How many times does a phenomenon need to occur before it officially becomes an industry trend? While that remains a philosophical question for developers and tech enthusiasts alike, a fascinating new movement has firmly taken root among early adopters experimenting with advanced AI assistants. A growing group of developers is pushing the boundaries of what is possible by leveraging artificial intelligence to build extraordinarily ambitious, highly functional projects packaged within astonishingly small file sizes.

The tech community has already marveled at a handful of these ultra-compressed creations over recent months. One recent and widely discussed example involved a developer who successfully managed to compress a No Man’s Sky-style sci-fi exploration game featuring an incredible 150 million procedurally generated star systems into a single, highly optimized HTML file measuring just 299 kilobytes. Now, another developer has captured the attention of the online community by showcasing a breathtaking, fully navigable 3D Atlantis scene that clocks in at an even tinier 96 kilobytes—proving once again that generative AI is unlocking entirely new dimensions of digital design and data efficiency.

This Atlantis Scene Fits in an HTML File Smaller Than Most Pictures

Over on the popular ClaudeAI subreddit, a GitHub user known by the handle Noniv recently shared their latest creative endeavor: an immersive 3D Atlantis project built using Claude. The interactive HTML file allows users to seamlessly fly around a meticulously detailed, visually striking underwater-themed city. When looking at the intricate architecture, the sprawling buildings, and the dynamic lighting, most observers would reasonably assume that rendering such a scene requires downloading a massive software package or heavy asset bundles. Instead, the entire graphical experience fits comfortably inside a file weighing a mere 96KB—making it smaller than many standard compressed digital photographs found on the web today.

What makes this technical achievement even more remarkable is how these types of lightweight 3D browser demos typically function. Historically, miniature browser-based 3D visualizations are purely passive spectacles. They are generally designed as static or animated environments meant strictly to be looked at from a fixed perspective, entirely lacking any form of physical interaction, dynamic user control, or environmental boundaries. However, the developer discovered that Claude was capable of engineering a remarkably robust solution that introduced fully functional collision detection into the ultra-compact environment.

It Even Has Collision Detection

Rather than simply letting the camera clip freely through every digital polygon in the scene, the AI-generated code successfully maps physical boundaries across the entire environment. According to the developer, this means users do not have to worry about phasing directly through the architecture while exploring the virtual city.

People are still doing amazing filesize feats with Claude, as this Atlantis scene clocks in at 96KB

"The ground, the water, quays, bridges (you can fly under their arches), the podium, the crystal, the towers, the domes and the buildings are solid: you glide along them instead of flying into them," Noniv explained when detailing the project to the online community. "Trees, bushes and rocks you can fly through. Press C to switch collision off when you want to pass through something."

This clever implementation of navigational physics transforms the project from a simple graphical tech demo into a genuinely engaging digital space that users can actively explore and pilot through without breaking the immersion of the virtual world.

The secret behind how Claude managed to pack all of these sophisticated 3D graphics, lighting systems, and physics calculations into a microscopic 96KB of digital space lies in the fundamental way the file is constructed. Rather than attempting to bundle heavy pre-rendered 3D asset files, audio files, or complex texture maps—which traditionally inflate the file size of modern video games and web applications—the HTML file instead contains detailed mathematical instructions and procedural code for how the user’s local device should construct the scene in real time.

Because the web browser and the local hardware shoulder the computational responsibility of generating the geometry, lighting, and physics based entirely on the provided code instructions, the project entirely bypasses the need for large file transfers. This efficient approach demonstrates the immense potential of pairing human creativity with advanced AI models to maximize computational economy.

Curious users who want to examine the inner workings of the code can check out the Atlantis project directly on Noniv’s GitHub page, while those interested in piloting a digital flight through the legendary lost city can experience the interactive 96KB web application firsthand by visiting its dedicated live website.