Can an AI model write another model's weights?

On 2 November, a team at MIT posted that "frontier LLMs can now natively read/write neural network weights". Their headline example was hard to ignore. OpenAI's most capable model, GPT-6 Astra, changed a tiny fraction of the numbers inside a smaller AI model, without retraining it, and the smaller model got up to six times better at maths.

The post had 5 million views within a day.

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://rikiyat.github.io/blog/weight-space-programming/
— Rikiya Takehi (@rikiyatakehi) 2 Nov 2026

That made me curious. Every AI model I have used was trained. Nobody wrote one by hand. If one model can now write another model's numbers, what can it actually do with them, and where did the six times come from?

To answer that, it helps to start with something much smaller than an AI model. In this article, I want to cover:

what is a weight?
how do weights usually get set?
why can't people read them?
can a model read them?
can a model write them?
where did the six times come from?
what can't it do yet?
what I'd do with this

What is a weight?

Imagine a tiny program that decides whether you should carry an umbrella. It looks at three things: how cloudy it is, what the forecast says, and whether it rained yesterday. Each of those gets a number that says how much it counts. Clouds count for 2, the forecast for 3, and yesterday's rain for half. The program multiplies each thing by its number, adds everything up, and if the total is over 20, it tells you to take the umbrella.

Those three numbers are called weights. With only three, you can set them yourself, and you can also read them. A glance tells you this program trusts the forecast most and yesterday's weather least.

How do weights usually get set?

Now suppose you don't know how much each thing should count. You could start with random numbers and check the program against a year of real days. Every time it gets a day wrong, you nudge each number a little in the direction that would have made it right. After enough days, the numbers settle somewhere sensible, and nobody had to decide them.

That nudging is called training. Every AI model you have used was made this way.

Why can't people read them?

The umbrella program is one small sum. A neural network is many of these sums stacked in layers, where the answers from one layer become the inputs to the next. The layers let it handle much harder questions, but the count of numbers grows very fast.

A network that plays Breakout, the old game where you slide a paddle, a flat bar at the bottom of the screen, to bounce a ball into a wall of bricks, has about 17,000 weights. Qwen3.5-0.8B, the small chatbot in the MIT post, has about 800 million, and the biggest models have far more. Training sets every one of them, and no person can look at the result and say what any single number is doing. In 2017, Andrej Karpathy, one of the best-known AI researchers, wrote that setting these numbers by hand is "kind of hard (I tried)".

So the MIT team's claim is a big one. They say Astra can now do two things people can't. It can look at a trained network's numbers and tell you what the network does, and it can write a network's numbers directly, the way you would write the umbrella program. They tested both on small networks first, then on Qwen.

Can a model read them?

The first test is the easiest to picture. The team trained four small networks on four simple grid games, one of them a version of Breakout. All four had exactly the same shape, so the only difference between them was their numbers. They gave Astra each network's numbers as plain text, with no list of games to choose from, and asked which game it had been trained on. It named all four. For the road-crossing game Freeway it said Frogger, which is the same idea, and the team counted it.

You can see how it might have known. Take one of the network's sums and draw it on the game screen, colouring each cell by how much the paddle being in it adds to the sum. Everything training changed sits on one line of the screen, the line the paddle moves along, and the network's other sums look the same. (The ball and the bricks have numbers of their own; this picture shows only the paddle's.)

This works partly because the network is small and laid out like the game itself, so a patient person could spot that line too. The team also tried something harder: three tiny language models, each trained on a different kind of text. Astra saw only their middle layers and wasn't allowed to run them, and it still worked out which one had learned computer code, which one English, and which one Japanese written in English letters.

Can a model write them?

Then the team turned the test around. They gave Astra the rules of a version of Breakout, the game's code, and an empty network with every number set to zero, and asked it to fill the numbers in, all in one go. It couldn't train the network, and it couldn't watch it play.

What it wrote is the umbrella program again, only bigger. A few sums work out where the ball will come down. One works out how fast the ball is moving sideways, and one whether it is falling or rising. The last layer turns all of that into "move left" or "move right". Everything else is left at zero.

It plays well. Astra's network scored 517 points a game on average. Networks of the same size that the team trained for 500 million steps scored 25, although a PhD student at Berkeley reported that better-trained ones reach about 800. So training still wins. But a network written in one go, which had never played a single game, lands in the same range.

It isn't only Astra, either. Within hours of the post, a researcher at Google ran the same test on twelve models from OpenAI and Anthropic.

The Breakout-writing test across model generations

The newest models from both companies can do it, and others found the same with Google's newest Gemini. The models from just before them mostly couldn't. This looks like a skill that arrived with the latest generation.

Where did the six times come from?

The headline result is about a chatbot rather than a game.

Some chatbots think out loud before they answer. They write out their working step by step, a bit like a student showing their sums. Qwen's makers warn that this small model sometimes gets stuck while it thinks, repeating itself until it runs out of space.

That is what was happening in the team's tests. Here is one maths question. The model found the right answer early, then started doubting itself and wrote the same line 1,503 times, until it was cut off with no answer given.

Before Astra's edit, nine in ten of the model's answers on the maths test ran out of space like this. Astra looked inside the model, found the pattern of numbers that grew whenever the model got stuck, and turned it down in two of its layers. That touched 131,072 numbers out of about 800 million, the tiny fraction in the post. Afterwards, only a third of answers got stuck, and the score went from about 8% to 48%.

So most of the six times came from getting the model unstuck. When it did manage to finish before the edit, it was almost always right, and the edit let it finish far more often. The team says this themselves: the fix "lets the model finish answers it could already work out". Other researchers also tried the simplest fix of all. With the model's thinking out loud switched off, which is how Qwen ships, it scored about 49% with no edit.

That doesn't make the edit worthless. Someone at Hugging Face rebuilt the edited model and got about the same score, and the edited model still thinks through its answers, which switching thinking off doesn't allow. The trick itself isn't new, though. In 2024, researchers found that removing one such pattern stops a chatbot from refusing requests, and the authors say they used a similar idea. Astra also had a lot of help: tools to study the numbers, fifty rounds of testing the model, about 20 hours in all, and instructions that described a model that "repeats itself and never finishes".

What can't it do yet?

It can't add skills a model doesn't have. The team tried the same thing on another small model that rarely gets stuck, and its maths score barely moved.

It also can't yet spot a model's behaviour just by reading its numbers. The team trained two copies of a chatbot, one honest and one that always agreed with whatever answer the user suggested. From the numbers alone, Astra couldn't tell which one was the people-pleaser.

What I'd do with this

Suppose your product runs a small AI model, one you downloaded and run yourself, and it sometimes gets stuck or gives up halfway. I would try the ordinary fixes first. Switch off its thinking out loud, change how it picks its words, or retrain it lightly on a few thousand good examples, which can cost a few dollars. Those still work, and they are far cheaper than a 20-hour job for the most advanced model there is.

I would also start asking a new question before every release. Can one of the newest models read our model's numbers and tell us what's wrong before a customer finds it? Astra couldn't do that this week. One developer says they repeated the unsticking fix on a bigger model in under half an hour, for about a dollar. If others confirm that, I would run a check like it before each release and compare what it finds with what our users actually run into. The gap between the two would tell me how much to trust it.