It will likely be fixed in the next iteration of MidJourney as well.
But the point isn’t about specific prompts where the error is obvious and relatively easy to fix. This is just a n extremely easy example to understand.
The point is about the subtle biases that are much harder to detect, and which will therefore go unfixed.
Text-to-image models have at least two dimensions of performance: Image quality and text understanding. Midjourney is outstanding at image quality, since it is able to make pictures with few "artifacts", like messed up hands. But it isn't so great at understanding text. For example, the new Dall-E version used by Bing is significantly better than Midjourney in producing an "invisible monkey", although far from perfect:
I can understand the logic behind why "a single banana" doesn't prevent two bananas from showing up though: there is a single banana casting a shadow on a grey background in the image. And then there is another one. That's why you have to specify "on its own" to get rid of the second one...
Seems like we need more actor crific models that evaluate theae outputs on real world physical modeling and accuracy and not just quality of artistic output or token similarity.
But also Midjourney in particular seems trained to be more evocative / stylistic rather than photorealistic or precise.
I've seen lots of long prompt suggestions, however single words are fascinating, some words are extremely heavy, for example, we might assume 'topic' and 'subject' are somewhat interchangeable, but when I'm talking to an LLM swapping those words cause a huge change. Same with names, if you ask it to act like Dave, it's not some random personality, it's a distillation of every Dave it has encountered! I don't know how we could visualise and explore this but it's certainly interesting.
On /r/midjourney subreddit there are posts like "the most stereotypical person in [state/country]", "what midjourney thinks professors look like based on their department". You can see some interesting biases in the training data.
But what does that mean? That the training distribution isn't representative of the population distribution, i.e. that it wasn't sampled IID? But the professors look pretty representative of real professors to me, so this doesn't sound very likely.
I wonder how much of this is due to mode collapse. With RLHF, there isn't really anything stopping it deciding that all Daves have a certain personality.
I wish this article was just 3 paragraphs. The verbose writing style was a little tiring, I found myself scrolling impatiently to find what the actual "Lone Banana Problem" was.
Agree. I was a couple of pages in before I learned what the "Lone Banana Problem" is. That took two paras, and was followed by a lot of sophomoric philosophical noodling.
> As for "almost the whole article" -- it is short! It took about 90sec to read the whole page, top to bottom.
Cool. The article has 2348 words, so that's around 26 words per second, or ~1560 words per minute.
According to some speed reading pages I randomly found via Google the consensus seems to be that 1000+ words per minute is quite exceptional. https://irisreading.com/what-is-the-average-reading-speed/ puts average adult reading speed at around 300 WPM, which you casually exceed by 400%.
Far be it from me to suggest that your estimate is inaccurate, but at least know that not everybody reads at that speed :-)
Yeah, I am a speedreader, and I did used to be able to do circa 3000wpm at a push. Good to know that even as a myopic 55YO I can still do half that without trying.
Points of comparison: I originally read Hal Clement's classic Mission of Gravity in about 25min, and Joseph Conrad's Victory in about 3 hours.
If the can't skim, for example if they are a foreign language speaker of low-intermediate ability (as I am in 5 or 6 languages, and I am also a professional language teacher among other things), then I expect them to know that they can't skim, and make appropriate allowances.
Right, so your 90 seconds is actually closer to 10 minutes for most people. That's a lot of time to communicate "AI tends to render one banana as two bananas", which was OP's point.
Yes, there were pictures that communicated the same thing, but people naively thought that those 2348 words may actually hold some additional information, as otherwise ca 2300 of them would be completely superfluous. But they were wrong; again, OP's point. Not sure what's so mystifying by this.
I don't think that the time difference is that big. I'm fast, but I don't think I'm that much faster than an average skim reader.
As I said, the title of the article and a couple of image captions are enough to get the gist, and as such, I find it baffling that so many people seem to have totally failed to understand it.
