What if you could turn every browser tab into a node in a distributed AI cluster? That's the proposition behind AI Grid, an experiment by Ryan Smith. Visit the page, run an LLM locally via WebGPU, and, if you're feeling generous, donate your unused GPU cycles to the network. Or flip it around: connect to someone else's machine and borrow their compute. It's peer-to-peer inference without the infrastructure headache.

The technical bet here is on WebGPU and WebLLM doing the heavy lifting. No downloads, no Docker containers, no cloud credits. Just a URL and whatever graphics card happens to be sitting in your laptop. The project also includes a memory test, which is a nice diagnostic touch for anyone wondering why their 8GB VRAM machine just flinched. There's something almost subversive about using browser sandboxing as the trust layer for a decentralized compute mesh.

It's a glimpse of what happens when WebGPU stops being a rendering curiosity and starts acting like infrastructure.