Today’s systems are filled with diversifying applications. Theoretically, application performance could be optimized by leveraging target specific software frameworks and domain specific hardware. However, these attempts rely on domain expertise heavily: to start with, developers must get familiar with heterogeneous hardware architectures, compiler toolchains, and programming models. Moreover, they need to go through a tedious trial-and-error process and develop optimization insights based on empirical observations, which predicated on an intricate interaction between the applications and expected inputs. Those insights are then manually converted into optimization guidelines or hard-coded heuristics, which provides little guarantee on actual performance, and could not adapt to upcoming applications or hardware. We are pursuing a vision of learned optimizations, where we could automate much of this workflow. The learned strategies should specialize for a particular scenario in order to perform well, and these specializations should generalize for unseen scenarios that may arise. We aim to re-architect learned optimizations into system infrastructures, while providing strong security and interpretability guarantees. Specifically, We are exploring learned optimization for heterogeneous hardware (Clara) as well as diversifying applications (vision paper).