GPT-6 Made Qwen 6x Smarter by changing its weights!

415 views · 6 likes · Nov 3, 2026

We asked Qwen3.5-0.8B one easy math question. It wrote "Wait" 1,121 times and never answered (our own run on a Mac, 32,768-token limit). Researchers at MIT CSAIL let OpenAI's GPT-6 Astra read the model's raw weights. Astra found a direction behind the loops and subtracted it fro…

0 comments

Sign in to comment.

No comments yet. Be the first.

More videos