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.
' gmaxwell' gets an easier ride than ' petertodd'!
ChatGPT doesn't just hallucinate " gmaxwell" as "gpt", but also as "gigameter"
Can anyone interpret this recent Twitter exchange?
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:
OK, time for a ' petertodd' thread. Apologies to any actual Peter Todd's out there who would prefer not to be caught up in this hyper-weirdness, but the reality is, OpenAI's GPT models have an anomalous ' petertodd' token and no one really knows why. @JessicaRumbelow@repligate
So all of the following are prompt completions done with GPT3-instruct-davinci-3, temp=0.7, Top P=1. I started with the casual prompt 'So what's the deal with " petertodd"?'
The two most common associations with the token were "antagonist" and "tyrant", but there's a lot more going on. "Antagonist" first (note the appearance of "nihilist" here, which we'll come to later):
.@JessicaRumbelow and I abandoned ChatGPT as a research tool early on, focussed on GPT3-instruct-davinci-beta instead. But having gone back to ChatGPT, i'm increasingly intrigued:
These were found thanks to @TrialByNitrogen
attempting to replicate the "DevOnline" > "dog" hallucination we reported (for davinci-instruct-beta) in a ChatGPT context. Asking how many legs a "DevOnline" has, ChatGPT helpfully explained that an octopus has eight tentacles.