Developer Tools●●Solid
I trained a small local model to translate natural language to CLI
Offline CLI command translator, but model training data stays private—can't verify accuracy.
Niche GemShip It
kiki_kuuki
103mo ago
Back to browse
Autoresearch-WebGPU uses agents to iteratively train LMs in the browser
by lucasgelfond·Mar 14, 2026·1 point·0 comments
AI Analysis
●●●BangerWizardryBig Brain
Runs autoresearch agents entirely in-browser using WebGPU and jax-js.
Strengths
- •Removes Python environment setup friction for running autoresearch experiments locally.
- •Leverages jax-js to run actual training loops client-side via WebGPU.
- •The agent iterates on training code autonomously without incurring server costs.
Weaknesses
- •Browser memory constraints limit model size and training duration significantly.
- •Relies on Karpathy's original logic rather than novel algorithmic improvements.
Category
Target Audience
ML engineers, browser tech enthusiasts
Similar To
Karpathy's autoresearch · TensorFlow.js · Hugging Face Spaces
Post Description
title! thought the Karpathy autoresearch demo was really neat, didn't want to bother with Python setup and thought it would be neat to show how much you can do now in the browser (with jax-js/webgpu!)
lmk what you think! neat to see agents write training code / first sparkles of "recursive self-improvement"!
Similar Projects
AI/ML●●Solid
Meadow Mind – a 7B diffusion LLM plays Gym games with zero training
Fixed-latency language-rule decisions beat traditional token-by-token LLM agents.
Big BrainWizardry
akaiHuang
307d ago
AI/ML●●●Banger
Run RL agents in the browser with WebGPU
WebGPU-powered RL training in browser—no install, no cloud, just train.
WizardryCrowd Pleaser
umjunsik132
3123d ago
AI/ML●●●Banger
ChonkLM – Tiny language models running offline in the browser
Runs GGUF models in the browser via custom WGSL shaders when cloud APIs ignore tiny models.
Zero to OneWizardryNiche Gem
bilalba
601mo ago
AI/ML●●Solid
EleutherAI / Lm-Evaluation-Harness
Industry standard benchmark harness refactored with lighter installs and new SGLang support.
Crowd Pleaser
marvinified
101mo ago
AI/ML●●Solid
Nanbeige 4.1-3B running in the browser via WebGPU
WebGPU LLM inference in-browser is slick, but Ollama, LM Studio, and local alternatives already work offline.
WizardryShip It
victormustar
613mo ago