GitHub designer builds a particle-cloud QR alternative with GPT-6 Astra

Why it matters

Octoglow shows how frontier models can turn a designer's screenshots and interaction concept into a working camera prototype, while leaving reliability as the hard part.

Maxime De Greve (@MaximeDeGreve), a designer and developer on GitHub's Copilot team, published a browser experiment on September 14th that replaces the familiar black-and-white QR square with a moving cloud of light generated and read on separate screens.

https://x.com/MaximeDeGreve/status/2099508706186588301

The project, called Octoglow, came out of a weekend spent using OpenAI's GPT-6 Astra through the GitHub Copilot app. De Greve described the result in a thread on X as an attempt to replace QR codes for screen-to-screen scanning.

De Greve is a Belgian designer and developer living in London. His GitHub profile identifies him as a designer on the Copilot team, while his portfolio lists earlier design and development work at collaborative-design startup Marvel. That hybrid background matters here: Octoglow treats the model as an implementation partner for a designer already comfortable working in React, Swift and other frontend tools.

A six-digit code made of light

Octoglow presents two modes: "Code" and "Camera." One screen generates a six-digit value and renders it as an animated particle cloud. A user opens the same page on a phone, switches to Camera mode and points the camera at the first display to read the value.

The design borrows from Apple's device-pairing interface. In Apple's description of proximity sign-in, a trusted iPhone or iPad pairs with another device by scanning an animated particle cloud. De Greve said that Apple's Quick Start experience was the starting point for his experiment.

The current Octoglow page exposes a regenerate button, the resulting six-digit code and a frames-per-second counter. It runs in the browser using HTML5 Canvas, which makes the experiment easy to test across a laptop and phone without installing an app.

De Greve has already identified the conditions most likely to break it. Responding to a user who asked about Pixel compatibility, he said glare and lower-quality cameras still presented rough edges. Those are central constraints for any visual code intended to work across different displays, lenses and lighting conditions.

The prompt started with screenshots

De Greve said his initial input consisted of a few sentences about iOS Quick Start and several screenshots. His first task for Astra was to work out how Apple's scanning interaction had been built. He then used the model inside GitHub Copilot to develop the browser prototype.

That workflow lines up with how OpenAI is positioning GPT-6 Astra. OpenAI began rolling out the model on September 3rd and says it has stronger visual judgment for websites, applications and rendered scenes, alongside improved software-engineering and computer-use capabilities. Its API documentation lists image input, a 1.05 million-token context window and pricing of $10 per million input tokens and $50 per million output tokens.

OpenAI's own numbers leave room between a polished demo and reliable design automation. Astra scored 50% on the lab's internal design tasks, compared with 47.4% for GPT-5.6 Sol. The improvement is measurable, though far smaller than the jumps OpenAI reports on several coding and computer-use evaluations.

Octoglow offers a more useful view of the model's role than another benchmark table. De Greve supplied the visual reference, interaction concept and taste. Astra helped convert those inputs into a functioning camera-based system over a weekend.

Pretty is the easy part

Octoglow does not yet match the engineering guarantees behind the symbol it is trying to replace. Denso Wave's QR specification supports thousands of numeric characters, uses corner markers for reading from any direction and includes error correction that can recover data when part of a symbol is obscured or damaged.

De Greve's experiment currently communicates a six-digit code and comes with acknowledged camera and glare limitations. That keeps Octoglow in prototype territory, particularly for payments, identity or other uses where a misread code carries consequences.

The narrower device-pairing use case is a better fit. A short code displayed briefly on one trusted screen does not need the capacity of a general-purpose QR symbol. It does need fast recognition across cheap cameras, reflective displays, awkward angles and uneven brightness. Octoglow shows the interaction working. Its next challenge would be proving that the particles remain readable after the weekend-demo conditions disappear.

For designers who can already specify an interaction and judge its output, the build is evidence of how quickly frontier coding models can collapse the distance between a visual reference and a testable system. The model supplied implementation speed. De Greve supplied the product decision that made the code worth generating.