Slava

This is absolutely wild! Researchers at MIT CSAIL have demonstrated something that could fundamentally change how we build and optimize neural networks: "Weight-Space Programming". Using GPT-6 Astra, they showed that an LLM can directly read, understand, and rewrite the weights of other neural networks, without gradient descent or retraining! In one experiment, Astra discovered a reasoning-loop problem inside Qwen3.5-0.8B and fixed it by modifying just 0.016% of its 800 million weights. The result? Its MATH-500 score skyrocketed from 7.9% to 48.1%, while instruction-following performance nearly tripled! Even more remarkably, Astra was able to construct neural networks from scratch, directly writing their weights to play video games. These networks outperformed models trained with 100 million steps of reinforcement learning! Think about the implications. Instead of spending enormous amounts of compute retraining models, future AIs could potentially inspect, debug, and directly reprogram the neural networks of other AIs. We're only seeing the first experiments, but this is a fascinating new direction for AI research.