It isn’t about font, it’s about (unicode) characters. Lucky for the model, most are named for “normal” letters they resemble, so it’s similar to a font problem.
Lots of edgy teens use these “fonts” (characters) on their Instagram bios as well, so thess things surely made the cut for training data.
well it’s not just a font. doing this does degrade performance, since rather than the words being common, individual tokens, each fancy cursed letter is processed as a really uncommon unique one. it will still answer correctly most of the time since in embeddings they are similar sentences, but it will probably answer worse (unless in thinking it rewrites your prompt correctly).
Because the font doesn’t matter, computers don’t read the shapes, except in the case of OCR, where the text is a raster image.
It isn’t about font, it’s about (unicode) characters. Lucky for the model, most are named for “normal” letters they resemble, so it’s similar to a font problem.
Lots of edgy teens use these “fonts” (characters) on their Instagram bios as well, so thess things surely made the cut for training data.
well it’s not just a font. doing this does degrade performance, since rather than the words being common, individual tokens, each fancy cursed letter is processed as a really uncommon unique one. it will still answer correctly most of the time since in embeddings they are similar sentences, but it will probably answer worse (unless in thinking it rewrites your prompt correctly).
This is akin to spelling errors. Which AI can also cover for.