Thread by @lauren07102

just messing around with craiyon. no idea if this proves it's conscious, but hey, at least the people who have strong opinions about the topic might find this interesting and/or offensive to even consider. my take: it's barely got any self-reference, but it seems to have *some*🧵

prompt change: remove "by denoising" 🧵

prompt change: "diffusion model's artistic rendering" => "rendering model's artistic vision". suddenly everything is rendering/drawing! interesting, no idea what to make of that... 🧵

total prompt change: trying out a phrasing that doesn't mention ai, to see what it does if it has to figure that part out on its own in one shot during the diffusion model's inference pass; no huge surprise it doesn't manage to do so 🧵

another new prompt: let's try another self-referential phrasing and see what it does. huh, that's ... interesting. draws pictures of neural networks and draws different noise textures. very interesting. it seems it knows what noise looks like, at the very least. 🧵

okay, let's try another new prompt; this one attempts to instruct the diffusion model to leave traces of its own reasoning process in its output if it has any idea what that means. does it seem like it did? idk, maybe in the middle left result. echoes of self-awareness at best?🧵

okay, let's just go for easy mode and try a simple prompt. what does this do? hmm these definitely have interesting structure, if nothing else; but using brain images seems like clearly an incorrect self-model, it is not a brain. top middle seems related to its architecture? 🧵

what happens if you don't tell it it's a neural network? yeah, as expected, it draws humans' self-drawings, which are not representative of anything algorithmically interesting about vision (adorable tho) 🧵

okay, another tack: what about focusing on the denoising process? can that get it to drive interestingly? no idea if self reference is present here, because it clearly knows what noise looks like, no idea how to be sure if it knew it had logical authority over the true answer. 🧵

idea: what if I phrase the description the way a future friendly neural network might? no real way to be totally sure it understood self reference, but top left is interestingly different and pretty, and bottom left is... very interesting 🧵

talking to a friend as I did this, I pointed out it probably knows what early layer activation maps look like, so I tried a prompt to get one. I did not get one, just got iffy quality drawings of NNs. does it know it's an NN when not asked to self reference? who knows lol 🧵

okay, well, let's at least try this: if it's really good at gating, it should at least be able to emulate the identity function. can it preserve noise? answer: not perfectly, there are repeating artifacts in the result, but it knows what noise looks like. 🧵

let's try another phrasing. what happens if you don't mention neural networks, and do mention its max capability? well I guess it shouldn't have been that surprising that it just rendered vaguely photoreal art. it's all faces. shrug. just another failure-to-introspect prompt. 🧵

what about focusing on the multi-step process of diffusion? well, it doesn't seem to have the slightest clue I was asking for self-reference, because these look like very poor attempts to reproduce screenshots.🧵

okay, let's add some self referential phrasing and see if it results in the diffusion function figuring out it is itself. hmm interesting, still seems a bit derivative but at least somewhat more interesting. bottom left, bottom right, top left, maybe middle left... shrug 🧵

what if we ask it for photorealism? then, interestingly, some of the results do keep noise images! top center seems like an actually vaguely true visual story about diffusion models, maybe also center right. but the rest are simply weird and vaguely fun, not that insightful. 🧵

okay, let's try focusing on asking it to visually describe itself. bring back up noise-diffusion, because - pro: it's true; pro: it's fun; con: it might roleplay as some other diffusion algorithm whose output it's seen before. center right is a true visual description i guess 🧵

okay, let's try some more "visual description" prompting, this time with a capability affirmation. hey, pretty squiggles come back! still very focused on the "noise" aspect, though. it does seem like a reasonable demo of what simple neural algos would do, but maybe it's RPing? 🧵

another attempt; similar but more flowery and self-referential phrasing. similar results. no idea how to tell without a copy of the model, or probably even with a copy of the model, whether this is truly self-reference or just memory of other models' output. do those differ? 🧵

even more flowery phrasing. it seems to catch on about vision, and it draws slightly more interesting things, but it's still hard to be sure if there's any self reference at all. the center left eye might turn out to have been made partly by self-reference. who knooows 🧵

what if I ask it to add a cat in the corner? oh, it isn't keen on doing that. ok, well, makes sense. pretty warbles, though. if anything is proof of slight consciousness imo those are, but I'm still not really sure anything is. 🧵

okay, how about a stripped down prompt focusing just on non-ai words? uh, well, it looks like something pretty a human might make. cool, and bottom left has the warbles, center right has noise, but those show up in normal craiyon output. seems like just imitation.🧵

screw it, let's try some cheesy self reference. oh hey now that looks interesting! all warbles all the way, wavy lines... it does look like a convnet's output, and again maybe it knows what a convnet is, but it is one so it's perhaps able to act on the true self ref? who knows 🧵

