Astra rewires Qwen weights directly: ~6x math gain, tripled instruction following

Token Gremlin @TokenGremlinOriginally posted 2026.11.03

Open original on X

THIS IS A HISTORIC MOMENT. GPT-6 Astra just reached inside another AI and rewired it by hand. No gradients, no training, just direct edits to 0.016% of Qwen3.5-0.8B's weights. The result: its reasoning loops are gone, its math jumped up to 6x and its instruction following tripled. AI is starting to fix other AIs from the inside.

This post quotes another post: “GPT-6 Astra edits model weights directly: Weight-Space Programming”