As I understand it, when starting up (beginning of stream) - the network does not compress, then guesses the next byte at every turn determininstically from input so far.

Decoding can "read" the fist byte, then proceed to learn and guess the second byte and so on.