Researchers built an ‘AI Scientist’ — what can it do?
With the introduction of the "AI Scientist" — a ground-breaking tool created in collaboration by researchers from Canada, the United Kingdom, and a team from Sakana AI in Tokyo— the goal of completely automating scientific discovery is getting closer to reality. It can handle everything from literature reviews to experimental testing, hypothesis development, and paper authoring. Therefore, this sophisticated system marks a major advancement in integrating artificial intelligence into the scientific process.
The AI Scientist is able to carry out extensive research cycles because of its advanced large language model (LLM). It starts by going over the current body of research to find gaps and develop theories. Then the system picks and produces academic papers by iteratively testing different types of machine learning algorithms, employing evolutionary computation—a technique influenced by Darwinian principles of natural selection. Amazingly, it can even conduct peer assessments of its own work automatically.
This ambitious system has been hailed as a turning point in the automation of research. “Nobody has yet achieved a fully integrated scientific system like this,” says Cong Lu, a researcher from the University of British Columbia who helped create the AI Scientist. The work recently generated interest and discussion in the scientific community when it was posted on the arXiv preprint server.
AI generated image of a possible future ’AI Scientist’
The AI Scientist has made amazing progress, but there aren't many uses for it yet. As of right now, the robot is limited to machine learning and is unable to conduct practical laboratory work, which is an essential part of conventional scientific methodologies. Moreover, the system still falls short regarding thorough scientific inquiry, according to Gerbrand Ceder, material scientist at Lawrence Berkeley National Laboratory. Ceder is nevertheless hopeful about the future and thinks that these kinds of technologies will be essential to the growth of science.
Despite being revolutionary, the work of the AI Scientist has come under fire. Critics have noted that the articles published mostly only show little steps forward and are skewed by "popularity bias," which favors well-cited studies over less well-known but potentially revolutionary ones. Several scientists have voiced doubts about the system's potential to make a substantial contribution to science, pointing out the value of unofficial information sharing and the complex understanding that results from interpersonal communication.
Computational social scientist Jevin West of the University of Washington argues that science is more than merely reading academic journals. "A quick chat can provide you with insights that are far more useful than spending hours reading literature," he claims. Despite this, West and his colleague Shahan Memon applaud the group for being transparent enough to let others evaluate and draw conclusions from the AI Scientist's findings.
Automating parts of scientific inquiry is not a novel idea. Since the early days of artificial intelligence, efforts have been made to automate experiments and expedite data analysis. Robotic chemists that can synthesize materials and the Automatic Statistician, which can evaluate data and generate reports, are only two examples of the tools that use AI to revolutionize research procedures.
While existing LLMs are not currently good at developing new scientific paths, Tom Hope of the Allen Institute for AI notes that they are excellent for automating repetitive work. While AI may not be able to develop creatively just now, he claims that it is capable of handling a sizable amount of mundane research labor. Cong Lu agrees, characterizing the AI Scientist as the “GPT-1 of scientific research” with enormous potential for advancement in the future.
A more thorough consideration of what qualifies as scientific research in the twenty-first century is prompted by the way artificial intelligence is being used in science. In closing, West says, "AI challenges us to rethink the essence of science and its future direction."
The process of incorporating AI into science is still in its early stages, but the AI Scientist is a positive first step toward a time when human creativity and artificial intelligence will work together to expand human understanding.
Sources:
Lu, Cong, et al. *Preprint at arXiv*. 2024, https://arxiv.org/abs/2408.06292.
Ghahramani, Zoubin. "Probabilistic Machine Learning and Artificial Intelligence." *Nature*, vol. 521, no. 7553, 2015, pp. 452–459.
Szymanski, Nicholas J., et al. "Robot Chemist Sparks Row with Claim It Created New Materials." *Nature*, vol. 624, no. 7980, 2023, pp. 86–91.