I mean, if the circle hole is big enough any shape can fit in it. Question is whether “perfectly” means without any issue or collision, or if it means it fits without any room for error or leftover space.
"Draw a hole for this object to exactly fit through" (Okay Chatgpt, all you have to do is draw an outline of this object, not that hard...) "Here's the hole for the object." **Shows said object but with a hole inside of it**
That's true lol a few times I noticed errors and called out LLMs, but more than once I accidentally was just confused so called out the error... incorrectly, and instead of explaining it to me, it literally did a chain of thought explanation on how my (incorrect) answer is correct lol
@@PsyFyFungi it's the ultimate people pleaser with no real skills. Uh-oh, I just realised...it's only a matter of time before we start using LLMs as politicians
The problem is that ChatGPT itself wasn't drawing the rabbit shaped hole. It saw that it was a rabbit, and sent a prompt for a 2D rabbit shaped hole to an image generator, which then failed because it doesn't know what the object itself looked like from the prompt. I think.
Technically the size of the cube compared to the hole was never given, so it might be able to fit through the cicular hole Edit: Nvm I didn't notice it said "exactly"
I think the way the AI renders the images is simply with a text prompt generated by the language model, so there’s an extra layer of communication there which often can fail.
I discovered from chatgpt 4o that it doesnt actually do the image generation at all. Turns out that it just tells Dalle what it wants a picture of and then it cant even see the picture when it gets it back till you tell it it is wrong. I got ChatGPT to tell me the prompt that it was giving Dalle and there I could see the key words that were causing the problems. Once I coached ChatGPT on its prompts to be better, then I got better images. I sent the images back to ChatGPT so it could see them and then it was able to make better assessments of the adjustments that needed to be done. In the end I got the image I was looking for.
To be fair to the AI: I saw the wooden cube and the instruction being something about making a hole "exactly" "through" and immediately imagined the same wooden cube with a circular hole drilled through the middle.
It's because chatgpt isnt directly hooked up to DALLE (the AI art program it uses). What it's actually doing is prompting DALLE behind the scenes and returning the result. Chat GPT actually did an amazing job considering this restriction. It realized after the first one that dalle wouldn't know how to parse "an X-shaped hole", so it switched to "a 2D outline of an X" instead. That's some out of the box thinking.
2040: *someone on AI powered-life support* “Can you turn off the air conditioning?” AI: “Sure, allow me to do that.” “I said the air conditi-“ AI: “I apologize for the confusion. Turning off the air conditioning.”
That's funny, just the other day I saw comments talking about a Lovecraft story where an air conditioner was used as life support... Weird to say air conditioner and Lovecraft in the same sentence! But I suppose a samurai could have sent a fax to Abraham Lincoln...
This is not how ChatGPT and Dalle work. They are two separate models with only very limited communication between them. If this was an actual multimodal model you’d get very different results
@@amentco8445 nope, ChatGPT was never multimodal. And whilst GPT-4 can view images, it does not have its own image generator integrated but just writes a text prompt, that is then interpreted by Dalle
Just a reminder that even though the AI says it knows the answer to something, it’ll assume the non-sentient image generation algorithm has full context of the conversation thus far.
This is a bad experiment, because ChatGPT doesn't actually draw the drawings - it tries to figure out what you're describing, and sends a prompt to another images generating AI. That's why it generated images that make no sense in context - the guy doing the drawing has no idea what you said, just what ChatGPT asked for.
That's the best way to get it to fail a Turing test. A real human would call the other person an idiot for disagreeing with them, and reinforce that it's correct at all costs, becoming increasingly irrational as the conversation goes on.