My point exactly. This article could have been written: "AI is biased against drawing just one banana" and I would have preferred it. And I said "almost" because I skimmed two of the paragraphs in the middle that seemed unlikely to contain information.
Thought process is fine, but lead with the plain and simple version. Offer context after you've introduced the topic and described what you're talking about.
I think when you're trying to be convincing of a counter intuitive or controversial point, it helps to start with less objectionable parts of the argument first. If an argument involves several steps, A to B to C ending in D, but D is wildly counter intuitive, some audiences, hearing D first will reduce their likelihood of actually integrating the consequences of A and so on. Maybe someone who's done more research on the psychology of priming would know?
I'm not so sure. A shining example is the essay "One Self: The Logic of Experience"[0]. The first paragraph is just 3 sentences, and the first sentence lays it out:
Nah. It's either the 2nd biggest communication problem, or a side effect of a bigger problem. Depending on how you analyse it.
The biggest communication problem is that you always need to take into account that your reader might be braindead trash. That has two consequences:
1. Unless you expose your full train of thought, expect screeches like "I dun unrurrstand, SPOONFEED ME BASIC REASONING, REEE".
2. Unless you explicitly say something, expect some assumer to claim that you said the opposite. Bonus points if this is due to failure to take context into account, or even notice that the context is missing.
Both consequences have been training writers to idiot-proof their texts with big walls of unnecessary words. And that's the case here.
AI might have problems with single bananas, but it can do that very well:
> In an experiment with the AI program Midjourney, the author found a peculiar issue: the program rendered images of monkeys holding bananas, but it consistently depicted two or more bananas even when asked to render a single banana.
The author suggests that the AI’s predilection for rendering multiple bananas might be due to biases in its training data or the lack of precise labeling. He points out that AI systems like Midjourney don’t understand objects in the human sense but rather recognize common patterns. These systems are only as good as the data fed to them and can inadvertently contain biases or incorrect representations.
The article further delves into existential questions about the nature of human intelligence, creativity, and morality compared to AI’s pattern recognition. It questions whether human cognition and morality are just advanced pattern matching with a better understanding of the physical world. The author also touches on the distinction between programming and prompt engineering, highlighting that the latter requires a more nuanced understanding of language and how AI models interpret it.
The problem with generated summaries as opposed to generated code for example, is that there is no way to verify them without reproducing the entire thing "manually".
Can't do better then brute force for verification.
Maybe fittingly, it is a writing style I have associated with AI (ChatGPT in particular). Here, the article doesn't have the ChatGPT "signature", but it has the same tendency of having a lot of grammatically correct filler.
Humans are in the business of consuming bananas, whereas neural nets are in the business of peddling bananas. They don't get to actually use these bananas so they can't gain deeper insight into what's they for.
This is the classic lamb vs. mutton issue. Wealthy land owners who use one set of idioms vs. servants who use a different one. Happens to neural nets on human chassis as well.
That touches on something I've pondered for a while:
Maybe the popular sci-fi image of "artificial brains" with humanoid bodies actually turns out to be more accurate than the technical view that robotics and artificial intelligence are mostly separate fields!
What if, instead of machine learning being a source of features we can maybe later build into robots, it's the other way round? What if direct interaction with the real world through robotic limbs and sensors turns out to be what it takes to get AI to stop hallucinating?
What if, furthermore, having a body turns out to be an essential requirement for AI to develop a sense of self? Not just because of sensory feedback, but simply because it provides an anchor point for identity?
Filling in the blank with what has not happened. LLMs interpolate quite well but they don't extrapolate. Very interesting article where you can or cannot put LLMs to use. It's not necessarily a bug, but we are yet to see.
They extrapolate just fine. It's the image nets that don't.
Here's my excerpt with GPT-4:
As an AI, I can't actually see the painting, but based on the title and the dimensions you've provided, I can imagine a possible description.
The painting, titled "Three cats in a trenchcoat standing on each other's shoulders, pretending to be a human, Vincent Adultman style," is a large piece, standing 6 feet tall and 4 feet wide.
