HuangShengZeBlueSky / MLP_AutoResearch

What is MLP_AutoResearch?

MLP_AutoResearch ports Karpathy's AutoResearch idea from his github projects like nanogpt and llm-council to simple MLPs in Python with PyTorch. It lets AI agents like Claude or GPT autonomously tweak model architectures, hyperparameters, and optimizers on MNIST digit classification, training fixed 20-epoch runs and keeping only improvements based on test accuracy. Developers get a hands-off loop: set rules via a prompt, run baseline experiments with a single command, and watch agents iterate via git commits.

Why is it gaining traction?

Unlike manual tuning in Karpathy's github llm or makemore repos, this runs on CPU without GPUs, making agent-driven optimization accessible beyond H100 clusters. The hook is pure experimentation—agents explore residuals, activations like ReLU or GELU, batchnorm, and schedulers in a greedy loop, logging results to TSV for easy tracking. It's a lightweight bridge from Karpathy's rnn effectiveness studies to MLP ablation.

Who should use this?

ML students replicating deep MLP papers on MNIST, researchers prototyping agent workflows before scaling to Karpathy-style llm-c or nanochat setups, or Python devs curious about autoresearch for hyperparameter sweeps. Ideal for those tired of scripting grid searches in notebooks.

Verdict

Fun teaching project with solid docs and easy uv/pip setup, but at 17 stars and 1.0% credibility score, it's immature—results are TBD, no tests. Try for learning Karpathy agent patterns, skip for production.

(178 words)