For the past few years, I have been part of a collaboration with top-tier mathematicians on a challenging project: teaching machines to assist humans with proving difficult theorems and conjecturing new approaches to long-standing open problems. My team has demonstrated that analysing and interpreting the outputs of (graph) neural networks is a concrete solution. Our efforts independently derived novel top-tier mathematical results in areas as diverse as representation theory and knot theory. The significance of our findings was recognised by the journal Nature, where it featured on the cover page. I will share our findings from a personal perspective, with key details of our team’s modelling work.