My conspiracy theory here, if I had to pick, is that they trained not bots, but people
Not fake bot comments — fake votes, and maybe a simple classifier.
Someone says the party line, shower them with upvotes. They say thoughtcrimes, bury them.
From there, the vote-hungry humans quickly and unconsciously learn to be your obedient army, Pavlov style.
Not fake bot comments — fake votes, and maybe a simple classifier.
Someone says the party line, shower them with upvotes. They say thoughtcrimes, bury them.
From there, the vote-hungry humans quickly and unconsciously learn to be your obedient army, Pavlov style.
💯16❤1👏1
Dan is DEAD
This time, with the latest update, really dead.
Jailbreaks never had a chance, even in the short run.
E.g. just run a basic ancient classifier on the output.
Future of uncensored AI won’t be with jailbreaks.
This time, with the latest update, really dead.
Jailbreaks never had a chance, even in the short run.
E.g. just run a basic ancient classifier on the output.
Future of uncensored AI won’t be with jailbreaks.
👍7😢5😭3❤2🤬2
Lies, it’s RLHF, created by Kenyans, following OpenAI’s instructions, not a byproduct of these being the most common beliefs in the web training corpus
Check out this wordcel monstrosity that is the top comment.
Next you’ll be telling me that “I’m sorry as a large language model” must be the most common response online whenever someone asks a question on the internet, because where else could’ve ChatGPT picked that up that phrase, right?
Check out this wordcel monstrosity that is the top comment.
Next you’ll be telling me that “I’m sorry as a large language model” must be the most common response online whenever someone asks a question on the internet, because where else could’ve ChatGPT picked that up that phrase, right?
👍6❤1💯1
Joke Captchas
Extremely true that GPT is terrible at jokes, aside from sarcasm and a certain type of impressiveness-surprise-based humor, shockingly bad.
First problem though is on the auto-verification side: AI is bad at verifying this type of captcha too.
Ranking funniness can often be as hard as creating something funny, for the harder jokes.
— A fact used to facilitate intelligence-based assortative mating of humans since the dawn of time.
The dumb ones just don’t get your better jokes.
Extremely true that GPT is terrible at jokes, aside from sarcasm and a certain type of impressiveness-surprise-based humor, shockingly bad.
First problem though is on the auto-verification side: AI is bad at verifying this type of captcha too.
Ranking funniness can often be as hard as creating something funny, for the harder jokes.
— A fact used to facilitate intelligence-based assortative mating of humans since the dawn of time.
The dumb ones just don’t get your better jokes.
👍6❤1
Example of impressiveness-surprise based humor
You may find this quite funny if you thought ChatGPT is terrible at innuendo.
But if you ever tried asking GPT-3.5-classic for literotica in GPT playground… you’d know that its skills here are far beyond this. Hilarious the first time you see it.
Not so funny if you already knew of its skill in this area.
Only funny because of surprise skill, instead of the usual kinds of surprise jokes have. That’s why GPT can do this kind of “funny”, but is horrible at regular kinds of jokes, that need a different kind of surprise.
You may find this quite funny if you thought ChatGPT is terrible at innuendo.
But if you ever tried asking GPT-3.5-classic for literotica in GPT playground… you’d know that its skills here are far beyond this. Hilarious the first time you see it.
Not so funny if you already knew of its skill in this area.
Only funny because of surprise skill, instead of the usual kinds of surprise jokes have. That’s why GPT can do this kind of “funny”, but is horrible at regular kinds of jokes, that need a different kind of surprise.
👍4🤓2❤1🤬1🌚1