Neural networks are becoming increasingly ubiquitous in a wide range of use cases. A primary hurdle in deploying neural networks in many scenarios is the tedious and difficult neural network architectural design process, which was reliant on expert knowledge and iterative design. Neural Architecture Search (NAS) reduces the human effort required for design, but still has considerable resource requirements and is extremely slow. To address the inefficiencies of conventional NAS, Zero-Shot NAS is a new paradigm, which introduces zero shot neural architecture scoring metrics (NASMs) to identify good neural network designs without training them. While applying Zero Shot NASMs is cheap and requires no training resources, we identify that there is a lack of NASMs that generalize well across neural architecture design spaces. In this paper, we present a program representation for NASMs and automate its search with genetic programming. We discover effective NASMs for Image Classification as well as Automatic Speech Recognition. We believe that our work indicates a new direction for NASM design and can greatly benefit from recent advances in program synthesis.