they're editable. #AI

曝光 48 · 评论 4 · 点赞 1 · 书签 0