In the computational literature intrinsic motivations have been connected to the possibility of developing more autonomous and versatile agents. Despite the growing theoretical understanding of the distinction between functions and mechanisms of intrinsic motivations, the implications of the distinction have not been exploited in specific models. In particular, knowledge-based mechanisms are widely used to implement intrinsic motivations signals for the acquisition of competences, leading to inappropriate learning signals. In this paper we analyse and compare, with the support of simple grid-world simulations, different mechanisms that can be used to implement competence acquisition through intrinsic motivations, describing their limits and strengths and highlighting which features are best suited for the acquisition of competence.