The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise

This project is a high-quality research contribution addressing a critical 'black box' in modern LLM architecture (MoEs). It challenges the industry-standard narrative that experts in a Mixture-of-Experts model specialize in semantic domains (e.g., 'the math expert'). Instead, it provides empirical evidence that routing is a byproduct of high-dimensional geometry. While the insight is significant for model architects at firms like Mistral, OpenAI, or DeepSeek, it lacks any traditional software moat. The defensibility is low (2) because the value is the insight itself, which is freely available via the paper; once the community absorbs the 'geometry-first' understanding of MoEs, the specific code implementation becomes a historical reference. The 3 forks within 3 days of release, despite 0 stars, indicates immediate interest from the research community (likely peer researchers or early adopters in LLM training). Frontier labs are unlikely to 'compete' with this but will likely internalize these findings to optimize their own routing algorithms, making the 'frontier risk' medium—not because they will build this tool, but because they will solve the problem it identifies at the training level. Platform domination risk is low as this is an academic/interpretability contribution rather than a service-based product.