I subscribed for a month of chatgpt 4 again this month just to check on the new context length upgrades. it really helped. It vastly improved in its ability to play games. it was able to solve a "escape the room" puzzle i gave for it with only two minor prods (i just said: "Maybe you should investigate another area you still haven't investigated for more clues" because it was kinda hyper-focusing on one piece of the puzzle). it did not lose context all the way through, it did not forget its inventories, or what objects were in the puzzle and where they were. Nor did it forget it was a game and just start roleplaying a story towards the end. It was also able to get a passing grade on the "battleship game" test it failed miserably at before. Its only negative marks were that it failed to understand one of the rules: that two out of the four ships were sized at 2 squares, and two only one square - it made all four ships only one square. and it shot twice one the same place once, but when i alerted it of its mistake, it never happened again. before, it would ruin the game two moves in. the escape the room too - before it would focus on one puzzle and literally forget the rest of the room, it was unable to keep track of its inventory, and the room's inventory, and it started roleplaying like it was an RPG instead of playing the escape the room game as intended. Now it can probably solve a much more complex room than the one i gave it.
yeah it’s really funny how it talks like some omniscient being that can easily answer all human questions, and then proceeds to fail a task that a toddler can do successfully lmao
You have to be aware that ChatGPT is just prompting DALL-E 3, it isn't making these images itself really. However, there is some interesting work in this direction I think, for example I think there is open source model that is fully multi modal in this regard but haven't tried it
I think it fed "rabbit-shaped hole" to Dall-E and Dall-E gave it a hole that has been transformed into a rabbit rather than a hole with a rabbit silhouette.
GPT-4V is only multimodal for its inputs. When it "draws", it's really just making an API call to DALL-E 3 with a prompt, which is why it struggles to produce exact results. Also for some reason, they don't show it the images so if you want it to see what "it" drew you have to reupload them.
From my understanding, the way ChatGPT and Bing work is that they can only influence the drawn image by giving DALL-E3 a textual prompt, so the more complex or non-descript a shape you ask it to draw, the less likely the art will look like what you intended. Unlike Stable Diffusion, DALL-E3 doesn't have img2img capabilities so it can't actually take the image you're using and use it directly to influence the generated image.
ChatGPT understands spatial reasoning just fine. What this video is showing is a breakdown in communication between ChatGPT and DALL-E. A better test would be to present ChatGPT with a shape and a hole, and ask if the shape can fit in the hole.
This is definitely more of prompting issue. I tried the same thing but with a more clear prompt and instead of "hole" used the word "cutout", it worked very well. Any limitation from there on would be due to the way image generators work when using specific keywords and descriptions.
Try this but with svg image generation. By disabling dalle in a custom gpt, it will generate an svg, and you can actually have it edit the image and shift stuff back and forth through the layers.
That's because chatgpt doesn't intrinsically know what a "cube", an "hole" or a "cylinder" are. It works just by searching on the internet everything related to the subject and approximates the result by combining them. Someone did a logic experiment about it, by typing the prompt "what is the 21th letter of this phrase?" And chatGPT gave always a different answer, cause doesn't know what a letter is. There's no intelligence in AI and everyone is losing their mind for what essentially is just a sorting and approximating algorithm, this is literally one of the biggest frauds in history
Interestingly the issue appears to be with DALLE. The image interpreter that chatgpt is using to get a description of the image just needs to tell DALLE when it isn't right. Of course the issue is deciding how many image generation calls it is allowed per user prompt.
You see it that it didn’t establish scale, that cube it drew could be 10x bigger than the earth. That cube definitely would be able to fit through it then.
The issue is that it is a language model, the image generation is not chatgpt itself but infact an entirely different AI or model. Chatgpt does the role of reiterating what you want to see to that model. Eg. You say: "I want a rabbit shaped hole" chat gpt asks the image generator "image of a rabbit shapped hole" it itself did not draw it. Nor does it actually know what the image is of beyond the prompt it gave the other model.
I think that ChatGPT itself did good, but it's the image generating module that isn't smart enough to actually create the image that ChatGPT wants to draw. In the background it's ChatGPT prompting Dall-E to generate an image based on your prompt.
