Sakana Labs has just made a big splash in AI. They introduced the world's first AI scientist. This system can automate many parts of scientific research. Founded by Leon Jones and David Ha, Sakana Labs is a Tokyo-based startup. They focus on creating AI models inspired by nature. They use ideas from schools of fish and beehives to make their AI flexible and efficient.
Leon Jones and David Ha have impressive backgrounds. Leon co-authored the famous 2017 paper "Attention is All You Need." This paper introduced the Transformer architecture. David Ha led research at Stability AI and Google Brain. Together, they raised $30 million in seed funding for Sakana Labs.
So, what does the AI scientist do? It starts by brainstorming a set of ideas. It checks how new these ideas are. Then, it writes code to test the new ideas. Next, it runs experiments to gather results. It crafts a scientific report explaining the results. Finally, it generates an automated peer review. This helps refine the research and guides future projects.
The AI scientist follows four key steps. First, it generates ideas. Second, it executes experiments and visualizes the results. Third, it writes up the findings in a clear report. Fourth, it uses an automated reviewer to evaluate the paper. This process creates a feedback loop to improve the quality of research over time.
This system has some fascinating aspects. It can produce research papers for just $15 each. This could democratize research and speed up scientific progress. Imagine if these papers cost just a dollar or even a few cents. We could automate many areas of research.
But are the papers any good? The AI scientist has generated papers with some novel insights. However, the quality is mixed. Some papers present new ideas and approaches. For example, a paper on diffusion models showed improvements in sample quality. Another paper proposed a new architecture for diffusion models. But overall, the papers are like early-stage work. They need human experts to develop them further.
A key model used in this research is Claude Sonet 3.5. It consistently produces the best papers. Some even reach the acceptance threshold for top machine learning conferences. Sakana Labs believes that all advanced models, including open ones, will keep improving. The competition among models will lead to better capabilities.
One limitation is that the AI scientist can't fix visual issues. It can't read plots or adjust layout problems. Adding multimodal models could fix this. The AI also sometimes makes critical errors. For example, it may struggle to compare numbers accurately.
Sakana Labs has open-sourced the AI scientist. You can find all the details on their GitHub repository. This means anyone can use the system to conduct their own research. It's an exciting development in the world of AI. With advancements like these, we never know what's coming next.