In this talk, I will consider the problem of learning a predictor from multiple sources of information, a situation common in many domains such as computer vision or bioinformatics. I will focus primarily on the multiple kernel learning framework, which amounts to consider one positive definite kernel for each source of information. Natural unanswered questions arise in this context, namely: Can one learn from infinitely many sources? Should one prefer closely related sources, or very different sources? Is it worth considering a large kernel-induced feature space as multiple sources?