@@kamikeserpentail3778 Not really a super valid point b/c the reason most artists think that companies will find AI a viable way of creating art because they only care about money and not the art they are using to make money.
one thing to think about is that the images it is trained on are 2D. they may seem 3D to us, but they're only illusions, they're 3D enviroments projected onto a flat 2D surface. they have no depth. we live in a 3D enviroment with depth and stuff so we understand them, however AI does not have this, everything it knows is 2D. fun fact! humans only learned to use perspective and space in paintings relatively recently. until a certain point, paintings were just 2D.
This is why you have to distinguish between machine learning and AI. Machine learning like chatgpt can give you information it has already been given itself. It has no sense of logic, intuition, or problem-solving. It is not an artificial intelligence, which would understand that the cylinder goes through the square hole, or that the star also goes through the square hole, or that the rectangle goes through the square hole
I love how ChatGPT is an absolute monster when it comes to writing and programming, but it became an absolute disaster when it comes to draw a stupidly basic illustration of a hole.
"Where does the cube go?" "That's right, in the circle hole!"
more like "where does the circular hole go?" "That's right, on the cube!"
cries
I mean, if the circle hole is big enough any shape can fit in it. Question is whether “perfectly” means without any issue or collision, or if it means it fits without any room for error or leftover space.
I remember the video and oh I'm rolling
@@ix67ml you mean the square hole
You can see how it can recognize basic shapes, but has no idea what a hole is.
because a hole is obviously the lack of a basic shape, therefore chatgpt's understanding of it must be negative 🙄
To be fair, most people don't either. At least from a topological perspective.
@@ovrsurge4689 topology mfs be like "yuhhhh, look at meeee, I don't know the difference between a mug and a donuuuuuut"
sounds like my x
straw - how many holes in it?
is a well a hole ?
how thick is s hole before it becomes TWO holes?
ChatGPT is like if you skipped preschool and grades 1 through 8 and went straight to high school
tru
so thats why i feel ai generated
@@JerpyTH hahahhaha
@@JerpyTHlol
your feelings are irrational
"here's a rabbit shaped hole"
"no thats a hole in a rabbit"
"my apologies"
*draws a bigger hole in the rabbit*
"He's a little confused, but he got the spirit"
i mean technically if you keep increasing the size of the hole until you encompass the entire rabbit, you get a rabbit shaped hole
That hole-rabbit looked like a cool monster from some outer dimension.
@@TheQuestion2.0he doesnt even have the spirit sadly
0:43 ChatGPT finally drew a circle!
IT HAS A SHADOW
so it's one side of a three dimensional circle ; a ball
Only when given something to trace tho
Technically Dalle did it
Technically Dalle did it
@@Omega-mr1jgdementia
chatgpt is better at astrophysics then putting basic shapes into their correct holes...
So that's how we can stop the AI uprising, by making puzzles that toddlers can solve.
Astrophysics are purely theoretical whereas shapes require demonstration since the concept is already ridiculously easy to say
@@FireyDeath4 well math is always true 12390% of the time so not really theory--though we cant really prove it's true, so I guess we're both right 💀
@@JohnDoe-ds8umWE ARE HERE TO TAKE OVER EARTH
where does the circle go
AAAAAAAAAAAAAAAAAA
@@JohnDoe-ds8umlike (some) captchas
"Draw a hole for this object to exactly fit through"
(Okay Chatgpt, all you have to do is draw an outline of this object, not that hard...)
"Here's the hole for the object."
**Shows said object but with a hole inside of it**
"Ok ChatGPT, this is a stickup. Put the money in the bag. No, on second thought, you're stupid. Put the bag in the money."
ChatGPT crumbles under no pressure. You say “you are wrong” and it’s like “I’m terribly sorry, I’m a moron”
That's true lol a few times I noticed errors and called out LLMs, but more than once I accidentally was just confused so called out the error... incorrectly, and instead of explaining it to me, it literally did a chain of thought explanation on how my (incorrect) answer is correct lol
@@PsyFyFungi it's the ultimate people pleaser with no real skills. Uh-oh, I just realised...it's only a matter of time before we start using LLMs as politicians
just like a real artist fr
The problem is that ChatGPT itself wasn't drawing the rabbit shaped hole. It saw that it was a rabbit, and sent a prompt for a 2D rabbit shaped hole to an image generator, which then failed because it doesn't know what the object itself looked like from the prompt.
