Cells are controlled by the complex and dynamic actions of thousands of genes. With the sequencing of many genomes, the key problem has shifted from identifying genes to knowing what the genes do; we need a framework for expressing that knowledge. Even the most rigorous attempts to construct ontological frameworks describing gene function (e.g., the Gene Ontology project) ultimately rely on manual curation and are thus labor-intensive and subjective. But an alternative exists: the field of functional genomics is piecing together networks of gene interactions, and although these data are currently incomplete and error-prone, they provide a glimpse of a new, probabilistic view of gene function. We outline such a framework, which revolves around a statistical description of gene interactions derived from large, systematically compiled data sets. In this probabilistic view, pleiotropy is implicit, all data have errors and the definition of gene function is an iterative process that ultimately converges on the correct functions. The relationships between the genes are defined by the data, not by hand. Even this comprehensive view fails to capture key aspects of gene function, not least their dynamics in time and space, showing that there are limitations to the model that must ultimately be addressed.