Подтвердите e-mail

Для публикаций, комментариев, реакций и сообщений подтвердите адрес.

Публикация

One way of thinking about genAI is: Suppose you steal every English word on the internet, and then look up what words are most likely to follow the word "throbbing." And then whenever you see "throbbing" you slap on a likely word You're gonna end up talking about throbbing dicks in weird contexts

Обсуждение

8 прямых ответов · 31 сообщений

That's not hallucination, because that's literally it working exactly as intended: outputting a statistically probable next word. And the corpus, the entire English internet, contains a whole lot of throbbing manhoods.

Ответ для Pookleblinky

So you try to stop this "hallucination" You hardcode it so if it tries to output the phrase "throbbing manhood" it throws an error or something. But oops, you literally stole the entire internet. People were creative when writing sex scenes.

Ответ для Pookleblinky

So it still outputs "throbbing erection" when generating a transcript of, say, a cancer patient describing the pain of their lung cancer.

Ответ для Pookleblinky

So you go nuclear, and make a whole list of banned words. Oops, you literally stole the entire internet. Now a kid is handing in a history essay containing the phrase "throbbing, quivering thighs"

Ответ для Pookleblinky

A police officer is generating a transcript of an arrest and now it contains the phrase "throbbing essence of grizzly bear lust" Oh no, you've removed all the *most* likely outputs so it's now it's skimming the weird parts of AO3 that you stole while literally stealing the entire internet.

Ответ для Pookleblinky

And that's the problem: when you steal the entire Internet to create a statistical algorithm for determining the next word, you're going to end up outputting the statistically most likely next word. From the *internet*

Ответ для Pookleblinky

It's not hallucination, it's the statistical analysis doing exactly what it is supposed to: slurping up the Omegaverse to learn what word is likely to follow this word.

Ответ для Pookleblinky

Mmm. Spicy autocorrect, indeed.

Ответ для Pookleblinky

This moose is unable to stop giggling at the sheer "be careful what you ask for" of this thread. They'll end up with something like PLAYGOL[1] that only generates obscenities without using any "rude" words or human input. 3:O))> [1] Fictitious programming language from "The Devil's DP Dictionary".

Ответ для Pookleblinky

And of course if you use prompts to try to keep it in line it will weight out those prompts on a long enough timeline until it "forgets" them and you end up with an AI teddy bear talking about fetishes with your toddler.

Ответ для Pookleblinky

People get mad at me when I tell them that AI sucks because I wrote ~3% of all reddit posts from 2011-2022 but it's true. I shitposted so hard I defeated Rocko's basilisk.

Ответ для Pookleblinky

(I've mentioned before how an AI evangelical badgered me into trying out his chatbot. Having recently been playing D&D I asked it about mimics, and was slightly surprised to get a summary of Everybody Loves Large Chests. Although knowing the guy who pestered me, maybe not too surprised.)

Ответ для Mina the Husky (and her human)

(For those who don't know, ELLC was a LitRPG focused on an unusually intelligent mimic. Also there were some really weird erotica scenes in there. No, like, weirder than you're thinking. NSFW in case it's not obvious already. Also several kinds of content warnings apply.)

Ответ для Pookleblinky

"slurping up the omegaverse" 😔

Ответ для Pookleblinky

Have you guys SEEN the “internet”? *shudder*

Ответ для Pookleblinky

This is the part I've never understood. The internet is wrong about tons and tons and tons of shit, and it's sarcastically ironic about even more stuff. Signal to noise ratio anyone?

Ответ для Skrrt Russell
Ответ для Tim E. o’Danaos & Don A. Ferentis (Structural Woodworkers) Ltd.
Ответ для Pookleblinky

One thing: "throbbing essence of grizzly bear lust" likely predates AO3 thanks to the (in)famous award winning Can-lit novel "Bear"

Ответ для Pookleblinky

Lol yup.

Ответ для Pookleblinky

Ever read The Eye of Argon? Were it not from an age decades ago, it would be dismissed as AI. The lust interest is described as having a "lithe, opaque nose"...

Ответ для Pookleblinky

90% of the internet is about throbbing dicks and less than 1% of dicks are throbbing at any given time so AI changes nothing

It's autocomplete with delusions of grandeur

I’d correct this by saying if all you give it is “throbbing”, youre gonna be talking about dicks eventually. If you give it a paragraph it’s “attention” considers ALL of. So, not unlike a person, if you just say “throbbing” and then wait smirking, expectantly, the results will be similar.

Ответ для Ford “Crash” Mandicoot
Ответ для Pookleblinky

They’re not for assessment; they are for structuring or parsing unstructured information and very reliable at it.

“throbbing shitgibbon has profoundlt stupid ideas about minorities”

Or gristle.

“Gristle” may also have some incongruous appearances