The AI Scientist
Self-hosted AI research agent that runs experiments and writes papers on Linux with NVIDIA GPUs, using models such as GPT-4o and Claude.
The AI Scientist is a self-hosted AI agent for researchers studying questions they can test through code. It generates research ideas, runs experiments, produces plots and writes scientific papers. Its scope is computational research.
The supplied research templates cover NanoGPT language modeling, low-dimensional diffusion models and grokking, which studies neural network generalization and learning speed. Researchers can create templates for other subjects. Community contributions include infectious disease modeling, MobileNetV3 image classification and quantum chemistry, though the project maintainers don't maintain those templates.
The software runs on your own Linux machine or server with NVIDIA GPUs, CUDA and PyTorch. The supplied experiments aren't practical on CPU-only hardware. It can distribute separate research ideas across multiple GPUs.
Model support includes GPT-4o and o1 through OpenAI, Claude through Anthropic, Amazon Bedrock or Vertex AI, plus DeepSeek, Gemini and Llama through OpenRouter. Those API connections require internet access and provider credentials, and send model requests to external services even though experiments run on your hardware. Semantic Scholar supplies literature searches, novelty checks and citations; those phases can be skipped.
The paper review component reads PDFs and returns scores, weaknesses and acceptance recommendations. Successful paper generation depends on the model, template and complexity of the idea. The custom AI Scientist Source Code License requires prominent disclosure of AI use in resulting scientific manuscripts.
Logotyper
