Comment by dofm
9 hours ago
Something that bothers me about AI generated content more broadly is how unmemorable it is. I don't mean as in bad. I mean literally, as in hard to remember or recall.
Despite seeing a lot of them, I cannot think of one AI-generated photo that I can picture clearly in my mind; a few are partial but elusive. Whereas I can recall (visualise) a whole bunch of traditional photographs.
The same is true of AI generated text. Only the annoyances stick. I cannot recall real details of text I have generated, until I commit it to memory some other way.
I don't think this is about ephemerality either. If we assume it's about celebrated/famous/infamous images, there are definitely non-ephemeral, cultural moments in AI generated images in particular, like Boris Eldagsen's Sony Prize winner:
https://petapixel.com/2023/04/14/artist-refuses-prize-after-...
This really should be memorable, but isn't. I had forgotten the second person is in the image.
Or Jason Allen's fake painting:
https://petapixel.com/2022/09/01/ai-generated-artwork-wins-f...
I had already forgotten there's more than one figure in it, and I only looked at it a few weeks back. I remember the colour, the bright circle, some vague hints of texture; one figure. And that is it. Only the crudest shape elements.
For me, something about AI-generated text and images confounds recall. It is really peculiar.
That's the cost of an AI work that is derived from the outputs of others.
There's no real edge to it. Same as with the writing. The stuff that you'd latch onto (and thus remember) is simply not there, precisely because those image or word choices would be just outside its latent probability space. But because they're well inside it, your mind sees nothing novel to register.
This is also why I think human output will actually increase in value. When any AI can just "phone it in", something genuinely human will stand out (to us, not the AI) and become a bellwether.
This will literally help us realize what it means to be human.
I also don't think the solution is simply to "make responses more random", either. That might help solve novel problems (the same way that throwing darts randomly at a dartboard eventually hits the bullseye of the dartboard right next to it that no one considered), but I don't think it will help it "seem more creative".
> There's no real edge to it.
Yes, as if it is in some weird hidden dimensional sense completely uniform.
ETA: suddenly reminded of the Bateson quote about information being “the difference that makes a difference”.
That checks out, right? We're instructing LLMs to choose from the least "surprising" tokens at each step (modulo some temperature). It feels smooth, slippery - no friction to the eyes that glaze over as they skim the text, and no bumps to snag in your memory. Like waking from a dream, recalling only the barest shapes of a few concepts or themes, until it vanishes with your morning coffee.
The examples I've seen of AI music (though I avoid it on principle) seem the same.
I have this same idea about why it's hard to remember dreams, but even more so, why it's hard to remember my kid's or spouse's sleep-talking.
Sometimes I'll check in on my sleeping kid and she'll sit up in bed and say some utter nonsense. I'll find it hilarious, giggle silently to myself, and kiss her goodnight again and she'll close her eyes and lie back.
Why I try to tell her about her sleep-talking in the morning, though, I find that the words she said have completely disappeared from my memory, no matter how funny I thought they were at the time.
In my head-canon, this is because it's dream language, and slips away as easily as dreams. But, like the AI art, it could be because it's bullshit: completely devoid of content, all signifiers and no signified.
I went through a phase of leaving notepads next to my bed to try to write down my dreams and it simply never worked.
I stopped after I wrote something on the pad while I was still asleep. Woke up to text with letters that were backwards, upside down, weird words — so close to real words that I was sure I ought to know what they meant and had really meant to write them down.
Scared me. Literally too weird to keep. I tore up the page.
In a way I think this is one part of the same continuum. There are thoughts that have meaning and can have no words, and words that look like they should have permanent memorable meaning and don't.
> Scared me. Literally too weird to keep. I tore up the page.
??????
1 reply →
The memories of dreams evaporate remarkably quickly. Much faster than anything else. It can't be accidental; I think our waking brains must be purging those memories as a matter of urgency. If we remembered them, we'd "learn" from them, which would be a catastrophe.
You know how you try to run in a dream, and you don't get anywhere? Imagine if your motor cortex started to expect that, when you're awake. It would destroy you.
On the other hand I'm still trying to forget the weird thing happening with this one guys eyes in a WickedAI video I was only able to watch half of.
I remember seeing some very strange, deliberately creepy images that were generated to accompany a two-paragraph creepypasta about a 19th Century Belgian expedition into the jungle.
They were actually rather good in a sort of "fake collodion image" sense, and the eerie early-DALL-E quality to them really helped the spookiness.
But I can only remember this technicality and the feelings with any clarity, not any of the details except in the broadest sense. I cannot bring these images to mind in any meaningful way.
They were deeply wrong and it's only the wrongness I really remember. It confounds memory.
Modern image generators have ironed out all the structural wrongness.
I wonder what you mean by ephemerality here, since those images are definitely sloptastic as hell. Compare to works in similar style, like Dorothea Lange[1] or Gerome’s orientalist pictures[2].
Reason why those images are flat and boring is that they are just statistical guesses making a composition averaging whatever the model has been trained with. They would be technically brilliant (if made in oil), but superficial and meaningless, same as so much Sunday painting is.
Same goes with language. Nobody is trying to communicate anything with you, so it just words after another. You can create meaning out of it if you want of course, we homo sapiens -apes excel at that, but what’s the point? Language Jones on YT has pretty good video on this[3].
[1] https://media.mutualart.com/Images/2024_01/12/12/124216388/d...
[2] https://uploads4.wikiart.org/00339/images/jean-leon-gerome/t...
[3] https://m.youtube.com/watch?v=ORgKY9AlybA&ra=m
Well the examples I gave rise above their ephemerality due to the circumstances that make them memorable. I can remember the details of the story around them — the way the prize winners reacted in each story — in such a way as to contrast them.
The way my memory works (especially as an amateur photographer) I would thus normally have a very good chance of remembering some key details of the images; some fascinating element of each would connect with the rest of the memory.
But it does not happen. Whereas I sometimes remember photos with clarity while forgetting where I even saw them.
Yeah, its really weird. Maybe its a cognitive bias that says "an AI made this, so it isn't important," but I can remember perfectly the events of a book I read 10 years ago, and a book I read 1 year ago, and another I finished 2 months ago. Meanwhile I can't remember what claude told me yesterday.
I think it is because they are on some latent level cognitively uniform and unchanging.