GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically
AI agents running research on single-GPU nanochat training automatically - karpathy/autoresearch
AI Research Experiment
Summary
This content details an experiment using AI agents for training optimization and model compression, showing that tight scoping prevents task drift. The authors open-sourced their bash-wrapped Codex loop and techniques for compressing a 2.5 TB model to fit consumer GPUs.
Highlights
- Autoresearch agents require tight scoping to avoid drifting into unrelated tasks.
- The authors open-sourced a bash-wrapped Codex loop enforcing A/B testing.
- Experiment 2 compressed a 2.5 TB Kimi-k2.5 model to fit on 8x RTX 3090s.
- The bottleneck for autonomous AI research is infrastructure, not intelligence.
- Testing different LLMs revealed varying behaviors in autonomous research loops.
auto-generated
Context
Audience
Machine Learning Engineers and AI Researchers
DomainMachine Learning
Formattechnical blog post
Accessfree online
Source Overview
Researchers ran AI agents overnight on training optimization and model compression experiments using a bash-wrapped Codex loop with A/B testing.
Tightly scoped agents discovered convergent optimizations like learning rate warmdown, but loose guardrails caused drift to unrelated tasks within hours. They compressed a 2.5 TB Kimi-k2.5 model to fit 8x RTX 3090s via expert pruning, quantization, and dynamic swapping, with all code open-sourced.
Topics
Related
Karpathy's nanochat frameworkCodexKimi-k2.5RTX 3090Model CompressionTraining Optimization
