Machine learning is playing an increasingly large role in much of science. In high energy physics it has already revolutionized many aspects of experimental and theoretical collider physics. Collider physics is well suited to machine learning partly because the problems are largely numerical. However, much of high energy theory is largely symbolic. Can machine learning help in these areas too? Inspired by the impressive success of large language models in recent years, progress on symbolic problems should be possible. In this talk I will review some elements of symbolic learning, including reinforcement learning and transformer networks, and discuss applications to the field of scattering amplitudes. In particular, I will discuss how the task of simplifying a scattering amplitude to a simpler symbolic form.