OpenAI Co-Founder Andrej Karpathy Open-Sources Autoresearch to Automate LLM Research With AI Agents

Large language model research typically requires researchers to repeatedly modify code, train models and compare results, a time-consuming process that is difficult to scale continuously. OpenAI co-founder Andrej Karpathy has open-sourced Autoresearch, which lets AI agents automate the workflow within a fixed framework and highlights the shift in agent technology from content generation toward research automation.

Autoresearch's latest experiment had an AI agent continually adjust the model and its training configuration, retaining effective changes under a consistent evaluation method. Results released by Karpathy showed that the training time needed for the model to reach GPT-2-level performance fell from 2.02 hours to 1.8 hours, a reduction of 0.22 hours, or about 10.9%. The project disclosed neither an investment amount nor a specific release date.