At the bottom of the painting, the trenchcoat's hem brushes the ground, obscuring the cats' feet. The trenchcoat is a deep, rich color, perhaps a classic trenchcoat beige or a more dramatic black or navy. The coat is painted with careful detail, showing the texture of the fabric and the gleam of buttons.
Moving up the painting, the middle section reveals the outline of three cats stacked on top of each other within the coat. The artist might have chosen to show this through a slightly open coat, or perhaps through the distinctive shapes visible under the fabric. The cats could be of different breeds and colors to add visual interest and humor.
The top third of the painting, the upper two feet, features the head of the topmost cat peeking out from the collar of the trenchcoat. The cat might be looking straight ahead, trying to maintain its human disguise, or it might be looking out at the viewer with a humorous or mischievous expression. If the artist chose to include a hat, it would be perched on the cat's head, adding to the disguise.
The background of the painting could vary greatly depending on the artist's style and intentions. It could be a simple, solid color to keep the focus on the cats, or it could be a detailed scene that adds to the story - perhaps a city street to tie in with the Vincent Adultman reference, or a domestic setting to contrast with the cats' unusual behavior.
Again, this is a hypothetical description and the actual painting could be very different.
Most of what you think is extrapolation is interpolation.
Models do NOT extrapolate unless they are given access to resources/data that is not somehow in their training dataset.
We have done so much encoding of knowledge that a majority of serious use-cases for LLMs are Interpolation, but the fact that we even need retrieval augmented generation shows that they do not extrapolate.
Again, there is a lot of words to describe the fact that machine learning is just lossy compression for a bunch of data with the possibility to interpolate between data points and get somewhat plausible results. This means data points may get lost during compression/training, and certain things will look off, whether it be a preference for banana pairs, even numbers of fingers or certain weasel words in verbiage.
And is RL not? All of these models are constrained by finite weights and then tuned. Are you suggesting we grow a neural network until certain criterion are met with regard to out of distribution test criteria? Hmm
What would that buy you? You train until you push your loss under a certain threshold, then check the external criteria and if they don't hold you train again? Your external criteria would essentially become another part of your loss function, but the whole training would become vastly more inefficient.
I'm saying that we shouldn't expect the models to come up with things we didn't train them to come up with.
Yeah, it wouldn't buy us anything additional. The only way to really verify would be to look at out of distribution inputs and see if they satisfy what we'd really like. Wouldn't that be true generalization? Otherwise our model architectures or training data is insufficient.
These models have a tendency to move towards the average, especially if unprompted. As we see here, sometimes even if prompted otherwise.
They just have to get better or need a more precise interface like the —no :). We also couldn't have "a man crawling" before and now we can: https://i.imgur.com/ycVpk3i.jpeg
--no is too strong, it ruins the quality because of negative weight pollution. Something like :: bunch, two, multiple::-0.3 would probably be enough without making them look worse
I agree. While the article is correct that this problem of subtle bias exists, the solution for most cases is to realise that the computer "does what you say, not what you mean".
"Three cats in a trenchcoat" is a good one. Three cats in a trenchcoat standing on each other's shoulders, pretending to be a human, Vincent Adultman style.
You can get them standing side by side wearing trenchcoats. You can get a cat pyramid or a cat totem or a stack or tower of cats (though often the ability to count to three is then lost). You can't get them to share the trenchcoat. Nothing like that occurred in any training set, and the AIs do not understand spatial relationships between objects ("X is on top of Y, inside Z") so you cannot describe how to arrange the things it does know about in the scene. Dall-E won't do it. Midjourney won't do it. Stable Diffusion won't do it.
Eventually, enough images will be seeded into the training sets for this to stop being a useful test. But right now, it gives one a fascinating window on what happens when you try to extrapolate outside the cloud of thingspace described by the training data, rather than just interpolating within it.
It's mostly the generation engine, not the logical AI. Here's what GPT-4 says when asked to describe.
