GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights to fix its math
No gradients, no training, just a frontier model reading and rewriting another network's parameters.
Rikiya Takehi says his team asked GPT-6 Astra to edit the weights of Qwen3.5-0.8B to make it better at math, and it did, changing 0.016% of the weights with no gradients and no training. He reports math improving up to 6x and instruction following 3x. The team calls the technique Weight-Space Programming, and the claim is that frontier models can now natively read and write neural network weights.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights!
The numbers underneath come from a summary of the MIT study posted by Latent.Space. Qwen3.5-0.8B often burned 32K tokens on a math problem without ever reaching an answer. GPT-6 Astra traced those loops to two layers, edited them, and with sampling held fixed the small model's MATH-500 score rose from 7.9% to 48.1%.
Why the fixed sampling matters
That last detail is the one worth holding onto. Small models' benchmark scores move around a lot depending on temperature and decoding settings, so holding sampling fixed is what separates a real change in the network from a change in how you are reading it. The jump is attributed to the edit itself, not to a kinder evaluation setup.
The reaction from people who build with these models was mostly stunned. Theo's entire response was a single wide-eyed emoji. swyx reached for a Ted Chiang reference, saying that with weight-space programming we are now at paragraph two of Exhalation, the story in which a being opens its own head and inspects the machinery doing its thinking.
What this is not
This is one study on one 0.8B model aimed at one capability, and it is not something you can run on your own checkpoints today. Nobody in the sources has reproduced it, and there is no claim here about larger models or about edits that survive further fine-tuning. What is new is the existence proof. Until now, improving a model meant training it. This is the first concrete claim that a frontier model can treat another network's parameters as a codebase, find the bug, and patch two layers.
GPT-6 Astra traced the loops to two layers and edited 0.016% of Qwen’s weights. With sampling held fixed, its MATH-500 score rose from 7.9% to 48.1%.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.