This looks interesting but the size isn't a fair comparison, from my skim it's storing embedding generated by an existing model, the 100kb is some embeddings that the procedure needs, it still needs a full on text-to-image model to go with it. It's closer to saying i can store some prompts in a 100k text file. Not to diminish the work, just the presentation exaggerates meaningless comparison points.
Before Threads or Tiktok, before ChatGPT, before Stack Overflow, before Google, before the Internet, before USB flash drives, before CD-roms, we had floppy disks to store and move data around on. They came in 8-inch, 5¼-inch, and 3½-inch floppy sizes, and held 2 MegaBytes, tops*. They were loud and slow, and prone to corruption, but that's what we had, back in the day.
And you could boot a small linux to act as a firewall/nat service [0] for your brand new cable modem service on a cheap 486. Ran that for 3-4 yrs before nat routers became a thing.
10 comments
[ 4.2 ms ] story [ 36.1 ms ] threadBefore Threads or Tiktok, before ChatGPT, before Stack Overflow, before Google, before the Internet, before USB flash drives, before CD-roms, we had floppy disks to store and move data around on. They came in 8-inch, 5¼-inch, and 3½-inch floppy sizes, and held 2 MegaBytes, tops*. They were loud and slow, and prone to corruption, but that's what we had, back in the day.
https://obsoletemedia.org/data/floppy-museum/
* 2 MiB raw, 1.44 MiB formatted, the way I recall, for the 3½ ones. There were later formats but they never quite caught on.
[0] https://en.wikipedia.org/wiki/Linux_Router_Project
https://arxiv.org/pdf/2305.01644.pdf