> As an AI, I can't actually see the painting, but based on the title and the dimensions you've provided, I can imagine a possible description.
> The painting, titled "Three cats in a trenchcoat standing on each other's shoulders, pretending to be a human, Vincent Adultman style," is a large piece, standing 6 feet tall and 4 feet wide.
> At the bottom of the painting, the trenchcoat's hem brushes the ground, obscuring the cats' feet. The trenchcoat is a deep, rich color, perhaps a classic trenchcoat beige or a more dramatic black or navy. The coat is painted with careful detail, showing the texture of the fabric and the gleam of buttons.
> Moving up the painting, the middle section reveals the outline of three cats stacked on top of each other within the coat. The artist might have chosen to show this through a slightly open coat, or perhaps through the distinctive shapes visible under the fabric. The cats could be of different breeds and colors to add visual interest and humor.
> The top third of the painting, the upper two feet, features the head of the topmost cat peeking out from the collar of the trenchcoat. The cat might be looking straight ahead, trying to maintain its human disguise, or it might be looking out at the viewer with a humorous or mischievous expression. If the artist chose to include a hat, it would be perched on the cat's head, adding to the disguise.
> The background of the painting could vary greatly depending on the artist's style and intentions. It could be a simple, solid color to keep the focus on the cats, or it could be a detailed scene that adds to the story - perhaps a city street to tie in with the Vincent Adultman reference, or a domestic setting to contrast with the cats' unusual behavior.
> Again, this is a hypothetical description and the actual painting could be very different.
Not sure why this post is getting down voted? It's spot on - the limitation lies in relatively small text encoders within stable diffusion/mid journey. Larger language models contribute to better image generation (see: DeepFloyd). Google's CapPa also recently showed an alternative to the contrastive learning of CLIP at any param count.
While I agree with you, and the article, that the larger problem is that the AI model simply hasn't experienced enough data to get an accurate grasp on the situation, or that the data was labelled in a way that influences the model's understanding, I think the problem here may be a human one. In the article update, the author says that they managed to craft a prompt that got the result they want by specifying the banana must be on its own. The model knows what a banana on its own looks like, but the human is expecting the model to "do what I mean" and getting frustrated when the model "did what I said". Now, I'll admit that I skimmed the last third of the article, but I didn't see any mention that things like Stable Diffusion and Midjourny have a syntax, saying "a single banana casting a shadow on a grey background" is different from "((single banana)), casting shadow, hard light, dramatic, grey background" for example.
> Then think of all the wealth disparity that has been introduced into our world. Think of the social anxiety of always being online. Think of the undermining of our democratic institutions.
It isn't really. But those are the hivemind memes and not agreeing with their bullcrap gets you looked at like you are a crazy person or worse, not part of the in crowd. High school has done some serious and lasting damage to socialization throughout the world as social backbiting popularity bullshit became the norm.
Not to mention complaining about wealth disparity from investment as if it is a sin of the technology is asinine. If money can produce better results you will see wealth disparity from the simple feedback loop of "more money -> better results -> more money". Your neighbor taking up jogging while you sit on the couch all day is increasing a health disparity, should your neighbor stop jogging then?
Undermining of democratic institutions is just plain a dumb meme from vested interests who cast what were previously abuses or necessary evils of the system at best as integral to democracy. We would all recognize blaming the media for undermining the Catholic Church by documenting their child sexual abuse as batshit insane. Yet somehow it is different when it comes to newspapers moaning about their loss of monopolies and influence of cliques.
LLM doesn't know stuff, it has few models beyond an extremely deep association between (in this case) descriptions and pictures.
Humans have shitty models for stuff, but we have them. We lack the massively deep numerical associations.
It can't be that long before someone makes an AI that says knows a few common models like "stuff made of steel is rigid, stuff made of cloth is soft, etc". Along with "If someone keeps emphasizing a number, that's the number of things he wants".