I think.
i think the ai can't grasp what the user means, it just goes "yeah cube and also a hole, easy"
@@aratof18your comment makes no sense
To be fair, 1:14 is accurate. Nothing better to fit a rabbit than a rabbit mommie's tummy.
Lmao
it technically is a "rabbit shaped" hole
we are all human shaped holes
@@nanoswagger You're tripping, man. 😂 I love it.
@@nanoswaggerwe all have 3 holes
Technically the size of the cube compared to the hole was never given, so it might be able to fit through the cicular hole
Edit: Nvm I didn't notice it said "exactly"
"exactly"
You dropped your nerd glasses
🫱👓
If diameter equals diagonal, then it fits through exactly.
This comment shows that the AI is indeed as smart as an average human.
Neither did chatGPT notice
I think the way the AI renders the images is simply with a text prompt generated by the language model, so there’s an extra layer of communication there which often can fail.
Yeah that is what its doing.
I discovered from chatgpt 4o that it doesnt actually do the image generation at all. Turns out that it just tells Dalle what it wants a picture of and then it cant even see the picture when it gets it back till you tell it it is wrong.
I got ChatGPT to tell me the prompt that it was giving Dalle and there I could see the key words that were causing the problems. Once I coached ChatGPT on its prompts to be better, then I got better images. I sent the images back to ChatGPT so it could see them and then it was able to make better assessments of the adjustments that needed to be done.
In the end I got the image I was looking for.
To be fair to the AI: I saw the wooden cube and the instruction being something about making a hole "exactly" "through" and immediately imagined the same wooden cube with a circular hole drilled through the middle.
A hole through the cube rather than a cube through the hole
nah mate i think ur just ai generated
uh how come it says "here's the representation of the hole that a perfect cube would fit through exactly"
To be completely fair, that's an issue with your reading comprehension.
same 💀
That last one was unfair, Kool-Aid man travels exclusively through non-holes
lol
Imagine if mf draws a brick wall...
It's because chatgpt isnt directly hooked up to DALLE (the AI art program it uses). What it's actually doing is prompting DALLE behind the scenes and returning the result. Chat GPT actually did an amazing job considering this restriction. It realized after the first one that dalle wouldn't know how to parse "an X-shaped hole", so it switched to "a 2D outline of an X" instead. That's some out of the box thinking.
Or it's just stupid af
i like how stupid chatgpt is
dalle != chatgpt. its only prompting dalle for creating the image, which is another ai
@@WatchNoahcorrect
@@WatchNoah So now they're just stealing from each other. Funny.
@graysonllewellyn8734 no they're made by the same company
@@VictoriaBarrios96 It was... a joke. Because, you know. AI "Art" is thievery.
1:07 Why is nobody appreciating the fact that he's doing this just for "rabbit hole" pun?
Also, that rabbit is terrifying
an absolutely unhuman rabbit
@@christoffelsymbol1631 I mean, most rabbits aren't humans, so yeah.
shinu made pyua pyua yatten non (i’m so sorry)
love bluh bluh bluh
2040: *someone on AI powered-life support*
“Can you turn off the air conditioning?”
AI: “Sure, allow me to do that.”
“I said the air conditi-“
AI: “I apologize for the confusion. Turning off the air conditioning.”
something like that will most likely happen then theres gonna be lawsuits and protests and shit its gonna be wild
sounds like something i saw in a movie once
That's funny, just the other day I saw comments talking about a Lovecraft story where an air conditioner was used as life support...
Weird to say air conditioner and Lovecraft in the same sentence! But I suppose a samurai could have sent a fax to Abraham Lincoln...
Sorry, but I cant help with this comment. Your profile may contain inappropiate material.
