jdfr

OP seems to have run a programming language detector on the generated texts, and made a graph of programming language frecuencies: https://pbs.twimg.com/media/Gx2kvNxXEAAkBO0.jpg?name=orig

As a result, OP seems to think the model was trained on a lot of Perl: https://xcancel.com/jxmnop/status/1953899440315527273#m

LOL! I think these results speak more to the flexibility of Perl than any actual insight on the training data! After all, 93% of inkblots are valid Perl scripts: https://www.mcmillen.dev/sigbovik/

The beginning of the text feels really odd, almost GPT-3-like at some points. At least, that's the way it seems when you do not know the jargon. It just rambles on and on. Thankfully it gets more coherent later on, but I am not sure what the author intended to say in this piece.

This video (wanderers) does not give a reason, but it sums up quite nicely the way I feel it: https://youtu.be/YH3c1QZzRK4

Whoa! Quite similar to the "magic" behind compressed sensing (as noted by the authors of the paper linked by the OP):

https://en.wikipedia.org/wiki/Compressed_sensing

Nice to see that similar mathematical tools are being developed for similar problems.

Edit: typo

Vitter's algorithm is for when you need to generate a random list of unique items with sequential reads. If you drop that requirement, you can just use a substitution-permutation network: https://en.wikipedia.org/wiki/Substitution-permutation_netwo...

A SPN will generate a unique random permutation, from which you can draw your random list of unique items. With a little of post-ptocessing, you can generate random permutations of arbitrary size.

I have implemented it in matlab and python:

http://nl.mathworks.com/matlabcentral/fileexchange/36626-alg...

https://gist.github.com/jdfr/46633c630471494d67ae