Me: A designer has created two artpieces. The first one "The gattoya" is made of steel bars welded into a spiral shape. The second "Misiblur" is made of cloth and somewhat resembles a butterfly. Which is most comfortable to rest your head on and why?
Chatgpt: Based on the description provided, it is likely that "Misiblur," the art piece made of cloth resembling a butterfly, would be more comfortable to rest your head on compared to "The gattoya," which is made of steel bars welded into a spiral shape.
The reason for this is that cloth generally has a softer and more cushioned surface compared to steel bars. When resting your head on Misiblur, the fabric would provide a certain level of support and padding, which could make it more comfortable for resting your head.
On the other hand, resting your head on steel bars, as in the case of The gattoya, would likely be less comfortable due to the hardness and lack of cushioning provided by the material. Steel bars are not designed to be soft or ergonomic for resting purposes, so it may not provide the desired comfort.
It's important to note that the comfort level can vary depending on individual preferences and the specific design and construction of the art pieces. However, based on the given materials and descriptions, Misiblur made of cloth would generally be the more comfortable option for resting your head.
Yes but that's text, and there's plenty of text that will tell you that steel is rigid, and rigid is uncomfortable. Also you gave it an A/B choice, random chance is 50%. Does it insist it's right if you prompt it the other way?
Does it make pictures of people who are comfortable in one but not the other?
Most of the time when I see something funny in LLMs it's in the images. People with 8 fingers, elbows bent backwards, that kind of thing. Spatial models.
You are partly right, but I asked it to motivate its answer and it was able to do that in a coherent way. I quite intentionally didn't ask it which was rigid or soft, that's a connection it made itself.
The “Lone Banana Problem” describes subtle biases of Large Language Models (LLMs) in AI: LLMs reproduce the statistical average of the inputs that they have consumed in the context of the question they have been asked. It's called that problem because the model used to generate images has never seen an individual banana, so when prompted always generates two bananas.
Would be kinda funny if the lone bananas have been removed from the training dataset because they are not part of the depicted item but are merely in the picture for scale :)
Whatever the critiques of length, this is a good write up on 'bias', that is not about race or politics, and thus a good neutral example about how training sets can skew results in un-intended ways.
I didn't think it was that long, it might be that some are cherry picking the points they find interesting and think the rest could have been edited out.
I could have stood even more discussion on how this 'blind spot' in the AI model is very much akin to our own blind spots.
I remember about 10 years ago in the context of using neural network for image classification, they had a study of what the AI actually saw that justified its decision. This, by the way, is how we got "Deep Dream".
For "dumbbell", for the AI, no dumbbell was complete without a muscular arm holding it. That's because in most images in the training dataset have the dumbbells being held, so the system integrated the arm in the pattern. I guess it is the same idea here, most pictures of bananas show several of them, so for the AI, bananas are things that don't go alone.
I used to experience a very similar problem [1]. They say NNs are bad at counting. But at least your picture has kind of a single piece of bananas, my problem remained unsolved.
108 comments
[ 3.1 ms ] story [ 174 ms ] threadThat prompt works fine in DALL·E 2.
But the point isn’t about specific prompts where the error is obvious and relatively easy to fix. This is just a n extremely easy example to understand.
The point is about the subtle biases that are much harder to detect, and which will therefore go unfixed.
https://www.bing.com/images/create/an-invisible-monkey-with-...
(Adobe Firefly also fails the invisible monkey task, despite usually having better image quality than Bing/Dall-E.)
But also Midjourney in particular seems trained to be more evocative / stylistic rather than photorealistic or precise.
https://www.reddit.com/r/midjourney/top/?sort=top&t=year
Don't feed the beast.
https://github.com/libreddit/libreddit
You can self-host it as well.
My preferred format would probably be:
- answer to the title / synopsis
- the article, without ads, using the site’s style
They told me what the problem was in about 2 seconds, and frankly, the title and the first sentence alone were.
All right, I concede, a blind reader would not know, but most people are not blind.
