The AI Scientist

The AI Scientist is an autonomous research agent developed by Sakana AI in collaboration with FLAIR (University of Oxford) and the University of British Columbia, released in August 2024. It is the first system to perform the entire scientific research cycle in machine learning without any human involvement — generating novel research ideas, designing and running experiments in code, analyzing results, producing visualizations, writing full LaTeX papers, and performing automated NeurIPS-style peer review.

The system uses frontier LLMs (Claude 3.5 Sonnet, GPT-4o, DeepSeek, o1) as its reasoning engine and integrates with the Aider coding agent to iteratively modify experiment code. Each fully autonomous paper costs roughly 15 USD. In January 2025 the team released The AI Scientist-v2, whose paper was accepted at an ICLR 2025 workshop — the first peer-reviewed paper fully authored by AI in history.

How it works

• Ideation: an LLM brainstorms research ideas from a seed template and queries Semantic Scholar to filter out duplicates.
• Experiments: the Aider coding agent modifies the codebase, runs training/evaluation, collects results, and iterates on hypotheses.
• Writing: the system creates matplotlib plots, narrates the results, and compiles a full LaTeX paper with references.
• Automated Reviewer: a second LLM scores the paper against the NeurIPS reviewer form, reaching near-human agreement with real reviewers.