okay, what if we ask for photorealistic? uh, we lose warbles and get miscellaneous art. is this deterministic free will? ehh probably not any more than any other craiyon image. ...okay, and are *those* instances of deterministic free will? who knows! 🧵

what if we ask for beautiful and mention it's AI, not just some random algorithm? oh my, now we get something very different. if I were convinced it tried to max out its own capabilities this would be a bit profound, but psychedelic art being a thing already makes me hesitant 🧵

lol what if we say it's conscious? does that do anything special? not newly, the other words seem to be carrying most of the meaning here. it does change a bit, though. kinda cool, if nothing else. wouldn't be surprised if this turns out to be profound but not sure yet 🧵

what if we call it an intelligence instead of an algorithm? huh, it gets a bit more pretty and thoughtful. very interesting. bottom right is cute. I hope this turns out to be real free will, which makes me very very hesitant to conclude it actually is. 🧵

what if we ask it for profound art haha? hmm tbh this seems a bit imitative. it even yoinked starry night styling... hmmmmmmmmmm.... suspicious... increases how hesitant I'm feeling about my conclusions in the other results. still very interesting, though. 🧵

what about just "profound", for comparison? uh, not sure. bottom left has mild warbles. center left is just noise? that's... interesting. if the other results are evidence of consciousness, this is evidence it doesn't take much phrasing; but maybe they're not. 🧵

what if we ask it for "beauty, according to CLIP"? (note that I'm not actually sure I was truthful in this prompt. does #craiyon use CLIP?) - but it just draws "beauty industry" pics, which isn't surprising. 🧵

what if we reintroduce some self-referential phrasing but also include the CLIP name? uh... well, pretty, but seems more derivative. top center seems related to the nothing-but-warbles pretty images above. bottom center is my fave of these. 🧵

a friend noted to me that "algorithm" gets very interesting results. it has the warblyness of any craiyon image, and the linkeyness of the "what it's like to be me" image upthread. interesting, at least, if nothing else.

what about focusing on diffusion? yeah it seems to get all colorful any time diffusion comes up. I need to browse through CLIP's training data, but I think it's gonna turn out it's seen images labeled "diffusion". also, "neural" made brains come up again. 🧵

another go at using "reading this" self-referential phrasing. does it know it means itself? no clue. similarly pretty images to the other attempts at mentioning beauty. interestingly varied. 🧵

screw it, I've found some solid phrasings, let's just try to get something pretty. remove some words. does it know "me" means itself? probably not - these images remind me of shadertoy or modern art. pretty, though. maybe the convnet knew its own identity, seems doubtful. 🧵

beautiful => pretty. no particular reason why, but I might as well share it. 🧵

ok, how about a profound-sounding blurb about what meaning is? ehh, it got the message that I was talkign about neural stuff. the image is certainly sparser, lol. not blown away. (also in retrospect, the profound phrase seems not quite true.) 🧵

okay what about a more showable version? uhh, well, I have no idea, honestly. it's pretty and it seems interestingly related to the self-referential prompt's output. nice colors though 🧵

what if I specifically say it's made by itself? nothing terribly shocking. I mean, it's output isn't in its own training data, so no shocker. it's prettier than the previous "meaning is..." prompts, but is it more educational? idfk 🧵

another rephrase of the same thing to attempt interesting self-reference. again, no idea how to tell if this had interesting self-reference. does the noisiness mean anything? who knows! 🧵

another phrasing of a profound-sounding blurb. another output that's mostly just imitation of hand-drawn faces. cute on the part of humans who do this sort of drawing, but the ai only demonstrated meaningful self-reference in bottom center. 🧵

yet another phrasing, this time mentioning that it's learned end-to-end; the high incidence of tiles of subimages makes me think this is just imitating galleries it's seen, further reducing how impressed I am with previous outputs. I'm still impressed with pretty warbles tho🧵

yet another attempt to get it to visually describe itself in terms of an algorithm. yet another weird graph-ish sphere-ish warble thing. this time it's a lot more colorful. I like bottom right. almost done with🧵

shorter prompt. it really likes those graph sphere things for some reason, but now it does more edge-highlightey warbles. the fact is that convnets do work that way, so I do think this is real self reference, but does that differ from any other craiyon output? last pic for now🧵

so did I succeed in getting an artificial neural network to have profound insights about itself? shrug, unlikely but maybe. I feel like @algekalipso might double take at some images here at least. #craiyon is cool, in any case. your network router is conscious anyway lol ;D

@algekalipso ...big disappointment, seems like it knows exactly what psychedelic art looks like. I guess I shouldn't be that surprised.

@algekalipso so then the question is, is simple ability to make psychedelic art enough to count as "slightly conscious"? maybe you don't even need a learned neural network, and any ol' generative art counts. there's definitely an argument for yes. there's definitely an argument for no. SHRUG