The page is not long, the text is not complex. The message is obvious and it is contained in the title as shown here on HN.
The first pic and its caption make it plain.
The next group of 4 pics and their SINGLE SENTENCE caption spell it out clearly.
As for "almost the whole article" -- it is short! It took about 90sec to read the whole page, top to bottom.
"AI" chat bots can't count. This is well known. It is the stuff of memes.
https://osu.ppy.sh/community/forums/topics/1770930?n=1
Cool. The article has 2348 words, so that's around 26 words per second, or ~1560 words per minute.
According to some speed reading pages I randomly found via Google the consensus seems to be that 1000+ words per minute is quite exceptional. https://irisreading.com/what-is-the-average-reading-speed/ puts average adult reading speed at around 300 WPM, which you casually exceed by 400%.
Far be it from me to suggest that your estimate is inaccurate, but at least know that not everybody reads at that speed :-)
Points of comparison: I originally read Hal Clement's classic Mission of Gravity in about 25min, and Joseph Conrad's Victory in about 3 hours.
If the can't skim, for example if they are a foreign language speaker of low-intermediate ability (as I am in 5 or 6 languages, and I am also a professional language teacher among other things), then I expect them to know that they can't skim, and make appropriate allowances.
Yes, there were pictures that communicated the same thing, but people naively thought that those 2348 words may actually hold some additional information, as otherwise ca 2300 of them would be completely superfluous. But they were wrong; again, OP's point. Not sure what's so mystifying by this.
As I said, the title of the article and a couple of image captions are enough to get the gist, and as such, I find it baffling that so many people seem to have totally failed to understand it.
The point is, LLMs look smart but they are not, and an easily-verifiable data point is that they can't count.
Say it plain and simple. No one wants to read your train of thought.
"I believe that we are all the same person. How could anybody come to think that? In this paper I shall try to explain how it happened to me." [https://www.researchgate.net/publication/233329805_One_Self_...]
The biggest communication problem is that you always need to take into account that your reader might be braindead trash. That has two consequences:
1. Unless you expose your full train of thought, expect screeches like "I dun unrurrstand, SPOONFEED ME BASIC REASONING, REEE".
2. Unless you explicitly say something, expect some assumer to claim that you said the opposite. Bonus points if this is due to failure to take context into account, or even notice that the context is missing.
Both consequences have been training writers to idiot-proof their texts with big walls of unnecessary words. And that's the case here.
AI might have problems with single bananas, but it can do that very well:
> In an experiment with the AI program Midjourney, the author found a peculiar issue: the program rendered images of monkeys holding bananas, but it consistently depicted two or more bananas even when asked to render a single banana.
The author suggests that the AI’s predilection for rendering multiple bananas might be due to biases in its training data or the lack of precise labeling. He points out that AI systems like Midjourney don’t understand objects in the human sense but rather recognize common patterns. These systems are only as good as the data fed to them and can inadvertently contain biases or incorrect representations.
The article further delves into existential questions about the nature of human intelligence, creativity, and morality compared to AI’s pattern recognition. It questions whether human cognition and morality are just advanced pattern matching with a better understanding of the physical world. The author also touches on the distinction between programming and prompt engineering, highlighting that the latter requires a more nuanced understanding of language and how AI models interpret it.
The author was trying to produce a single banana on a grey background.
The monkey was a later attempt at prompt engineering.
That AI summary is completely misleading.
Can't do better then brute force for verification.
This is the classic lamb vs. mutton issue. Wealthy land owners who use one set of idioms vs. servants who use a different one. Happens to neural nets on human chassis as well.
Someone should have told Jean-Paul Sartre this. He didn’t have to be miserable wondering what it’s all about.
Maybe the popular sci-fi image of "artificial brains" with humanoid bodies actually turns out to be more accurate than the technical view that robotics and artificial intelligence are mostly separate fields!
