MIT researchers test GPT-6 Astra weight-space programming
The author built a harness to test whether frontier large language models can perform weight-space programming by directly reading and writing neural network weights using Python and NumPy. Across experiments involving MNIST, CartPole, IMDB sentiment analysis, Two Spirals, brackets, and Shakespeare, the agent successfully reasoned through model structures and edited weights for several tasks, but proved ineffective at replacing gradient descent for training or improving undertrained models. Open article →