Four researchers at MIT's Computer Science and Artificial Intelligence Laboratory published a blog post this week called "Sparks of Weight-Space Programming", and it has become one of the most discussed pieces of AI research in recent times. They gave OpenAI's GPT-6 Astra the raw weights of Alibaba's small Qwen3.5-0.8B model and asked it to improve the model directly, without any training. Astra performed surgery on over 130,000 of its weights (its internal numerical neurons), and the smaller model's score on the MATH-500 benchmark rose from 8% to 48%.
Within hours, commentators were comparing the work to gene editing for neural networks, and threads on Reddit and Hacker News filled with speculation about models that repair other models and, eventually, themselves. Some readers were more cautious. They noted that the work is a blog post rather than a peer-reviewed paper, that the strongest results without tools came from very small networks, and that the Qwen edit fixed one known fault, a tendency to fall into repetitive reasoning loops. When the team tried to improve a Gemma model that did not have this problem, the gains did not appear.
The method is easiest to place against the established ways of changing what a model does, listed in order of how far each reaches inside the model:
Four ways to change what a model does
Ordered by how far each method reaches inside the model, from the request to the weights themselves. The last is the method in MIT's report.
Prompting
Weights unchangedInstructions, examples or reference documents go in with each request. The weights stay as they are, so the change costs nothing to apply and lasts only for that request.
Fine-tuning
All weights, or an adapterThe model is trained further on a set of examples, and its weights move a little in whatever direction the data pushes. It needs a dataset, training compute and tests that nothing else has got worse. Adapter methods such as LoRA do the same through a small add-on set of weights.
Activation steering
Run time onlyA chosen direction is added to the model's internal activity while it runs, or one feature is clamped at a set level. Nothing is saved to the weights, so the change can be adjusted or switched off, and it needs access to the model's internals.
Weight-space programming
Written by a modelA model is handed another model's weights, works out what they do and writes new values straight into them. There is no training data and no training run, and the editing model rather than a human-written procedure decides what to change.
We wanted to know how practical this would be for a small team, so we ran our own experiment on a small open 135M parameter model. We started from a familiar problem, AI gender bias. Given a sentence about various occupations, the model favoured "he" over "she" more strongly for professions like coach, doctor and engineer. With Astra designing the experiment, we recorded the model's internal calculations, tested which components carried the association and converted the most promising changes into fixed weight edits.
Our first attempts reduced the association by less than 2%. A broader search across 2,010 single edits and 207 combinations, which took about ten minutes on a local Blackwell GPU, produced a four-matrix patch that reduced our measure by 36%. It introduced no new errors on prompts that stated a person's gender, it could be reversed exactly, and it left general prediction quality almost unchanged. The improvement was uneven, however, because the six occupations used during the search improved by 49%, eight new occupations improved by only 15%, and chef and paramedic became worse.
Takeaways: The MIT research shows that a frontier model can read and change another model's weights. Our own experiment shows that it's possible to find a useful, reversible edit for a narrowly defined problem on modest hardware. The difficult work lies in defining the problem, searching widely and measuring what else the edit changes. Editing weights directly could well become a normal way to maintain and improve AI models, and the organisations that benefit most will be the ones with the strongest evaluation practices.