Despite the vast amount of material about the MLIR project, misconceptions regarding its scope and implementation abound. This talk will clarify such misconceptions highlighting that MLIR is not in fact a compiler, let alone a machine learning system; that dialects are intended to be mixed together despite the misleading name; and that there is fundamentally no single optimization and lowering pass pipeline. This talk addresses audience members with varying degrees of familiarity with MLIR, from neophytes who would get a better conceptual understanding of the project to advanced long-term users and contributors who could question or additionally justify some of the design choices that led to these misconceptions.