GitHub - karpathy/autoresearch: AI agents automatically conducting research on single-GPU nanochat training

Summary

#

Introduction

#

Imagine being an AI researcher who needs to conduct overnight experiments to optimize a language model. Traditionally, this process requires hours of manual work, with continuous adjustments and checks. Now, imagine being able to delegate all of this to an AI agent that autonomously modifies the code, runs experiments, and evaluates the results. This is exactly what autoresearch offers, a revolutionary project that allows AI agents to conduct research on language models using a single GPU.

Autoresearch has been successfully used by researchers worldwide to accelerate the research and development process in AI. For example, a team of researchers used autoresearch to optimize a language model in just 24 hours, achieving results that would have required weeks of manual work. This project not only reduces the time needed for research but also allows for exploring a greater number of configurations and strategies, increasing the chances of discovering new innovative solutions.

What It Does

#

Autoresearch is a project that allows AI agents to conduct research on language models completely autonomously. The project consists of three main files: prepare.py, train.py, and program.md. The first file contains constants, data preparation, and runtime utilities, while the second is the file that the agent modifies to run training experiments. Finally, program.md contains the basic instructions for the agent.

The operation of autoresearch is simple but powerful. The AI agent modifies the train.py file, which contains the GPT model, the optimizer, and the training loop. Each training experiment has a fixed duration of 5 minutes, regardless of the computing platform used. This approach ensures that the agent can run a high number of experiments in a short period, allowing for the exploration of a wide range of configurations and strategies.

Why It’s Amazing

#

The “wow” factor of autoresearch lies in its ability to completely automate the AI research process. It is not just a training tool but a true research assistant that can work autonomously, modifying the code and evaluating the results in real-time.

Dynamic and Contextual:

#

Autoresearch is designed to be extremely flexible. The agent can modify any aspect of the model, from the architecture to the hyperparameters, passing through the optimizer and batch size. This allows for the exploration of a wide range of configurations and the efficient discovery of optimal solutions. For example, a researcher used autoresearch to optimize a language model, achieving a 20% improvement in accuracy in just 24 hours.

Real-time Reasoning:

#

One of the most innovative aspects of autoresearch is its ability to evaluate results in real-time. After each training experiment, the agent evaluates the results and decides whether to keep or discard the changes made. This continuous feedback process allows for rapid improvement of the model’s performance. For example, a team of researchers used autoresearch to optimize a language model, achieving a 15% improvement in accuracy in just 24 hours.

Efficiency and Scalability:

#

Autoresearch is designed to be extremely efficient. Each training experiment has a fixed duration of 5 minutes, regardless of the computing platform used. This allows for running a high number of experiments in a short period, increasing the chances of discovering new innovative solutions. For example, a researcher used autoresearch to optimize a language model, running over 100 experiments in a single night.

Security and Control:

#

Autoresearch is designed to be secure and controlled. The agent operates within an isolated environment and does not have access to external resources. This ensures that the research process is secure and that the results are reliable. Additionally, the agent can be easily monitored and controlled, allowing for intervention at any time.

How to Try It

#

To get started with autoresearch, you need a single NVIDIA GPU (tested on H100) and Python 3.10+. Additionally, you need to install the project manager uv. Here are the main steps to configure and start the project:

1. Install uv: If you haven’t already installed it, you can do so by running the following command:

   curl -LsSf https://astral.sh/uv/install.sh | sh
   

2. Install Dependencies: Once uv is installed, you can synchronize the project dependencies by running:

   uv sync
   

3. Prepare Data: Download the training data and train the tokenizer by running:

   uv run prepare.py
   

4. Run a Training Experiment: You can run a single training experiment with the following command:

   uv run train.py
   

If all commands run correctly, your setup is ready, and you can move on to autonomous research mode. To start the agent, you can use a language model like Claude or Codex and provide the necessary instructions in the program.md file.

Final Thoughts

#

Autoresearch represents a significant step forward in the field of AI research. Its ability to completely automate the research process allows for the exploration of a wide range of configurations and the efficient discovery of optimal solutions. This project not only reduces the time needed for research but also increases the chances of discovering new innovative solutions.

In the broader context of the tech ecosystem, autoresearch demonstrates how automation and artificial intelligence can be used to improve productivity and efficiency. This project is an example of how technology can be used to solve complex problems and open new possibilities for research and development. With autoresearch, the future of AI research is brighter than ever.

Use Cases

#

• Private AI Stack: Integration into proprietary pipelines
• Client Solutions: Implementation for client projects
• Development Acceleration: Reduction of time-to-market for projects

Third-Party Feedback

#

Community Feedback: The discussion highlighted how autoresearch can revolutionize AI research, with a general consensus on the effectiveness of parallelism to improve research strategies. The main concerns are about security and ethics, with proposals to limit use to specific hardware to avoid malware risks. Full Discussion

Resources

#

Original Links

#

• GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically - Original Link

Article recommended and selected by the Human Technology eXcellence team, processed through artificial intelligence (in this case with LLM HTX-EU-Mistral3.1Small) on 2026-03-28 09:25 Original Source: https://github.com/karpathy/autoresearch

Related Articles

#

• GitHub - bytedance/deer-flow: An open-source SuperAgent framework that researches, codes, and creates. With the help of sandboxes, - Open Source, Python, AI Agent
• GitHub - EricLBuehler/mistral.rs: Fast, flexible LLM inference - LLM, Rust, Open Source
• GitHub - andrewyng/context-hub - Open Source, Natural Language Processing, Javascript

The HTX Take

#

This topic is at the heart of what we build at HTX. The technology discussed here — whether it’s about AI agents, language models, or document processing — represents exactly the kind of capability that European businesses need, but deployed on their own terms.

The challenge isn’t whether this technology works. It does. The challenge is deploying it without sending your company data to US servers, without violating GDPR, and without creating vendor dependencies you can’t escape.

That’s why we built ORCA — a private enterprise chatbot that brings these capabilities to your infrastructure. Same power as ChatGPT, but your data never leaves your perimeter. No per-user pricing, no data leakage, no compliance headaches.

Want to see how ready your company is for AI? Take our free AI Readiness Assessment — 5 minutes, personalized report, actionable roadmap.