The field of Machine Learning is witnessing a fundamental re-evaluation of neural network architectures with the emergence of Kolmogorov-Arnold Networks (KANs) as a promising alternative to the long-dominant Multi-Layer Perceptrons (MLPs). Inspired by the Kolmogorov-Arnold representation theorem, KANs propose a significant theoretical departure: instead of fixed activation functions on nodes ("neurons"), KANs feature learnable activation functions on edges ("weights"), which are parametrized as splines. This means KANs do not use linear weight matrices at all; every weight parameter is replaced by a univariate function. This architectural innovation is reported to offer substantial advantages, demonstrating superior accuracy and interpretability compared to MLPs on small-scale AI + Science tasks, exhibiting faster neural scaling laws. The ability of KANs to be intuitively visualized is particularly noteworthy, as it facilitates human understanding of the model's decision-making process, enabling KANs to function as "collaborators" with scientists in the (re)discovery of mathematical and physical laws. This development suggests a growing emphasis in ML research on not just raw performance but also on transparency, understanding, and the active facilitation of scientific discovery.