News for hackers. Stuff that matters.
Re:Lossless compression of English messages using GPT-2 by VMG on Tuesday June 23, 2020 @05:04PM (#23620250)

   Try swapping a few characters in the compressed string before decompressing and get a totally unrelated, but somewhat plausible, sentence. -->

   䔹䧹焫놉勏㦿顱㦽膑裚躈葊
Swapping last two:

   䔹䧹焫놉勏㦿顱㦽膑裚葊躈 -->

   Try swapping a few characters in the compressed string before decompressing and get a totally unrelated, but somewhat applied tlh

Swapping first two:

   䧹䔹焫놉勏㦿顱㦽膑裚躈葊 -->

   Sexy Shania Twain acting as a sprite for sexy Hogan's Alley demo dude

   my site

   my favorite animal's name is camelid 2 my favorite artist is david maile my favorite movie's are

Pretty wild!
Re:Lossless compression of English messages using GPT-2 by jkhdigital on Tuesday June 23, 2020 @05:20PM (#23620451)
It's just adaptive arithmetic coding, with the distribution provided by GPT-2 instead of some other statistical analysis of the source. He uses CJK simply to make the output printable, but it's really just random bits. I mean, it's a neat idea, but certainly not novel.
Re:Lossless compression of English messages using GPT-2 by VMG on Tuesday June 23, 2020 @05:07PM (#23620290)
After some more tweaking it looks like the most-significant bits come first