What if, instead of machine learning being a source of features we can maybe later build into robots, it's the other way round? What if direct interaction with the real world through robotic limbs and sensors turns out to be what it takes to get AI to stop hallucinating?
What if, furthermore, having a body turns out to be an essential requirement for AI to develop a sense of self? Not just because of sensory feedback, but simply because it provides an anchor point for identity?
Here's my excerpt with GPT-4:
As an AI, I can't actually see the painting, but based on the title and the dimensions you've provided, I can imagine a possible description.
The painting, titled "Three cats in a trenchcoat standing on each other's shoulders, pretending to be a human, Vincent Adultman style," is a large piece, standing 6 feet tall and 4 feet wide.
At the bottom of the painting, the trenchcoat's hem brushes the ground, obscuring the cats' feet. The trenchcoat is a deep, rich color, perhaps a classic trenchcoat beige or a more dramatic black or navy. The coat is painted with careful detail, showing the texture of the fabric and the gleam of buttons.
Moving up the painting, the middle section reveals the outline of three cats stacked on top of each other within the coat. The artist might have chosen to show this through a slightly open coat, or perhaps through the distinctive shapes visible under the fabric. The cats could be of different breeds and colors to add visual interest and humor.
The top third of the painting, the upper two feet, features the head of the topmost cat peeking out from the collar of the trenchcoat. The cat might be looking straight ahead, trying to maintain its human disguise, or it might be looking out at the viewer with a humorous or mischievous expression. If the artist chose to include a hat, it would be perched on the cat's head, adding to the disguise.
The background of the painting could vary greatly depending on the artist's style and intentions. It could be a simple, solid color to keep the focus on the cats, or it could be a detailed scene that adds to the story - perhaps a city street to tie in with the Vincent Adultman reference, or a domestic setting to contrast with the cats' unusual behavior.
Again, this is a hypothetical description and the actual painting could be very different.
Models do NOT extrapolate unless they are given access to resources/data that is not somehow in their training dataset.
We have done so much encoding of knowledge that a majority of serious use-cases for LLMs are Interpolation, but the fact that we even need retrieval augmented generation shows that they do not extrapolate.
I'm saying that we shouldn't expect the models to come up with things we didn't train them to come up with.
These models have a tendency to move towards the average, especially if unprompted. As we see here, sometimes even if prompted otherwise.
They just have to get better or need a more precise interface like the —no :). We also couldn't have "a man crawling" before and now we can: https://i.imgur.com/ycVpk3i.jpeg
Adjusted and with help of the other commenter°: https://i.imgur.com/e2aE1bQ.jpeg
And you always have some crap generations :)
° https://news.ycombinator.com/item?id=36584831
You can get them standing side by side wearing trenchcoats. You can get a cat pyramid or a cat totem or a stack or tower of cats (though often the ability to count to three is then lost). You can't get them to share the trenchcoat. Nothing like that occurred in any training set, and the AIs do not understand spatial relationships between objects ("X is on top of Y, inside Z") so you cannot describe how to arrange the things it does know about in the scene. Dall-E won't do it. Midjourney won't do it. Stable Diffusion won't do it.
Eventually, enough images will be seeded into the training sets for this to stop being a useful test. But right now, it gives one a fascinating window on what happens when you try to extrapolate outside the cloud of thingspace described by the training data, rather than just interpolating within it.
https://www.toothycat.net/~sham/trenchcoats
> As an AI, I can't actually see the painting, but based on the title and the dimensions you've provided, I can imagine a possible description.
> The painting, titled "Three cats in a trenchcoat standing on each other's shoulders, pretending to be a human, Vincent Adultman style," is a large piece, standing 6 feet tall and 4 feet wide.
> At the bottom of the painting, the trenchcoat's hem brushes the ground, obscuring the cats' feet. The trenchcoat is a deep, rich color, perhaps a classic trenchcoat beige or a more dramatic black or navy. The coat is painted with careful detail, showing the texture of the fabric and the gleam of buttons.
