MIT: GPT-6 Astra rewrites other models' weights directly, lifting Qwen math accuracy from 7.9% to 48.1%
The Decoder · rss · 2026-11-04
MIT researchers show that OpenAI's GPT-6 Astra can read and deliberately rewrite the weights of other AI models without traditional training. In one experiment, Astra boosted a Qwen model's math accuracy from 7.9% to 48.1% by changing just 0.016% of its weights. The approach could fundamentally simplify AI optimization, though it still hits clear limits.
More from Models
- Zero-shot classification is here, yet people won't even fit a threshold on a dev set — antoine_chaffin · 2026-11-05
- Retweet: Google's recurrent transformer could make sub-10B models viable soon — rickasaurus · 2026-11-05
- GPT-6.2 Rumors Hold Up; Launch Could Come as Soon as Tomorrow — kimmonismus · 2026-11-05
- GPT-6.2 Could Launch as Soon as Friday in Astra, Sol, and Cyber Variants — kimmonismus · 2026-11-05
- Nathan Lambert launches free open model economy dashboard with paid analysis series — natolambert · 2026-11-05
- Nathan Lambert on measuring open model inference share: exponential growth vs Google scale — Interconnects (Nathan Lambert) · 2026-11-05