The AI Scientist is a system designed for fully automated scientific discovery, enabling foundation models like LLMs to conduct research independently. It targets researchers and developers seeking to automate knowledge generation, offering a framework for LLMs to brainstorm, experiment, and write scientific papers across various domains.

How It Works

The system utilizes a templated approach, where each template defines a specific scientific domain (e.g., NanoGPT, 2D Diffusion, Grokking). Within a template, the AI Scientist generates research ideas, writes code to execute experiments, analyzes results, and compiles findings into a scientific paper. This modular design allows for extensibility with community-contributed templates and supports a wide range of LLMs for different stages of the research process.

Quick Start & Requirements

• Install: Use conda to create an environment and pip install -r requirements.txt. Requires texlive-full for PDF generation.
• Prerequisites: Linux, NVIDIA GPUs with CUDA, PyTorch. CPU-only machines are infeasible.
• LLM API Keys: Requires API keys for supported models (OpenAI, Anthropic, DeepSeek, Google Gemini, etc.).
• Setup Time: Installing texlive-full can be lengthy. Template setup involves cloning repositories and installing dependencies.
• Docs: Paper, Blog Post, Drive Folder (link placeholder).

Highlighted Details

• Supports a wide array of LLMs, recommending frontier models like GPT-4o and Claude 3.5 Sonnet.
• Includes three core templates: NanoGPT, 2D Diffusion, and Grokking, with community contributions welcomed.
• Offers functionality for reviewing generated papers using LLMs.
• Provides a Dockerfile for containerized execution.

Maintenance & Community

The project is associated with SakanaAI. Community-contributed templates are listed, with links to pull requests.

Licensing & Compatibility

The repository's license is not explicitly stated in the README. Compatibility for commercial use or closed-source linking is not detailed.

Limitations & Caveats

The codebase executes LLM-generated code, posing risks related to package usage, web access, and process spawning. Containerization and restricted web access are strongly advised. The success rate of idea generation and paper completion varies by LLM and template complexity. Support for non-Linux OS and non-NVIDIA GPUs may require significant adjustments.