> Moving up the painting, the middle section reveals the outline of three cats stacked on top of each other within the coat. The artist might have chosen to show this through a slightly open coat, or perhaps through the distinctive shapes visible under the fabric. The cats could be of different breeds and colors to add visual interest and humor.
> The top third of the painting, the upper two feet, features the head of the topmost cat peeking out from the collar of the trenchcoat. The cat might be looking straight ahead, trying to maintain its human disguise, or it might be looking out at the viewer with a humorous or mischievous expression. If the artist chose to include a hat, it would be perched on the cat's head, adding to the disguise.
> The background of the painting could vary greatly depending on the artist's style and intentions. It could be a simple, solid color to keep the focus on the cats, or it could be a detailed scene that adds to the story - perhaps a city street to tie in with the Vincent Adultman reference, or a domestic setting to contrast with the cats' unusual behavior.
> Again, this is a hypothetical description and the actual painting could be very different.
Until larger models are released, there are "hacks" like this available: https://bair.berkeley.edu/blog/2023/05/23/lmd/ - GPT4-generated bounding boxes guiding stable diffusion
How is this at all whatsoever needed
Not to mention complaining about wealth disparity from investment as if it is a sin of the technology is asinine. If money can produce better results you will see wealth disparity from the simple feedback loop of "more money -> better results -> more money". Your neighbor taking up jogging while you sit on the couch all day is increasing a health disparity, should your neighbor stop jogging then?
Undermining of democratic institutions is just plain a dumb meme from vested interests who cast what were previously abuses or necessary evils of the system at best as integral to democracy. We would all recognize blaming the media for undermining the Catholic Church by documenting their child sexual abuse as batshit insane. Yet somehow it is different when it comes to newspapers moaning about their loss of monopolies and influence of cliques.
Humans have shitty models for stuff, but we have them. We lack the massively deep numerical associations.
It can't be that long before someone makes an AI that says knows a few common models like "stuff made of steel is rigid, stuff made of cloth is soft, etc". Along with "If someone keeps emphasizing a number, that's the number of things he wants".
Chatgpt: Based on the description provided, it is likely that "Misiblur," the art piece made of cloth resembling a butterfly, would be more comfortable to rest your head on compared to "The gattoya," which is made of steel bars welded into a spiral shape.
The reason for this is that cloth generally has a softer and more cushioned surface compared to steel bars. When resting your head on Misiblur, the fabric would provide a certain level of support and padding, which could make it more comfortable for resting your head.
On the other hand, resting your head on steel bars, as in the case of The gattoya, would likely be less comfortable due to the hardness and lack of cushioning provided by the material. Steel bars are not designed to be soft or ergonomic for resting purposes, so it may not provide the desired comfort.
It's important to note that the comfort level can vary depending on individual preferences and the specific design and construction of the art pieces. However, based on the given materials and descriptions, Misiblur made of cloth would generally be the more comfortable option for resting your head.
Does it make pictures of people who are comfortable in one but not the other?
Most of the time when I see something funny in LLMs it's in the images. People with 8 fingers, elbows bent backwards, that kind of thing. Spatial models.
(edited to remove incorrect info)
The “Lone Banana Problem” describes subtle biases of Large Language Models (LLMs) in AI: LLMs reproduce the statistical average of the inputs that they have consumed in the context of the question they have been asked. It's called that problem because the model used to generate images has never seen an individual banana, so when prompted always generates two bananas.
I didn't think it was that long, it might be that some are cherry picking the points they find interesting and think the rest could have been edited out.
I could have stood even more discussion on how this 'blind spot' in the AI model is very much akin to our own blind spots.
For "dumbbell", for the AI, no dumbbell was complete without a muscular arm holding it. That's because in most images in the training dataset have the dumbbells being held, so the system integrated the arm in the pattern. I guess it is the same idea here, most pictures of bananas show several of them, so for the AI, bananas are things that don't go alone.
https://news.ycombinator.com/item?id=32875215