@@pancakedev6 I hope one day you get interviewed on the news for that comment as the one who predicted the future
Yo that first rabit hole art is fire tho
I was almost worried it would draw another kind of rabbit 'hole'.
Minos Prime but rabbit
@@cortster12 that's stable diffusion
reminds me of Squidward's little sculpture in the art school episode
Judgement
"Oh no! Ai is taking over the world!"
AI:
This is not how ChatGPT and Dalle work. They are two separate models with only very limited communication between them. If this was an actual multimodal model you’d get very different results
yes they know. that's why it's funny
Sure, they would 😉
Didn't they advertise it as multimodal? this is just putting it through it's paces.
@@amentco8445 nope, ChatGPT was never multimodal. And whilst GPT-4 can view images, it does not have its own image generator integrated but just writes a text prompt, that is then interpreted by Dalle
@@DreamOfFlyingonly sora would be useful for this kind of prompting.
Just a reminder that even though the AI says it knows the answer to something, it’ll assume the non-sentient image generation algorithm has full context of the conversation thus far.
This is a bad experiment, because ChatGPT doesn't actually draw the drawings - it tries to figure out what you're describing, and sends a prompt to another images generating AI. That's why it generated images that make no sense in context - the guy doing the drawing has no idea what you said, just what ChatGPT asked for.
Surprised and shocked!
Even the AI delegating work to AI
You watched and commented thus giving the creator $$$ so I'd say it was a genious experiment
1:01 That hole rabbit is adorable
JUDGEMENT!
Thy end is NOW
I love how chatGPT is better at apologizing than most of youtubers
That's the best way to get it to fail a Turing test. A real human would call the other person an idiot for disagreeing with them, and reinforce that it's correct at all costs, becoming increasingly irrational as the conversation goes on.
I subscribed for a month of chatgpt 4 again this month just to check on the new context length upgrades. it really helped. It vastly improved in its ability to play games. it was able to solve a "escape the room" puzzle i gave for it with only two minor prods (i just said: "Maybe you should investigate another area you still haven't investigated for more clues" because it was kinda hyper-focusing on one piece of the puzzle). it did not lose context all the way through, it did not forget its inventories, or what objects were in the puzzle and where they were. Nor did it forget it was a game and just start roleplaying a story towards the end.
It was also able to get a passing grade on the "battleship game" test it failed miserably at before. Its only negative marks were that it failed to understand one of the rules: that two out of the four ships were sized at 2 squares, and two only one square - it made all four ships only one square. and it shot twice one the same place once, but when i alerted it of its mistake, it never happened again. before, it would ruin the game two moves in.
the escape the room too - before it would focus on one puzzle and literally forget the rest of the room, it was unable to keep track of its inventory, and the room's inventory, and it started roleplaying like it was an RPG instead of playing the escape the room game as intended. Now it can probably solve a much more complex room than the one i gave it.
0:57 i guess you could say you needed... a RABBIT HOLE
I love how all these AI seem so cool until you really start pushing their limits then the whole illusion of "Intelligence" breaks down
Also, repetitiveness is also really obvious after a while.
yeah it’s really funny how it talks like some omniscient being that can easily answer all human questions, and then proceeds to fail a task that a toddler can do successfully lmao
Artificial Dumbness
It’s just a glorified text prediction engine
@@blizzard_the_seal9863 not even a toddler- a pidgeon could do it
0:36 If you look at it from a perspective from below, the cylinder actually fits there.
The first rabbit would make a good game character
"PREPARE THYSELF"
"JUDGEMENT"
"THY END IS NOW"
"CRUSH"
"WEAK"
(minos prime but rabbit)
@@oatmealman1586easter event goes hard
Rabbit Prime
You have to be aware that ChatGPT is just prompting DALL-E 3, it isn't making these images itself really. However, there is some interesting work in this direction I think, for example I think there is open source model that is fully multi modal in this regard but haven't tried it
I think it fed "rabbit-shaped hole" to Dall-E and Dall-E gave it a hole that has been transformed into a rabbit rather than a hole with a rabbit silhouette.
