I will describe a new approach that we are currently developing to describe developmental dynamics. We are using simple machine learning techniques to project the dynamics of the Drosophila gap genes onto a low dimensional space, allowing us to build "geometric" models. We uncover a relatively simple dynamics in latent space, where we identify two modes, which allows us to capture with very good precision the entire gap genes dynamics. We interpret the dominating mode as a "maternal" contribution, while another mode accounts for "networks" effects. I will also discuss implication for positional information encoding, and interpretation of mutants.