interactive projects
AI Demos
Experiments you can run in the browser, with the implementation and technical details alongside them. Every model here is running live on my GPU.
What does a language model actually remember?
RWKV-7 carries what it reads in a fixed-size recurrent state instead of a growing attention cache. Goose Lab explores images and motion read from a language model's recurrent memory. The research journal follows the project from its first CIFAR and shapes experiments through 7.2B still images and instruction-driven synthetic video, with measured results, original samples and open questions. It stays open while the live demos are asleep.
the picture in the memory, one frame per token
Neural Wave Sandbox
Draw on the water. A physics simulator seeds the first frames, then a ~18M-parameter video transformer I trained rolls the ripples forward one frame at a time on my GPU and streams them to your canvas. It learned its physics in a walled tank, so the waves bounce off the edges and interfere on the way back.
draw a splash →
preview · the real thing is a transformer
Spectral Text Encoder
Sentence embeddings you can watch: a transformer trained from scratch turns every token of your sentence into sine waves — amplitude, frequency, phase — and their superposition is the embedding. Type a sentence and the model synthesizes its signal live on my GPU, with the full technical report below the demo.
try it →
wave(“a sentence becomes a wave of meaning”) · channel 4
This Website + Self-Hosted Model Inference
A FastAPI + Tailwind site running in Docker on my own machine, exposed through a Cloudflare Tunnel, with live GPU inference, an admin control plane, and adaptive serving limits. Bigger models, like Goose Lab's, run in their own sandboxed container and are loaded on demand from the admin panel.