MIT researchers say an AI fixed another AI by editing its 'brain' directly, without retraining
In an unreviewed study, OpenAI's GPT-6 Astra changed 0.016% of a small model's settings and its math scores jumped from 8% to 48%.
Researchers at MIT say a top AI model was able to read the inner settings of other AI programs and edit them by hand, much like a programmer fixing a bug, according to a study posted online Monday. The work has not been reviewed by other scientists.
What are 'weights'?
An AI model is made of millions or billions of numbers called weights. They are set during training, a slow, costly process in which the computer adjusts them bit by bit until the model gets better at a task. Humans can't look at those numbers and tell what a model does. That is why AI is often called a "black box."
What the researchers found
The team from MIT's Computer Science and Artificial Intelligence Laboratory gave OpenAI's GPT-6 Astra the raw weights of small AI programs. Just by reading the numbers, Astra correctly named which video game each one had been trained to play, the researchers wrote.
They also asked Astra to improve Qwen3.5-0.8B, a small free AI model from China's Alibaba, at math, without any training. Astra found that the model often got stuck repeating itself and never reached an answer. It traced the problem to one pattern in the weights and removed it, changing only 0.016% of the numbers.
The model's score on a standard math test rose from 7.9% to 48.1%, and its ability to follow instructions roughly tripled, the researchers said.
The limits
The trick didn't work on every model. When the team tried the same thing on Google's small Gemma model, which doesn't get stuck in loops, Astra could not make it better at math. The researchers wrote that editing can remove a problem hiding a skill, but couldn't add a skill the model lacked.
Why it matters
If it holds up, the method could let companies fix flaws or check models for hidden problems without the cost of retraining. The authors, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, called their results "very early" and released full logs for others to check.
Source: "Sparks of Weight-Space Programming," blog post by MIT CSAIL researchers, Nov. 2, 2026.