The outputs of AI models are the result of a complex interplay between neurons, weights, training, and inference, and are therefore difficult to follow.
But in some ways they're starting to look more like open books. Recent work by Anthropic safety fellow Joshua Fonseca Rivera has revealed how certain neurons in a neural network can be associated with a specific persona.
Now, computer scientists from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have found that frontier AI models like OpenAI's GPT-6 Astra can rewrite the weights of another model directly, as opposed to indirectly through training, and alter their behavior.
"We find early signs that frontier LLMs are surprisingly good at natively reading, writing, and reasoning about neural network weights," said Rikiya Takehi, a PhD student at MIT CSAIL, in a social media post. "Instead of training, models may treat raw weights like readable, editable code: diagnosing failures and patching them directly."
To demonstrate that, Takehi and colleagues Jayden Teoh, Alex Zhang, and Omar Khattab asked GPT-6 Astra to edit Qwen3.5-0.8B weights to make it better at math.
Astra changed only 0.016 percent of the model's weights, reducing its reasoning loops and improving its math performance by as much as sixfold on the MATH-500 benchmark, while boosting its instruction-following score on the IFEval benchmark roughly threefold.
There was no training using gradient descent; Astra just revised the numerical values that constitute Qwen's weights.
"All of this is possible without allowing Astra to write training loops," the researchers wrote in a blog post. "It uses its reasoning to figure out what each part of a neural network computes, or should compute, and changes the weights accordingly. We call this Weight-Space Programming.
"Throughout the history of deep learning, we encoded our simplistic inductive biases into tiny symbolic optimizers like Adam or Muon. This is changing: our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights."
Astra's ability to program weights extends to being able to write a neural network from scratch, zero-shot. The researchers told Astra: "write, from scratch, the weights of a policy network that plays [Breakout] as well as possible." And it did so.
Takehi suggests this technique shows promise for improving models without the consequences of catastrophic forgetting, which can occur when a model is revised to include new knowledge.
More generally, the authors expect Weight-Space Programming will open the door to new ways to understand models and optimize them, and to implement recursive self-improvement. The technique may also prove useful for reducing model sycophancy and improving interpretability and alignment.
The next step, the authors say, involves building a benchmark and a harness for carrying out Weight-Space Programming on LLMs. "We believe this will provide a standardized evaluation to measure and accelerate progress in Weight-Space Programming, turning what currently seems like black magic into a rigorous, reproducible engineering discipline," they say. ®