Tokenisation: Who decides what a token is anyway?

How does a neural network understand text? It's a question so basic, you can be forgiven for forgetting about it. Words and letters are human-readable, but are discrete (even if infinite in combination) and not compatible with the continuous representations that neural networks use.