A team from Sakana AI, a Tokyo-based lab founded by former Google researchers David Ha and Llion Jones, has built something that actually works like a real scientist. Their system, called "The AI Scientist," takes a research topic and runs with it. It generates ideas, writes the code, runs experiments, plots results, writes a full paper, then reviews what it produced. One of those AI-written papers passed the first round of peer review at a NeurIPS workshop with a 70% acceptance rate.
The system works in two modes. Focused mode starts from human-provided code templates to explore specific topics. Template-free mode lets the agent roam wider, using search to find its own direction. Both produce complete papers. The team also built an automated reviewer that performs about as well as human reviewers at predicting conference acceptances. That let them test different configurations at scale. Better base models and more compute both improve output quality, which suggests future versions will keep getting better.
This isn't just a demo. The research landed in Nature. Sakana has raised over $130 million from Lux Capital and NEA, hitting unicorn status. Llion Jones co-authored the "Attention Is All You Need" paper that gave us Transformers, so the team has real credibility. The obvious concern is what happens when these systems flood already strained infrastructure with generated papers. The authors argue that responsible development could accelerate discovery instead of just adding noise.