As nearly as this 20th-century old geezer understands it, there has been a sort of feedback loop between internet memes, artificial intelligence and image trawling that some people are claiming has produced a self-emerging viral entity. A “demon” or virtual life form. Is this accurate?
Can you cite a non-ironic example of this purported phenomenon? Cuz this is the literal first I’ve heard those neologisms.
First I’ve head of them, as well.
Apparently, they’re not self-emerging. Both Crungus and Loab are results of new-fangled AI art software. There is at least one long thread in Cafe Society where Dopers have been playing around with such software and uploading their images for the board’s viewing pleasure.
I ran Crungus (using a prompt of “crungus, artstation”) through my local version of Stable Diffusion, and it came up with some interesting ones:
It definitely inferred a kind of creepy monster from the prompt, but none look like the examples from the Futurism article.
I know less than nothing about this sort of stuff, so my question is probably way basic - If the next person asks the Art Thing to draw ‘Crungus’, will it just find these pics you’ve posted and produce something that is vaguely reminiscent and referential?
And next dumb question - how is iterative copying supposed to be a ‘self-emerging viral entity’? Or am I missing some profound subtlety?
What you’re describing is a potential risk in other situations, but not here. A few things:
- The training sessions aren’t continuous. They typically happen only once a large number of new images have been collected. The training is also expensive–the Stable Diffusion net allegedly cost $600,000 to train, and is one of the cheaper ones. It’s not something that’s happening daily, or even monthly.
- The data comes from specific data sets that gather text in a specific way. The LAION-5B dataset uses alt-text for the text description of the images, which works well for sites like ArtStation but not here (since I did not provide alt-text).
- They probably wouldn’t be crawling a webforum like this in the first place. They want large collections of known good images, like ArtStation, Getty Images, DeviantArt, and so on.
- Many of the outputs are watermarked. Sometimes invisibly, but in a way that the training program can distinguish. They’ll want to throw away that data.
It’s kind of a silly phrase, but I think the idea is just that it starts as something completely random, but eventually the more interesting versions of Crungus will be selected for, and people share these images and so on, making the association even stronger, until the neural net has a very strong association with this particular version of Crungus. I don’t think that’s what’s happened here, though.
Thanks for the detailed reply. I think I’m understanding what this is about a bit better.
My WAG would be that the AI, not recognizing the word “Crungus”, is comparing it to similarly-spelled words much like how Google will auto-suggest search terms if it thinks you’ve misspelled something, and it’s settled on an amalgamation of “Krampus” and “Chungus”.
I think it’s probably going a step deeper than that. It’s likely picking up an association with multiple parts of the word, finding something about them (maybe the initial “k” sound, or endings of “us”) that is associated with monsters. Not looking for misspellings exactly, but rather picking up that there are some “monster-sounding” words.
BTW, “chungus” gives me pictures of fat rabbits (as expected). All different styles, though, and none that really look like Big Chungus.
Crungus and Loab together. I like them as game pieces.
I sometimes toss nonsense words or namish sounding word pairs at the AIs. I recently tried “toogiewhuffle”.
Two of those bottles of toogiewhuffle have mold growing on them…
Similar, for “Loab”, if I check Google images for that result, and ignore the ones that reference the AI cryptid, I see these three references:
I’d venture to guess that it’s merging the images with the Korean lady taking the lead spot.
And, if the AI is continuing to learn from the Internet then the more “loab” images that are out there, the more it will reinforce the image.
@Darren_Garrison This is what Midjourney gave me when I typed in “toogiewhuffle”.
Looks like a castle growing out of a massive tree.
Probably where the sauce is made.
I would definitely consult Crungus & Loab for all my legal problems.
I often refer to the firm of Various & Sundry.
The lettering on the sauce bottles doesn’t look dissimilar to the script used in Zelda: Windwaker.
Benny Hill employed the distinguished firm of Martin, Barton, Parton, & Fargo.
Anyone try entering “Leopold and Loeb”?