We introduce the Discovery Engine, a general purpose automated system for scientific discovery. It combines deep learning with state-of-the-art interpretability techniques to identify complex, non-linear relationships in arbitrary datasets. This technology enables a shift from hypothesis-driven to data-driven discovery, and massively accelerates scientific enquiry by allowing a full exploration of the space of possible insights, free of bias and assumptions. It is hundreds of times faster than manual analysis, domain-agnostic, and data efficient – requiring only hundreds (rather than hundreds of thousands) of samples, and thereby making AI for science accessible in domains where data is limited or costly to obtain. In this paper we describe the Discovery Engine system and contrast it with contemporary approaches to AI for general scientific discovery (particularly LLM-driven methods, which form the bulk of alternative solutions).