lucasgelfond / autoresearch-webgpu

What is autoresearch-webgpu?

This TypeScript/Svelte app ports Andrej Karpathy's autoresearch to the browser, letting AI agents run research experiments and train small language models entirely on your device's GPU via WebGPU. You pick manual mode to tweak and run training code yourself, or auto mode where Claude generates code, evaluates bits-per-byte loss on Shakespeare data, and iterates via a leaderboard of experiments. No Python setup – just load the live demo at autoresearch.lucasgelfond.online and watch loss curves drop in real time.

Why is it gaining traction?

It stands out by ditching heavy ML stacks for instant browser training: visualize live loss charts, compare experiments on a leaderboard, generate text samples from trained models, and export CSV/ZIP results. The hook is "github let's build from here" – Claude handles the heavy lifting like "let agents transform data quality," making ML iteration feel like a game without installs or cloud costs.

Who should use this?

ML hobbyists prototyping tiny transformers, web developers exploring on-device training, or educators demoing agentic workflows like "let github copilot's agent mode handle the heavy lifting." Perfect for quick experiments on Shakespeare or byte-level text, especially if you're on Apple silicon with WebGPU.

Verdict

Fun proof-of-concept for browser ML – try it for "let's do automation" vibes, but with 13 stars and 1.0% credibility score, it's early-stage; expect rough edges in docs and stability. Solid for tinkering, skip for production.

(187 words)