Machine learning and data mining, as a whole, try to distill patterns from large amounts of past data in the hopes of predicting future event. This model is problematic when predictions cause actions which influence the future or when the future cannot be assumed to be like the past. This is true, for example, in most security or abuse prevention applications. In the online model, one operates in an adaptive mode without assumptions of data stochasticity. The goal is to act in a way that is provably close to optimal for any sequence of inputs. Interestingly, this is often possible without standard ML machinery. This talk will present some basic concepts and an introduction to this fascinating subfield of computer science.