Summary
Matthew Berman breaks down a new MIT paper on weight space programming, in which a general-purpose frontier model such as OpenAI’s GPT6 Astra reads and writes another neural network’s raw weights directly, without gradient descent. He begins with a quick primer on neural networks, weights and why they are considered a black box.
The video then steps through the paper’s experiments. In the first, Astra was given only the weights of four networks trained on Breakout, Pong, Pac-Man and Freeway, and correctly named each game with no options provided. In another, it identified whether small models had been trained on English, Japanese or Python using only intermediate layers, and it predicted internal components that matched measured behavior.
The most striking experiment asks whether Astra can repair a failing network. Given a model’s weights, a scene description and one recorded episode, it explained why a simulated car kept getting stuck on a ramp and wrote new weights, changing 3.5% of the parameters. The car then succeeded in all 1,024 episodes.
Berman discusses the implications for interpretability and for self-improving AI, since a model that can edit another model’s weights might eventually do the same for itself.
📺 Source: Matthew Berman · Published November 04, 2026
🏷️ Format: News Analysis
