Karpathy Open-Sources Autoresearch Tool

Andrej Karpathy, the creator of Autoresearch, is a prominent figure in the AI world, having served as the Director of AI at Tesla where he led the Autopilot vision team. He was also a founding member of OpenAI, a leading AI research lab. This background in both cutting-edge research and real-world application provides the context for his focus on automating machine learning experiments. The Autoresearch tool operates on a simple yet powerful loop: a human provides high-level research goals in a text file, and an AI agent takes over from there. The agent repeatedly modifies the training script, runs a short five-minute experiment, and evaluates the results. This process allows for rapid, autonomous iteration, with one report suggesting it could run 100 experiments overnight. Positive changes to the model's performance are automatically committed to a git branch, creating a record of successful modifications. This automates the tedious process of tweaking parameters and architectures that consumes a significant amount of a machine learning researcher's time. The evaluation is based on a "bits-per-byte" (BPB) score, where a lower score indicates a more accurate model. This project is a practical example of "agentic AI," where AI systems move beyond simply responding to prompts and begin to reason, plan, and execute complex tasks autonomously. Instead of writing every line of code, the developer's role shifts to that of an orchestrator, directing AI agents that perform the hands-on work. Karpathy himself has referred to this shift as "agentic engineering."