Thread with 8 posts
jump to expanded post
you know how llms will overgeneralise their training data and start hallucinating/extrapolating stuff?
this happens to me all the time with emoji. the original set in ios 4(?) was tiny. i could remember what was and wasn't an emoji. but there's too many now, so i “hallucinate”
like there are so many fruit emoji now that i subconsciously assume any fruit i could want has an emoji, even though that's not the case. and it gets so bad that there are specific emoji i can picture in my mind that don't exist and never have…
the one that really gets to me is that there is no edamame emoji! but i can picture in my mind what it should look like! maybe i saw a tiny illustration of a soybean pod once and it got filed into my memory as “emoji” even though it wasn't. yes, i know there's a peapod emoji now
makes me wonder if llm hallucination is fundamentally unsolvable. humans are also prone to getting things wrong if you expose them to too much data
omg ios 4 is old enough now that people could have nostalgia for the old, tiny emoji set, and when emoji weren't in unicode yet and you had to use a special app to get access to them if you weren't in japan…
@hikari recently I was making a badger badger song reference and thought “there’s no way there’s a badger emoji, right?” but there totally is: 🦡
I kind of can’t shake the feeling it JIT-unicode-consortium’d on the spot
@0xabad1dea @hikari 🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🍄🍄🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🍄🍄🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🦡🐍🐍🐍🐍
@jernej__s @0xabad1dea @hikari https://m.youtube.com/watch?v=C0apsOhWV2M obligatory badgers metal remix posting