Urithiru Tests Whether an AI Scientist Can Start With Data Instead of a Goal

Summary

Urithiru is an open-source agent designed to explore scientific hypotheses from existing tabular datasets rather than beginning with a predefined research goal. Its creator spent six weeks building and testing the system, which grows a hypothesis tree using Monte Carlo tree search and Bayesian surprise: branches receive more analysis when new evidence changes the model’s literature-grounded beliefs. Each branch uses separate agents for exploratory data analysis, literature review, code-based experiments, and optional checks against an independent dataset. The system records several belief stages, including a model-specific prior, a literature-based estimate, a conclusion from executed code, and an external-data result. In one example about vigorous exercise, BMI, and long-term blood sugar, the initial belief was 0.77, literature reduced it to 0.35, and analysis of 5,141 NHANES participants raised the code-based estimate to 0.94 after finding a 38% lower BMI–HbA1c slope among vigorously active adults. A separate Mexican health survey found virtually identical slopes for active and inactive adults, lowering the external estimate to 0.11 and illustrating how disagreement becomes a search signal. Several experiments exposed limitations. A Markdown research wiki created more context but performed worse as model-facing memory because synthesis removed statistical qualifications and added polished but distracting text. Linking datasets to grant narratives changed proposal wording but not executable proposal rates: 269 of 325 with the grant versus 267 of 325 without it. External-dataset searches worked for almost half of frozen claims, while other claims failed to replicate or lacked the needed measurements. The project is open source, supports local Docker and Google Cloud execution, checkpointing, resumption, and a small tree-inspection interface. The article presents dataset-first discovery as an exploratory aid and a way to test results before collecting new data, not as a replacement for scientific objectives or a final evidentiary verdict.