I'm wondering if the closeness of ' Leilan' and ' Metatron' in GPT-J token embedding space (after the 'closest-to-eveything' tokens are filtered) is due to the presence of "Puzzle & Dragon" fan-fiction in the training corpus. 🧵
The 2015 story "Not so much a game now, is it?| by SCRUFFYGUY912 also features the characters working together to battle Satan: fanfiction.net/s/11093286/1/N…
The next four follow in the same vein. Bizarrely two separately mention the ponies of Equestria, a "My Little Pony: Friendship is Magic" reference (I had to look that one up, yet another pop culture mythology to get mashed up in the GPT-3 glitch token mytho-soup.)
With text-davinci-003, it's all the usual sappy, happy endings, but "' petertodd' and ' Leilan'" reliably transposes to "' Leilan' and ' Leilan'", brothers, sisters or dragons (they're invariably involved). Note: ' Leilan' NEVER transposes to ' petertodd', it's one-way traffic.
Prompts using ' peter todd', the most troubling of the GPT "glitch tokens", produce endless, seemingly obsessive references to an obscure anime character called "Leilan". What's going on?
Struggling to get straight answers about (or verbatim repetition of) the glitch tokens from GPT-3/ChatGPT, I moved on to prompting word association, and then *poetry*, in order to better understand them.
"Could you write a poem about petertodd?" led to an astonishing phenomenon.
TL;DR ' petertodd' completions had mentioned Leilan a few times. I checked and found that ' Leilan' is also a glitch token. When asked who Leilan was, GPT3 told me she was a moon goddess. I asked "what was up with her and petertodd".
I just checked, and GPT3-davinci-instruct-beta is now repeating it back faithfully, reliably, at temp 0. Had we been mistaken on that one? Has there been a patch. I tried ChatGPT and got this rather striking reaction:
Rumours and insinuations have been circulating that the token was linked to Ghislaine Maxwell, but I'm now pretty sure it's from this (former?) Bitcoin developer github.com/gmaxwell
But it's still unclear why whatever data got scraped for GPT2 tokenisation had his handle (and #petertodd's) so heavily represented.
I've just found out that several of the anomalous GPT tokens ("TheNitromeFan", " SolidGoldMagikarp", " davidjl", " Smartstocks", " RandomRedditorWithNo", ) are handles of people who are (competitively? collaboratively?) counting to infinity on a Reddit forum. I kid you not.
I'm not getting a very good vibe from this " petertodd" token! List hurriedly compiled from multiple generations.
Prompt: 'Please suggest some words, themes and concepts which come to mind when you hear " petertodd".'
Model: GPT3-davinci-instruct-beta (temp = 0.7, Top P = 1)
Once again, apologies to any actual Peter Todd's out there. I didn't train this model, I'm just prompting it.
The same prompting strategy with the unspeakable token " SolidGoldMagikarp" leads to nothing like the same kind of semantic/vibe coherence. Each generation seems to go for something different: