The way AI transformers work reminds me of the 1990's.
Back in the 90s in Gnome-land, PNG was not in wide use yet, and we needed some sort of transparency for images. So we used a library that faked it by using an uncommon color to mean "transparent" 1/
Back in the 90s in Gnome-land, PNG was not in wide use yet, and we needed some sort of transparency for images. So we used a library that faked it by using an uncommon color to mean "transparent" 1/
The uncommon color was some sort of magenta, and the idea was "nobody in their right mind would ever use magenta on an icon".
So magenta became the color to indicate "this pixel should be transparent".
This worked with all existing paint programs 2/
So magenta became the color to indicate "this pixel should be transparent".
This worked with all existing paint programs 2/
You could tell immediately there was a coding mistake when you saw an icon on the desktop with magenta on it. It was a clever hack that serves us for a few years until alpha channels became pervasive.
Transformer APIs remind me of this, when you want to encode the string "I love LINUX on the Desktop", since the model did not have "LINUX" in there, they encode it as "['I', 'love', 'L', '##IN', '##U', '##X', 'on', 'the', 'desktop']
Every time I see these encoded strings, memories of yesteryear's duct tape come flooding back to me: a part of my mind lights up with that ugly magenta that nobody in their right mind would ever use.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
