‘The Karpathy Loop’: Former OpenAI researcher’s autonomous agents ran 700 experiments in 2 days—and gave a glimpse of where AI is heading

In March 2026, Andrej Karpathy, a prominent AI researcher and former key figure at OpenAI and Tesla, captured widespread attention on social media platform X. His recent experiment involved autonomous AI agents that conducted a staggering 700 experiments over just two days, providing insights into the future trajectory of artificial intelligence.

Andrej Karpathy: A Brief Background

Andrej Karpathy has established himself as a leading voice in the AI community. With a substantial following of 1.9 million on X, his insights and experiments are often regarded with high esteem. Currently, he operates as an independent researcher and is the founder of Eureka Labs, an initiative aimed at developing educational frameworks suitable for the AI era.

The Experiment: Autoresearch

Karpathy’s recent experiment focused on enhancing the training of a small language model through an AI coding agent. This agent was allowed to run continuously for two days, during which it explored various optimizations. The results were impressive: the agent identified 20 distinct adjustments that improved the training time. When applied to a larger language model, these optimizations resulted in an 11% reduction in training time.

Key Features of Autoresearch

• Continuous Operation: The AI agent operated for 48 hours without interruption.
• Extensive Experimentation: Conducted 700 experiments to discover optimizations.
• Effective Results: Achieved a notable 11% speed increase in training larger models.

The Implications of Autoresearch

Karpathy’s autoresearch system is significant not only for its immediate results but also for its potential to reshape how AI research is conducted. He suggests that this methodology could become the norm in AI labs, accelerating the pace of innovation. He referred to this approach as the “final boss battle” in AI research, indicating its transformative potential.

Future Directions

Looking ahead, Karpathy envisions a future where multiple AI agents collaborate asynchronously to explore various optimizations in parallel. This collaborative model aims to simulate a research community rather than just a single researcher. He emphasized that any metric that can be efficiently evaluated could be optimized through this agent swarm, broadening the scope of problems that can be addressed using autoresearch.

The Karpathy Loop Explained

The concept of the “Karpathy Loop” has emerged from discussions surrounding this experiment. It consists of three main components:

1. An Agent: An AI agent capable of modifying a single file.
2. A Metric: An objective, testable metric that the agent aims to optimize.
3. A Time Limit: A fixed duration for each experiment, ensuring timely results.

This framework not only illustrates the mechanics of Karpathy’s autoresearch but also serves as a guide for effectively interacting with AI agents in various contexts.

Critiques and Comparisons

Despite the enthusiasm surrounding Karpathy’s work, some critics argue that he has merely rebranded existing methodologies, particularly AutoML (Automated Machine Learning). AutoML has been utilized by major tech companies like Google and Microsoft for years, employing optimization loops and experiments to enhance AI models.

Differences from AutoML

While AutoML focuses on random variations and evolutionary algorithms, Karpathy’s autoresearch leverages a more sophisticated AI agent capable of reading research papers and formulating hypotheses for improvements. Karpathy defended his approach, stating that traditional AutoML methods, such as neural architecture search, lack the power and flexibility of his autoresearch system.

Conclusion

Karpathy’s experiment with autoresearch not only showcases the potential of autonomous AI agents to optimize processes but also raises important questions about the future of AI research. As these systems evolve, they may redefine the boundaries of what is possible in AI development. The implications of the Karpathy Loop extend beyond mere experimentation; they suggest a future where AI systems can autonomously enhance their capabilities, potentially leading to significant advancements in the field.

Note: The developments in AI are rapid and ongoing. It is crucial to stay informed about these advancements and their implications for society.