Can an AI Scientist Start with a Dataset Instead of a Goal? (eamag.me)

A new open-source AI project named Urithiru has been launched, challenging conventional scientific inquiry methods by starting with datasets rather than predefined research goals. Developed over six weeks, Urithiru aims to stimulate hypothesis generation by exploring existing tabular data, fostering a unique approach to scientific discovery. This method contrasts with standard research practices, which typically require identifying important research questions and collecting new data to answer them. By leveraging the existing datasets, Urithiru proposes that new insights and hypotheses can emerge without the pressure of initial experimental design. Significantly, Urithiru employs a Monte Carlo Tree Search (MCTS) strategy similar to that seen in AlphaGo, utilizing Bayesian surprise metrics to prioritize hypotheses testing based on their potential to alter existing beliefs in the literature. The system features separate agents that examine data, generate and execute hypotheses, and document findings—offering a structured environment for continuous scientific evaluation. This approach not only expands the utility of existing datasets but also demonstrates the potential for AI agents to facilitate research without needing clear objectives from the outset, thus allowing for a more explorative research methodology that could benefit the wider AI/ML community.