They're discrete with such high cardinality that successful approaches will likely model them assuming they're basically continuous. Neural network layer activations are also discrete after all, but they're often 256+ dimensional vectors of float32s or float16s.