GitHub - Search code, repositories, users, issues, pull requests...: Apple Silicon (MLX) port of Karpathy's autoresearch — autonomous AI research loops on Mac, no PyTorch

Summary

#

Introduction

#

Innovation in the field of artificial intelligence and machine learning continues to amaze, and a recent example is the “Apple Silicon (MLX) port of Karpathy’s autoresearch” project. This project, available on GitHub, represents an important evolution in autonomous research in the AI field, allowing autonomous research loops to be run directly on Apple Silicon machines, without the need for PyTorch. This tool was shared on X with the intention of highlighting how new hardware architectures can significantly influence AI research results, offering new opportunities to optimize training processes.

The project was developed to make the most of the capabilities of Apple Silicon CPUs, offering an effective alternative to PyTorch and CUDA-based solutions. This makes the project particularly interesting for developers and researchers working on Apple machines, allowing them to run machine learning experiments more efficiently and autonomously.

What It Offers / What It Is About

#

The “Apple Silicon (MLX) port of Karpathy’s autoresearch” project is a port of Andrej Karpathy’s original autoresearch, adapted to run natively on Apple Silicon hardware. This means that there is no need to use PyTorch or CUDA, making the training process simpler and more accessible. The project maintains the same basic rules as the original: a modifiable train.py file, an evaluation metric (val_bpb), a fixed training budget, and a keep-or-revert mechanism managed via Git.

The project includes several key components:

• prepare.py: Manages data preparation, tokenizer, dataloader, and evaluation. This file should be considered fixed.
• train.py: Contains the model, optimizer, and training loop. This is the file that the agent modifies.
• program.md: Describes the protocol of the autonomous experiment.
• results.tsv: Records the history of experiments.

The autonomous research loop works by modifying train.py, running an experiment with a fixed time budget, reading the val_bpb metric, keeping the changes if they improve the results, and reverting them otherwise. This process repeats continuously, allowing the model to be optimized autonomously.

Why It’s Relevant

#

Hardware-Specific Innovation

#

The project was shared on X to highlight how new hardware architectures can influence AI research results. In particular, the porting to Apple Silicon has shown that smaller and faster models can outperform larger ones simply because they can execute more optimization steps within the available time budget. This is a concrete example of how hardware can influence AI model design choices.

Efficiency and Accessibility

#

The absence of dependencies on PyTorch and CUDA makes the project particularly interesting for developers working on Apple machines. This allows machine learning experiments to be run more efficiently and autonomously, without the need for complex configurations or specialized hardware. Additionally, the project offers a practical example of how hardware-specific optimization can lead to significant improvements in AI research results.

Usage Context

#

The project is useful for researchers and developers who want to explore new hardware architectures and optimize their AI models autonomously. The discoveries made with this project can be applied in various contexts, such as improving the performance of machine learning models, optimizing training processes, and researching new hardware-specific solutions.

How to Use It / Deep Dive

#

To start using this project, you need a Mac with an Apple Silicon CPU and Python installed. Follow these steps:

1. Install dependencies: Use uv to install the necessary dependencies. If uv is not already installed, you can install it with the command curl -LsSf [URL] | sh.
2. Prepare data: Run uv run prepare.py to prepare the data and tokenizer.
3. Run experiments: Run uv run train.py to start a one-minute training experiment.
4. Automation: Point a coding agent like Claude Code to program.md and let it manage the autonomous research loop.

To deepen your understanding, you can consult the GitHub repository and explore the configuration files and experiment results. Additionally, you can consult the official uv documentation and other related resources for more information.

Final Thoughts

#

The “Apple Silicon (MLX) port of Karpathy’s autoresearch” project fits into a broader context of hardware-specific innovation in the field of machine learning. The discoveries made with this project highlight the importance of adapting AI models to the specific capabilities of the hardware, allowing for significant improvements in performance. This approach can be applied in various contexts, such as optimizing training processes and researching new hardware-specific solutions. Additionally, the project represents a concrete example of how hardware innovation can influence AI model design choices, opening new opportunities for research and development in the field of machine learning.

Use Cases

#

• Private AI Stack: Integration into proprietary pipelines
• Client Solutions: Implementation for client projects

Resources

#

Original Links

#

• GitHub - Search code, repositories, users, issues, pull requests…: Apple Silicon (MLX) port of Karpathy’s autoresearch — autonomous AI research loops on Mac, no PyTorc - Main content (Github)- Original X post - Post that shared the content

Article reported and selected by the Human Technology eXcellence team, processed via artificial intelligence (in this case with LLM HTX-EU-Mistral3.1Small) on 2026-03-23 08:50 Original source: https://x.com/trevinpeterson/status/2030611877198221458?s=43&t=ANuJI-IuN5rdsaLueycEbA

Related Articles

#

• GitHub - karpathy/autoresearch: AI agents automatically conducting research on single-GPU nanochat training - AI Agent, Python, Open Source
• MicroGPT is a compact, open-source language model designed for efficient text generation and understanding. It is built to be lightweight and can run on a variety of devices, including personal computers and even some mobile devices. MicroGPT is intended for tasks such as text completion, summarization, translation, and more, making it a versatile tool for developers and researchers working with natural language processing. - Tech
• GitHub - z-lab/paroquant: [ICLR 2026] ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning in Large Language Model Inference - AI, LLM, Machine Learning

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.