Andrew Karpathy’s “autoresearch”: An autonomous loop where AI edits PyTorch, runs 5-min training experiments, and continuously lowers its own val_bpb. "Who knew early singularity could be this fun? :)"

Former OpenAI and Tesla AI researcher Andrew Karpathy has showcased a groundbreaking project called 'autoresearch,' which demonstrates an autonomous AI agent capable of conducting machine learning research without human oversight. The system operates in a continuous loop where an AI agent edits PyTorch training code on a dedicated git feature branch, runs 5-minute training experiments, and analyzes results to make further improvements. Each dot in Karpathy's visualization represents a complete training run, with the agent accumulating git commits as it discovers better configurations that yield lower validation loss (measured as val_bpb, or validation bits per byte).

The agent autonomously optimizes multiple aspects of the training process, including neural network architecture decisions, optimizer selection, and hyperparameter tuning. This creates what Karpathy describes as an 'autonomous loop' where AI can 'continuously lower its own val_bpb.' The system enables researchers to compare different prompts and agent configurations to determine which produces the fastest research progress. As Karpathy noted in his post, 'Who knew early singularity could be this fun? :)' suggesting this represents an early form of self-improving AI systems.

This approach fundamentally changes how AI research could be conducted, potentially accelerating discovery by orders of magnitude. Instead of human researchers manually tweaking parameters and running experiments, the AI agent can explore the optimization space continuously, 24/7, making incremental improvements through systematic experimentation. The 5-minute training cycle allows for rapid iteration, while the git-based version control provides full transparency into the agent's decision-making process and experimental history.