autoresearch-webgpu runs Claude agent training loop in browser

Stripping the autoresearch concept down to a zero-install browser demo is the right move — it turns an intimidating GPU workflow into something anyone can poke at in an afternoon.

• –The core loop mirrors Karpathy's original: LLM writes training code → runs experiment → reads loss → proposes next hypothesis → repeat; the WebGPU port just removes every infrastructure prerequisite
• –Relies on Eric Zhang's `jax-js` for browser-native GPU-accelerated tensor math — the real technical enabler here
• –Claude is the agent generating and iterating on `train.ts`; the project is a concrete example of LLM-driven code synthesis in a tight feedback loop
• –Part of a broader mid-March 2026 wave of autoresearch forks (distributed swarm variants, Triton kernel optimizers, etc.) — the WebGPU fork stands out for accessibility, not raw capability
• –Early stage: ~10 GitHub stars, solo maintainer, crashes on mobile — more proof-of-concept than production tool