I'd been mulling over this crazy idea for a while. Can programs be generated? Inspired by recent advances in world models, I wondered if we could do away with source code and generate pixels directly and interactively.

As an experiment to answer this, I set out to create a neural window manager, training a neural network to predict what the screen would look like next.

Basically, the idea was to generate the next frame based on the last two frames and the mouse position. That's it: moving windows without programming an event system, just a simple convolutional neural network guessing pixels.

To implement the experiment, I used Pygame to simulate a turquoise desktop background, a gray window with a navy blue title bar, a white cursor, and four colors in total. Then, a bot randomly dragged the window, and I recorded everything, processing the frames as color index matrices (not RGB, to avoid complications) and the mouse delta (dx, dy, click) that caused each transition. 8000 frames, a few minutes in Colab.

The model is a unitary neural network (UNET). The encoder compresses the stacked frames, the decoder reconstructs the next one, and the mouse vector coordinates are projected with a linear layer to fit the spatial size of the bottleneck. There, they are concatenated before decoding, so that motion information feeds each jump connection.

And it works! Which still surprises me a little. You can drag, and the window follows you; when you release, it stops. There's no internal state, no (x, y) coordinates anywhere. The model infers the position from what it sees, which works until it doesn't. But after a couple of seconds of strange movement, the window starts to distort.

This will probably improve with more computing power for training and more examples, but to narrow the scope of the experiment and test it within a web browser, I decided to abandon the rendering aspect and have the model predict primitives instead of pixels, simply converting the motion engine into a neural network.

Basically, I trained a small MLP to receive (distance to the title bar, distance to the resize point, click) and generate (dx, dy, dw, dh), with two separate heads: one for moving and one for resizing. The trick is that they share nothing except the click signal, so the model can't confuse dragging with resizing. I then exported it to ONNX as well, and now everything runs in the browser, without a server, just a canvas element and two small neural networks communicating with each other.

With this new approach, the renderer remains deterministic, with rectangles drawn in JavaScript, but the window's behavior (where it moves, how it resizes) is learned from examples. It feels like a peculiar middle ground between traditional and neural, so you can feel the space the network has learned by interacting with it: dragging near the title bar moves it, but approaching the corner resizes the window. There are no conditionals or hitbox code; the network simply learned where those areas are from examples.

Sometimes it gets confused near the edges, which, frankly, is more interesting than if it worked perfectly; you can perceive how the probability changes. This makes sense when you think about it, because no (x, y) coordinates are stored in these models; the position is implied in the activations. It works well for short sequences, but fails when asked to maintain state over time.

Update: A few weeks later, Meta published the Neural Computers article (2604.06425, it's worth reading). The premise is the same, but they go much further: cli and uis, real programs. Their failure modes are practically identical to those I found with the pure pixel version: "challenges persist with routine reuse, controlled updates, and symbolic stability." which is a fancy way of saying that the window blurs after a few seconds (that was the reason for choosing deterministic rendering).