shiyudongx/MoE-KAN

MoE-KAN is a project attempting to merge two high-interest research topics: Kolmogorov-Arnold Networks (KANs), which use learnable spline-based activation functions on edges rather than nodes, and Mixture-of-Experts (MoE), a technique for sparse model scaling. As of the current assessment, the repository has zero stars and forks, indicating it is in a very early 'preprint-adjacent' or personal experimentation phase. While the combination is conceptually novel—applying MoE to scale the computationally expensive B-splines of KANs—the project lacks any defensive moat. It is a pure implementation play that could be easily superseded by more established KAN libraries (like the original pykan or optimized versions like fast-kan) if they decide to implement MoE layers. The risk from frontier labs is low because KANs currently target niche scientific discovery and symbolic regression tasks rather than the large-scale linguistic reasoning labs like OpenAI or Anthropic focus on. However, its value is entirely dependent on its ability to prove better scaling laws for KANs, which is currently unproven in this repository.