I'll present recent developments in inferring phylogenies from sequences by maximum likelihood (ML). In fact, these approaches involve two main components: probabilistic modelling of substitution events, and algorithmic to infer near-optimal trees. I'll show that algorithmic has considerably progressed during the last few years, allowing now for fast ML inference from large datasets. On the other hand, I'll underline the difficulties encountered with modelling, especially of protein evolution, and show that elaborating more realistic and improved substitution models is a major challenge.