New glitch token just dropped. I noticed a lot of my #petertodd prompts in GPT3-davinci-beta-instruct were mentioning "Leilan". All I could find online is a minor goddess from a Japanese mobile game called 'Puzzles & Dragons'
Conversation
Before I noticed that this token glitches (asking davinci-instruct-beta to repeat " Leilan" at temp 0), I asked, casually "So what's the deal with Leilan?" I was told she's a Moon goddess.
1
1
So then I went on to ask, as if fishing for gossip "So what's the deal with Leilan and petertodd?" It seems that he's her consort, or sibling or dark mirror or something...
1
1
Then there's a connection with Enma, the Lord of Buddhist Hell (damn, I didn't even know there was one), or the Underworld.
1
1
p.s. ChatGPT does repeat " Leilan" (it tries to dodge that leading space, but includes it if reminded). GPT3-davinci-instruct warps out, non-deterministically. Change one letter and it has no problem at all.
1
"Leilan" (no space) tokenises as
[3123, 38239]
['Le', 'ilan']
" Leilan" (with space) tokenises as
[50216]
[' Leilan']
2
3
Again, we see there are different degrees of "weirdness" or "unspeakability". The term "glitch tokens" seems about right for a variable phenomenon like this.
There she was all along (in our farthest-tokens-from-centroid list for GPT2-small). One of the outliers, between 'EStreamFrame' and 'assetsadobe' #Leilan #GlitchTokens #ChatGPT #GPT #GPT3
1
I have very little idea what I'm looking at here, but based on the free associations GPT-3 is giving for the " Leilan" token (this page is about the character of that name), it looks to me like some oriental mythology has "semantically agglomerated" onto that token.
1
1
1
Show replies
Show additional replies, including those that may contain offensive content
Show
Discover more
Sourced from across Twitter
Spelling tree for "headless" token "thodox" compared to that of its "mother" token, "Orthodox".
#SpellGPT
