Summary
MIT researchers have shown that frontier AI models can read and directly edit the raw weights of another neural network, a technique they call weight space programming. This video from TheAIGRID walks through the research, led by a team that includes Omar Khattab, creator of DSPy, and Alex Zhang, known for work on recursive language models.
In the headline experiment, OpenAI’s GPT6 Astra was given the weights of a small Qwen 3.5 0.8B model and rewrote just 0.016% of them, about 131,000 weights, to remove reasoning loops. No retraining or gradient descent was involved. Scores on the MATH 500 benchmark rose from 7.9% to 48.1%, and instruction-following performance roughly tripled. The same approach failed to meaningfully help Google’s Gemma 3, which doesn’t suffer from those loops, suggesting the method removes obstacles rather than adding new abilities.
The video also covers independent replications within a day. Tests across model generations suggest the capability emerged only in the latest releases, with Claude Opus 5.5 and GPT6 Astra showing roughly 20x improvements over baseline. One replication on Qwen 3.5 4B took 27 minutes and about $1 of API usage. The video closes with expert skepticism about hype, what this means for conventional training, and whether it points toward recursive self-improvement.
📺 Source: TheAIGRID · Published November 04, 2026
🏷️ Format: News Analysis
