Neurons in the mammalian brain often exhibit complex non-linear responses to multiple task variables (mixed selectivity). The responses are so diverse that it seems hopeless to find some interesting organization in their response properties. However, it is often possible to observe an interesting structure in the representational geometry: task relevant variables tend to be encoded in approximately orthogonal subspaces. This is a signature of low dimensional disentangled representations, it is typically the result of a process of abstraction, and it allows a linear readout to readily generalize to novel situations. We show that these representations are observed in sensory and cognitive areas of the brain of rodents, human and non-human primates performing a variety of different tasks. In all these cases most of the neurons respond to multiple variables and we never observed segregated populations of highly specialized neurons. This means that typically the disentangled representations are not aligned to the neural axes.