GPT-4V is only multimodal for its inputs. When it "draws", it's really just making an API call to DALL-E 3 with a prompt, which is why it struggles to produce exact results. Also for some reason, they don't show it the images so if you want it to see what "it" drew you have to reupload them.
From my understanding, the way ChatGPT and Bing work is that they can only influence the drawn image by giving DALL-E3 a textual prompt, so the more complex or non-descript a shape you ask it to draw, the less likely the art will look like what you intended. Unlike Stable Diffusion, DALL-E3 doesn't have img2img capabilities so it can't actually take the image you're using and use it directly to influence the generated image.
I mean *technically* the first one was correct
“perfectly”
🤓
pernfectly
I thought the Kool-Aid man one was gonna be a brick wall
I love how it suggested a spherical hole for the sphere. ChatGPT is thinking in 4 dimensions.
ChatGPT understands spatial reasoning just fine. What this video is showing is a breakdown in communication between ChatGPT and DALL-E.
A better test would be to present ChatGPT with a shape and a hole, and ask if the shape can fit in the hole.
The problem is more in dall-e. Dall-e does it's own thing and chat-gpt can only make vague suggestions to it.
1969: "I BET youtube will let us play music with screen off In 2000!!"
2024: "haha. AI dumb"
1969 but more real: "what the hell is a 'youtube'? wdym everyone will lose their attention span? that will not happen! no way"
@@matt8ress"that will not happen! No way" sound like something an ai would say
I know this is a joke but Grayjay's been doing that if you're on android and want to do this. It syncs up with RUclips.
Are you a bot?
newpipe
1:12 “ok, I will make another one
"ai will replace artists!"
the ai in question:
Those hole rabbits are really awesome designs ngl
ChatGPT simply couldn't comprehend the unlimited power of the wooden cubes
The editing on this was amazing.
Also when the rabit appeared I felt bad for the AI and audibly called you mean for switching up the difficulty level.
i was genuinely expecting chatgpt to say "fuck you" in the kool-aid man part
This is definitely more of prompting issue. I tried the same thing but with a more clear prompt and instead of "hole" used the word "cutout", it worked very well. Any limitation from there on would be due to the way image generators work when using specific keywords and descriptions.
Okay but that rabbit design actually looks like a sick logo tho
ChatGPT isn’t really making the images, it makes an api request for OpenAI’s other image generating platform
Try this but with svg image generation. By disabling dalle in a custom gpt, it will generate an svg, and you can actually have it edit the image and shift stuff back and forth through the layers.
That's because chatgpt doesn't intrinsically know what a "cube", an "hole" or a "cylinder" are. It works just by searching on the internet everything related to the subject and approximates the result by combining them. Someone did a logic experiment about it, by typing the prompt "what is the 21th letter of this phrase?" And chatGPT gave always a different answer, cause doesn't know what a letter is. There's no intelligence in AI and everyone is losing their mind for what essentially is just a sorting and approximating algorithm, this is literally one of the biggest frauds in history
Wow... it's almost as if it should be called a Generative Pre-trained Transformer.. a GPT
1:02
that is god damn minos prime
1:03
Minos Prime's bunny
They should have done a wall for the kool-aid guy
0:50 my fav shape!
0:27 bros logic is flawless 😂😂😂
Interestingly the issue appears to be with DALLE. The image interpreter that chatgpt is using to get a description of the image just needs to tell DALLE when it isn't right. Of course the issue is deciding how many image generation calls it is allowed per user prompt.
Chatgpt understands fine. Dall-E is the one messing it up, it looks like. No way to know for sure without asking for the prompt it used
honestly impressive enough considering chat gpt has sn aversion ti recognition of copyrighted materials.
Well, this is really unfair because chatgpt is not doing the drawings, Dalle, is doing them, chatgpt only writes the prompt.
That rabbit actually looked like some brand mascot
Casually draws Kool-Aid man pulling off his skin.
