To meet the diverse demands for wireless communication, fifth-generation (5G) networks and beyond (B5G) embrace the concept of network slicing by forging virtual instances (slices) of its physical infrastructure. While network slicing constitutes dynamic allocation of core network and radio access network (RAN) resources, this article emphasizes RAN slicing (RAN-S) design. Forming on-demand RAN-S that can be flexibly (re)-configured while ensuring slice isolation is challenging. A variety of machine learning (ML) techniques have been recently introduced for traffic forecasting and classification, resource usage prediction, admission control, scheduling, and dynamic resource allocation in RAN-S. Albeit these approaches grant opportunities towards intelligent RAN-S design, they raise critical challenges that need to be examined. This article underlines the opportunities and the challenges of incorporating ML into RAN-S by reviewing the cutting-edge ML-based techniques for RAN-S. It also draws few directions for future research towards intelligent RAN-S (iRAN-S).