Karpathy Releases Autoresearch: Open-Source Project Lets AI Agents Design and Train AI Models Autonomously

Andrej Karpathy has released an open-source project called autoresearch that is generating significant buzz in the AI community. The project's core concept is ambitious: let AI act as its own scientist, autonomously designing model architectures, modifying optimizers, running experiments, analyzing results, and deciding on next research steps.

According to the project description, the workflow works as follows: humans only need to edit a program.md instruction file. The AI agent then reads the instructions, modifies train.py, runs approximately five-minute training experiments, checks validation bpb metrics, decides whether to keep or discard results, and proceeds to the next experiment. By morning, the system may have completed over a hundred experiments with detailed reports.

The repository includes a complete GPT model implementation, Muon + AdamW optimizer combination, and training loop code. All components are modifiable — architecture, hyperparameters, batch sizes, optimizer types — with the agent iterating autonomously based on experimental results. Karpathy stated on Twitter: "Research is now entirely the domain of autonomous swarms of AI agents running across compute cluster megastructures in the skies."

Due to tool environment limitations, we could not directly verify the GitHub repository details, but multiple Chinese tech media reports indicate the project garnered approximately 46.7k GitHub stars in its first week, described as one of the fastest-growing repositories in GitHub history. This trajectory reflects strong developer interest in the autonomous AI research paradigm.

The potential implications are significant. If this paradigm proves viable, AI research could shift from human-driven manual tuning and experimental design to massively parallel exploration by AI agent clusters operating 24/7. Karpathy also suggested a recursive enhancement scenario: better models produced by autoresearch agents could be used to train the next generation of autoresearch agents.

However, the community has also voiced caution. Cutting-edge AI research involves substantial engineering innovation, theoretical breakthroughs, and intuitive leaps into unknown territory — capabilities that today's AI agents may not fully possess. While autoresearch may excel at optimizing simple tasks autonomously, its real impact on the competitive landscape of frontier AI labs remains to be seen.