Crazy! GPT-6 just surgically altered the weights of an open model
⭐️ Try FeedHive, our AI social media automation tool: Use with MCP, Claude Code, OpenClaw, Grok Bot, and many more, or use our built-in AI. Try for free at https://www.feedhive.com. MIT just had GPT-6 Astra open up another AI model and fix it. No training, no new data. It traced Qwen's thinking loops to a tiny part of its weights (about 0.016% of them), subtracted it, and the model's math score jumped from 7.9% to 48.1%. In this video: what MIT actually did, why weight-space programming is such a big deal, and what it means if you want to run AI you own. Sources: MIT's post: https://rikiyat.github.io/blog/weight-space-programming/ Rikiya Takehi's announcement: https://x.com/rikiyatakehi/status/2117313742891327994 Simon Willison's test: https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/
