An AI rewired another AI’s brain. Math scores jumped sixfold.
MIT researchers let GPT-6 Astra edit a small math model directly — no training, no data. It changed just 0.016 percent of the model’s weights, fixing a thinking loop where the model repeats itself, and math scores rose up to sixfold on the MATH-500 benchmark. Then Astra wrote a game-playing brain from scratch: it never practiced Breakout, it just wrote the weights.
Storyline
Hook: An AI rewired another AI’s brain. Math scores jumped sixfold.
Context: MIT researchers let GPT-6 Astra edit a small math model directly. No training. No data. It changed just 0.016 percent of the weights.
Stakes: The fix targeted a thinking loop, a glitch where the model repeats itself. Then Astra wrote a game-playing brain from scratch — it never practiced Breakout, it just wrote the weights.
Payoff: Retraining costs millions and can erase old skills. Editing weights could fix one flaw without breaking the rest.
Watch next: Best AI Scores Only 12% on Real Jobs — https://simpleaiexplain.com/2026/11/02/best-ai-scores-only-12-on-real-jobs/
How we verified this: Facts checked against The Register’s November 3, 2026 coverage of the MIT CSAIL paper “Sparks of Weight-Space Programming” (Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab) and the paper listing. The sixfold figure is the reported MATH-500 improvement; instruction-following rose roughly threefold on IFEval. The authors describe this as early work on small models, not a replacement for training.
Links:
Instagram reel: https://www.instagram.com/reel/DfSNoADghbO/
Sources: https://www.theregister.com/ai-and-ml/2026/11/03/forget-retraining-ai-can-now-tweak-another-ais-weights/5307105 and https://www.alphaxiv.org/abs/2611.weight-space-programming
Instagram: https://www.instagram.com/simpleaiexplains/
Leave a Reply