karpathy/autoresearch
Andrej Karpathy's framework for autonomous AI-driven ML research. AI agents modify a single training script, run 5-minute experiments on a single GPU, evaluate results against a validation metric, and repeat — running roughly 100 experiments overnight while you sleep. Zero dependencies beyond PyTorch. One GPU, one file, one metric.
Why It Matters
AutoResearch represents a fundamental shift in how ML research gets done. Instead of a researcher manually tweaking hyperparameters and architectures over weeks, an AI agent systematically explores the design space at 12 experiments per hour. Karpathy — former Tesla AI director and OpenAI co-founder — designed this as the minimum viable autonomous research loop: agents edit train.py (which contains the full GPT model, optimizer, and training loop), train for exactly 5 minutes, check if val_bpb improved, keep or discard, and iterate. The simplicity is the point. No distributed training, no complex configs, no cluster requirements. The project has already spawned dozens of forks exploring extensions like multi-objective optimization and architecture search.