Researchers have developed a graphical notation system for designing interpretable AI architectures. The system, inspired by Penrose tensor notation, provides a global view of an architecture and can be directly translated into code using the PyTorch einsum function. This allows for easier comparison and understanding of different architectures, including those that are inherently interpretable such as concept bottlenecks and sparse probes. The system has been applied to Steerling-8B, a frontier interpretable language model, resulting in a 33-line PyTorch code translation. --- Why it matters: This matters because it provides a standardized way for researchers to design and compare AI architectures, making it easier to identify key components and understand how they interact. This can lead to more efficient development of new models and improved reproducibility in the field. Source: https://arxiv.org/abs/2608.18936