[2009.11243] Tasks, stability, architecture, and compute: Training more effective learned optimizers, and using them to train themselves

https://arxiv.org/abs/2009.11243