Recent developments in machine learning have highlighted the potential of Kolmogorov-Arnold Networks (KAN) as a more expressive alternative to traditional multilayer perceptrons (MLPs). A key paper on KANs discusses how these networks can allow for more expressive internal activations, enabling embedding vectors to convey richer meanings. Additionally, the concept of DropKAN has been introduced as a regularization method that masks post-activations to prevent co-adaptation of activation function weights in KANs. This approach aims to enhance parameter efficiency and generalization in large-scale models. Furthermore, a study on differentially private KANs explores their application in private model training using the DP-SGD algorithm. Recordings of a recent London Machine Learning Meetup, featuring a talk by Ziming Liu on KANs, are now available on YouTube, providing further insights into these advancements.