Neural network weights are usually treated as unreadable: millions of numbers produced by an optimiser, inspected only indirectly. A blog post from four MIT CSAIL researchers, picked up by the Two Minute Papers video channel, argues that GPT-6 Astra can read and write them more like code, and calls the idea weight-space programming.
The experiments start small. Four identical networks, each with 17,253 weights, were trained to play grid versions of Breakout, Pong, Pac-Man and Freeway. Given only the raw weights as text and no list of options, Astra named the right game for all four, in one case with 99% confidence. It identified what shape each of three small diffusion models drew, again from weights alone. With the input and output layers removed from three tiny character-level language models, it named the training language, Python, romaji Japanese or English, and found every copying head, matching what the researchers measured by running the networks.
Then it wrote weights. Handed a policy that drove a car straight into a ramp and stuck, Astra explained the failure and, in one attempt without tools, edited 3.5% of the weights so the car succeeded in all 1,024 test episodes. The headline result is on a real model: asked to make Qwen3.5-0.8B better at maths without training, Astra traced the model's known habit of looping in thinking mode to one direction written by MLP layers 15 and 19, and subtracted it, changing 0.016% of the weights. MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval, an instruction-following test, from 21.0% to 62.3%.
For readers of our GPT-6 Astra review, this is a striking demonstration of the model reasoning about structure rather than surface text. It is also the source of a quick follow-up: the Tree-Ising preprint, which we cover separately, used the agent's public logs to recast the same repair as a solver problem and reached a slightly higher MATH-500 score for about $0.90 of GPU time.
That follow-up is the right frame for the caveats. This is a blog post, not a peer-reviewed paper, and most experiments use networks far smaller than anything in production. The Qwen fix removes one well-documented failure mode rather than adding new ability, and a cheap optimiser matching it suggests the hard part was the diagnosis, not the edit. The capability looks real; how far it scales, and what it costs per task, is still open.