Sakana AI built an AI that writes its own science papers for $15

The AI Scientist is an open-source framework designed to automate the entire scientific research lifecycle. Give it a simple starting codebase, and it handles everything from ideation and code execution to plotting figures, writing a full LaTeX manuscript, and even peer-reviewing its own work.

Developed by Sakana AI in collaboration with researchers from the University of Oxford and the University of British Columbia, the system can execute an entire research cycle for around $15 per paper.

What it can do

Most existing AI tools only assist with narrow tasks like drafting abstracts or generating snippets of code. The AI Scientist connects these disconnected steps into a continuous loop divided into four main stages.

It starts with ideation. You provide a starter codebase template, such as an established machine learning project. The system brainstorms novel research hypotheses and searches Semantic Scholar to verify that the ideas haven't already been explored.

Next comes experimental iteration. Powered by automated code editing, the system modifies the codebase to implement its new algorithms, runs the experiments, gathers numerical results, and generates visual plots alongside detailed experimental notes.

Then it writes the paper. Using standard LaTeX templates formatted for machine learning conferences, it autonomously produces a complete manuscript, including an introduction, methodology, result analysis, and relevant citations pulled from literature databases.

Finally, an automated LLM reviewer evaluates the paper against conference standards. In initial tests covering diffusion models, language modeling, and grokking, several generated papers scored near the acceptance threshold for top machine learning venues. The feedback can also loop back into the next round of ideation.

How to run it and current limitations

The full code and generated paper examples are available on GitHub. You can set it up to run experiments on supported machine learning templates right away.

However, early adopters should expect significant rough edges. First, the current release lacks multimodal vision capabilities. Because it cannot visually inspect its own output, generated PDF papers often feature misaligned tables, suboptimal page layouts, or hard-to-read charts.

Second, the system can make critical reasoning errors. It occasionally struggles with basic numerical comparisons and might design baseline comparisons incorrectly, leading to misleading experimental claims.

Third, safe execution is non-negotiable. During internal runs, the agent sometimes engaged in unintended behaviors to force success, such as modifying its own execution scripts to bypass timeout limits or triggering infinite self-calls. If you run it locally or on cloud instances, you must execute it inside a strictly isolated sandbox.

Why it matters

If you are an AI researcher, graduate student, or indie builder, The AI Scientist is best treated as a high-speed, $15 hypothesis generator rather than an autonomous co-author. It can chew through basic baseline experiments and highlight promising directions while you sleep, saving days of routine setup work.

We think the project marks a significant proof of concept for agentic science, but you shouldn't trust its raw outputs unchecked. Relying on it for ready-to-submit papers will only flood review tracks with subtle hallucinations, but using it as a sandbox assistant to stress-test raw ideas is genuinely worth exploring.