Summarizing https://arxiv.org/pdf/1906.01820.pdf

Here's my try:

This paper introduces the concept of mesa-optimization, which refers to machine learning models that are also optimizers. The authors identify two key problems related to mesa-optimization: unintended optimization and inner alignment. They introduce a distinction between the two different, independent alignment problems that appear in the case of mesa-optimization. Mesa-optimizers need not be robustly aligned with the base optimizer that created them, only pseudo-aligned. The paper seeks to understand what sorts of machine learning systems are likely to exhibit mesa-optimization and what sorts are not. Furthermore, the analysis suggests that a time complexity penalty (as opposed to a description length penalty) is a double-edged sword. In section 2, we suggested that penalizing time complexity might serve to reduce the likelihood of mesa-optimization. However, the above suggests that doing so would also promote pseudo-alignment in those cases where mesa-optimizers do arise. If the cost of fully modeling the base objective in the mesa-optimizer is large, then a pseudo-aligned mesa-optimizer might be preferred simply because it reduces time complexity, even if it would result in suboptimal plans for the base objective.