Astra can read and write weights to improve Qwen
Points and comments are a snapshot, not live.
LLMs can read and edit neural network weights natively without gradient descent.
MIT researchers show GPT-6 Astra can identify games from raw weights, fix a Qwen3.5-0.8B model's math reasoning by editing 0.016% of weights, raising MATH-500 from 7.9% to 48.1%, and write game-playing networks from scratch that outperform PPO-trained ones. The method, called Weight-Space Programming, lets the LLM reason about what each part of a network computes and adjust weights accordingly.
What commenters are saying
Commenters are stunned by the result, calling it potentially as significant as 'attention is all you need.' Many note the implication for recursive self-improvement. Some question whether this capability was explicitly trained or emergent, noting that training data includes raw weights from ML repos. Others connect it to mechanistic interpretability work. Skeptics point out Astra is vastly larger than the networks it edits, so self-improvement remains unproven.