Self-optimizing interpreters are a promising technique to build languages based on simple abstract-syntax-tree (AST) interpreters. The Truffle framework provides a convenient way to build a wide range of dynamic languages. However, in the combination with Graal it relies on partial evaluation as meta-compilation technique, which is only one of the possible approaches to generate efficient native code. Trace-based compilation as used by the PyPy/RPython project is an alternative compilation approach. This talk discusses the practical differences between both approaches based on a self-optimizing interpreter implemented on top of Truffle as well as RPython.