The story behind the cartoon
MIT CSAIL researchers handed OpenAI's GPT-6 Astra the weights of Alibaba's small open model Qwen3.5-0.8B and asked it to edit them directly — no training run, no gradient descent.
Astra changed just 0.016% of the weights, 131,072 out of roughly 800 million, tracing the gain to a numerical direction in layers 15 and 19 that made the small model loop instead of answering, and subtracting it.
The edited model's MATH-500 score rose from 7.9% to 48.1%, and the share of answers running to the token limit fell from 93% to 33.1%; on Gemma 3 270M, which rarely loops, the edits produced no established improvement.
The future of AI training is surgery: one robot opens another robot's head, removes the one bad loop, and the patient wakes up six times better at math — no textbooks, no tuition, just a very precise pair of tweezers.

