well and the reason i really wanted to do that actually was to use something like unum and somehow figure out how to encode symbol transfer probabilities for a tANS compression implementation
because the crazy thing nobody seems to care about with tANS is that you can specify how much space you want to allocate for compression data instead of letting the data drive your allocation. and then you have this really neat way to specify your encoded message as the delta between transition table states!!!!
but i also believe duda's work on compression through tANS is limited to first-order state transitions instead of attempting to support e.g. slurping multiple symbols into one? or vice versa?
but a few weeks ago i started thinking i got that wrong. i think the inferential model he uses is very basic and clearly derived from his physics background (which is, incredibly, not derogatory) and i think we can do better is all. but it does kind of require me to incorporate matrix multiplication shit and i will simply have to do that later especially because compression is so much more effective with domain expertise guiding it
