The AI Scientist
VerifiedSakana AI's open-source agent that runs ML experiments and writes research papers
Product preview
The AI Scientist
What it is
The AI Scientist is a research automation system released by Tokyo-based Sakana AI in August 2024, built with collaborators at the University of Oxford and the University of British Columbia. Given a starting template, it uses large language models to brainstorm ideas, check novelty, write and execute experiment code, plot results, and write a paper in LaTeX, followed by an automated peer review step. The first version relies on templates such as NanoGPT, 2D diffusion and grokking. The AI Scientist-v2 drops the template requirement and uses an agentic tree search guided by an experiment manager. Both versions are on GitHub, need Linux with NVIDIA GPUs, and use API keys for supported models.
Key features
- Automated idea generation and novelty checks
- Writes and runs experiment code
- Drafts full LaTeX papers
- Automated reviewer for generated papers
What we liked
- + Open source on GitHub
- + Covers the full research loop
- + Well documented with example papers
Worth noting
- · Requires GPUs and paid model API usage
- · Runs model-written code, so sandboxing is needed
Editor's notes
More context from AI BirdView
The AI Scientist is one of the most ambitious open research agents available. Sakana AI built it to handle the entire cycle of a machine learning study, from forming an idea to running experiments and writing a paper, and the team shares example outputs so you can judge the quality for yourself.
- Generates research ideas and checks them for novelty against existing literature
- Writes experiment code, runs it and produces plots from the results
- Drafts complete papers in LaTeX and scores them with an automated reviewer
- Version 2 explores open-ended problems with an agentic tree search instead of fixed templates
- Sakana AI says a paper produced by v2 passed peer review at an ICLR 2025 workshop
The AI Scientist is best for machine learning researchers, labs and advanced students who want to study or extend automated discovery systems. A practical tip: follow the project's own advice and run it inside a sandbox such as a Docker container, since it executes code written by language models, and start with one of the provided templates before attempting open-ended runs.
Reader reviews
From people who've used The AI Scientist. Emails are never shown publicly.
No reviews yet. Be the first to share your experience.
Alternatives to The AI Scientist
Other tools we like in this space.
Elicit
Automate the grind of literature review
SciSpace
Read dense papers with an AI copilot beside you
Undermind
Deep scientific search that reads like a PhD colleague