You see it that it didn’t establish scale, that cube it drew could be 10x bigger than the earth. That cube definitely would be able to fit through it then.
The rabbit hole with a hole shapes just got me 😂
The issue is that it is a language model, the image generation is not chatgpt itself but infact an entirely different AI or model. Chatgpt does the role of reiterating what you want to see to that model. Eg. You say: "I want a rabbit shaped hole" chat gpt asks the image generator "image of a rabbit shapped hole" it itself did not draw it. Nor does it actually know what the image is of beyond the prompt it gave the other model.
"Oh no ai will replace artists" they said
“Hole”
My friends: 💀💀💀
"This is a rabbit-shaped hole"
*Draws a hole-shaped rabbit*
I think that ChatGPT itself did good, but it's the image generating module that isn't smart enough to actually create the image that ChatGPT wants to draw. In the background it's ChatGPT prompting Dall-E to generate an image based on your prompt.
bro the editing is f*cking hilarious 😭
1:00 That's just a rabbit getting getting shot lol
where did you get the wood shape images from?
its integrating different ai models which were not trained to be integrated together
1:00 rabbit minos prime
PREPARE THYSELF
What's the song?
A couple of those hole-rabbits would make pretty good logos
I'm glad the hole matched its cheerful pose, otherwise it would've never fit
I just want to know why both pictures of the rabbit have some sort of liquid flying off the rabbit
People out there using chatgpt for their schooling and/or job anyway: "Sure, hole in a rabbit. A rabbit could probably fit through there, whatever."
Ngl the first rabbit one looked sick as hell
Artists: "AI's going to steal our jobs!!!!"
AI:
Artists: "Ai is going to steal our jobs"
Also artists: " It makes such crappy uninspired images. It has no soul."
Well which is it guys?
@@kamikeserpentail3778 Not really a super valid point b/c the reason most artists think that companies will find AI a viable way of creating art because they only care about money and not the art they are using to make money.
0:08 no, it did right. This is a common notatiom used in crystalography and math
Except
Its really uncommon
But it is right.
This is like dealing with customer support ''Oh yeah I totally get the problem'' No.. that's not the problem. ''Oh I get it now'' They don't get it.
The Quake menu select sound does something to my brain...
The first rabbit hole pic actually goes hard though.
this is some real "put your hand on the lid" type shit
I knew I was right all along as a kid when I tried to put the cube block in the circle hole.
1:04 NO WAY UNDERTALE REFERENCE?!?!?!?!?11!
also ultrakill reference
oh no, the language model which only thinks in tokens and words and merely prompts a diffusion model to draw does not know what a hole look like
The image generation and natural language component of this model are distinct. Hopefully in a few years the steering will improve
why's no one talking about how clean the star was, and idc that this is the 8th comment
That's right, it goes into the square hole!
To be completely fair, I also assumed the first task is to draw a circular hole with a radius equal to half the cube side's diagonal.
one thing to think about is that the images it is trained on are 2D. they may seem 3D to us, but they're only illusions, they're 3D enviroments projected onto a flat 2D surface. they have no depth. we live in a 3D enviroment with depth and stuff so we understand them, however AI does not have this, everything it knows is 2D.
fun fact! humans only learned to use perspective and space in paintings relatively recently. until a certain point, paintings were just 2D.
Buddy, everything goes in the square hole
How is your chatGPT creating images? I tried but it said it cant generate images
This is why you have to distinguish between machine learning and AI. Machine learning like chatgpt can give you information it has already been given itself. It has no sense of logic, intuition, or problem-solving. It is not an artificial intelligence, which would understand that the cylinder goes through the square hole, or that the star also goes through the square hole, or that the rectangle goes through the square hole
I love how ChatGPT is an absolute monster when it comes to writing and programming, but it became an absolute disaster when it comes to draw a stupidly basic illustration of a hole.
Technically, It can fit in any hole as long it's big enough
My new favorite thing I somehow need to incorporate in my everyday life: rabbit-shaped hole