Imagine effortlessly testing complex theories about real-world entities without painstakingly gathering data. That's the promise of a new research paper exploring how Large Language Models (LLMs) can simulate tabular datasets, allowing for rapid hypothesis exploration. Traditionally, testing a hypothesis like "Do horror writers have worse childhoods than other writers?" would require manual sifting through biographies and interviews to quantify relevant factors. This new research proposes using LLMs to shortcut this process. How does it work? The researchers propose two methods: LLM-driven and Hypothesis-driven Dataset Simulation. In the first, you give the LLM a list of entities (e.g., writers) and properties (e.g., number of adverse childhood experiences). The LLM then estimates the property values for each entity, creating a simulated dataset. The second method goes further: you give the LLM a high-level hypothesis, and it generates both the relevant properties and entities before simulating the data. The researchers tested these methods on diverse datasets, including animal characteristics, country demographics, and athlete performance. They found that LLMs can simulate datasets with reasonable accuracy, especially as model size increases. In the Countries dataset, for instance, larger models showed significantly higher correlation between simulated data and real-world values. Furthermore, simulating data and then analyzing it proved more effective than directly asking the LLM to estimate relationships between variables. The results are promising, though the method has limitations. The accuracy of simulated data depends on the LLM’s knowledge of the subject and its overall capabilities. Also, the current method requires human intervention to specify the hypothesis. However, the researchers suggest future development toward a fully automated system where the LLM generates and explores hypotheses independently. This research is a stepping stone toward more automated knowledge discovery. By simulating datasets on the fly, LLMs could unlock hidden patterns in vast amounts of existing data, accelerating scientific progress across various fields. Imagine AI systems constantly searching for and revealing unexpected connections in the ocean of human knowledge—this research gives us a glimpse into that exciting future.