Yes, and it still feels a lot like Searle's Chinese Room. It's as if it skips a dozen steps. Well, that's exactly what happens, of course. But it does show that the network can match linguistic descriptions to images extraordinarily well.
Do you feel that the human mind is more than an "appropriately" trained "biological" neural network?
What do you consider the limits of a DALL-E like system compared to a "true" mind?
My personal opinion is that the Chinese Room argument is fancy handwaving that crucially relies on never being explicit about what it means by "understanding", combined with an appeal to intuition.
I strongly believe that there is nothing "magical" about the human mind or brain (that could not be replicated artificially), and thus that a comparably trained, appropriately designed system ("DALL-E successor") OR a copy OR a simulation of a human brain would be all just as capable and "understanding"/"conscious" as another human...
> My personal opinion is that the Chinese Room argument is fancy handwaving that crucially relies on never being explicit about what it means by "understanding", combined with an appeal to intuition.
There doesn't need to be anything "magical" about the human mind for human "understanding" not to be particularly closely approximated by performing mathematical transformations on a huge corpus of raster data, even though the latter approach produces very useful results (see also the utility of a pocket calculator vs attempting to skip a billenium or so of natural evolution to grow an organism motivated to learn multiplication in a petri dish, or parse human minds to figure out the elements of brain state associated with performing a calculation)
GPT-3 is far, far better at generating descriptions of loving humanity than the average dog. But it's pretty obvious that the dog's goal formation and excitement hormones aren't particularly similar to letter-by-letter ASCII output probability calculations on a database, and that GPT-3 has no more grasp of the dog's love of humanity in general and this human in particular than the dog has of Shakespeare. "Thought" and intelligible results are essentially orthogonal, which limits the feasibility of training the former...
> My personal opinion is that the Chinese Room argument is fancy handwaving
It isn't hand-waving. It's against it, really. It's a thought experiment that encourages a sceptic attitude towards jumps in understanding mental processes. The operator in the Chinese Room doesn't understand Chinese. While the translations are excellent, he or she wouldn't be able to go out in the street and ask for a glass of water if their life depended on it. Hence, a computer that mechanically translates Chinese cannot be automatically assumed to understand Chinese.
The argument doesn't need to explain exactly what understanding means. We all (sort of) know what it means. The same goes for e.g. attention. That's what makes it so hard to define what strong AI is and how to verify it. The Turing Test famously tries to decide this (without defining anything, I might add), and the Chinese Room is a good argument against it being the proper test.
> The argument doesn't need to explain exactly what understanding means. We all (sort of) know what it means.
Then what does "understanding" a language mean? Your "asking for a glass of water on the street" example implies to me that you demand a system with:
a) capability for internal intent
b) ability to express intent in target language
to grant it "understanding".
Basically, first you deliberately construct a system not capable of intent (pure stateless query-response algorithm), then you deny that it "understands" Chinese on the basis of being unable to express intent. That does not hold.
Give me a precise definition for what you mean by "understanding" and I'll dismantle the Chinese Room for you.
> Give me a precise definition for what you mean by "understanding" and I'll dismantle the Chinese Room for you.
You probably can't define "chair" in an exact enough sense. It would merely show the shortcoming of the definition rather than refuting the Chinese Room.
> a system not capable of intent
If I now would require a definition of "intent" I would be just as childish.
> That does not hold.
First, the system isn't stateless. It has the operator who has to remember what has happened before in order to come up with a correct translation.
Second, your argument implies that understanding requires intent. That alone is a tall order to prove. But it's the point of the Chinese Room: the mechanism doesn't understand, doesn't have intent, nothing. Yet it performs exactly as an intelligent, understanding, intentional translator. So would you not agree with the conclusion: you can't judge intent by only looking at I/O behavior?
Note that Searle (probably, this is just my understanding) doesn't mean that the Room is useless or "dumb" (in most senses of the word).
The sleight-of-hand in the Chinese room is that Searle asks us whether the man in the room understands Chinese. Of course not. This is like asking whether my CPU knows how to decode h264. The real question is whether the embodied process instantiated by the actions of the man, along with the other involved components in the room, understands Chinese. But the argument doesn't touch this claim.
If you ran it 1000 times and picked the best, you might get all good ones. I would want to see all 1000. It's like stock picker ads (person who called the market says XXX) where you only show the lucky ones.
Yeah exactly. The fact that a person who could never create such a picture on their own, now just has to go through a bunch of images and select one to get this result is already amazing. The goal posts keep shifting, glass half empty.
It reminds me of pop articles about the P=NP question which simplify it down to, "is being able to recognize creativity the same as being creative?" Turns out for practical purposes, yes (the simplification question not the mathematical question), as long as you can offload the work to an AI on a powerful compute cluster!
It's amazing and wonderful, but it's not the AI that's "creating art", it's the human choosing the best AI output that's doing the artistic part (often togehter with the person who input the prompt to the AI).
It's just like Dada poets did 100 years ago: you use a mechanical process to generate quasi-random output, and then you choose some of this output to present to other humans. The way you provide input to the mechanical process (e.g. what words you choose to put in the hat) and the curation are the real creative part of the art, not the mechanical process generating the text itself.
I bet I could give a phrase to a collection of art students in some university and DALL-E, and the human art would likely be more creative than a single run of DALL-E. What distinguishes the humans is that you can tell them to make art of their own choosing and they will, but DALL-E is unlikely to create anything interesting with no input.
Van Gogh invented Starry Night without any prompting despite it not being a real scene (much less anything he had ever seen and such abstraction was very rare in the 1880s). Picasso made Les Demoiselles d’Avignon in 1907; it was so radical even his fellow artists were unable to comprehend it.
It doesn't change the fact that DALL-E is pretty amazing tech, but it's still as far behind human ability as any AI is today. It is way way better than what came before, but that's true of most technologies.
Imagine a crowd sourced metadataset of which model output is "good". What happens when we try to incorporate that knowledge into the pre-existing model? Will it learn to generate more "good"looking output?
Are you aware that this is how human artists work, too? They make countless works that aren't comparable to their top works. Even more so if you count the practice when they were just starting out. I think some people just refuse to believe that real art can be generated without humans and selectively look for things that confirm their pre-determined conclusion. Witnessing this always feels like witnessing someone before the wheel was invented, when things were rolled and logs, assuring everyone that technology will never beat manual labor because look at how cumbersome it is to work with logs.
There is a crucial difference: the human artist does this choosing themselves.
If DALL-E just spits out 1000 images and then a human goes through them and picks the best 2-3, and those are good - it's impressive, but the human was still a crucial part of the process. On the other hand, if DALL-E were to generate 10 billion images, and choose the best 2-3 itself and give those as output, and if at least one of those 2-3 would be consistently great, then DALL-E could be indeed considered to be creating (good) art.
Sure, it might. But until then, I wouldn't say it makes sense to consider the AI as "being the artist".
When DALL-E x.0 does that, and when it also generates similar quality from much higher-level prompts ("paint a sad picture", or "social commentary on BLM" or something like this, instead of a description of what the picture should show and in what style), then I for one will be in complete agreement that it's indeed an artist in itself.
Personally, I don't expect this to happen in the next few decades, as I don't think the current approaches are very promising for the type of intelligence that you would need to actually do this type of reasoning, but that remains to be seen, and I am fully confident that it will happen some day.
I disagree. The human artist's tastes at least partially originate in other people, both individuals and general societies/cultures, and oftentimes the artist directly incorporates feedback into future work. Are you aware that students in art school, music conservatories, etc constantly get feedback from instructors and peers? I reject your premise entirely unless you can give me an example of a human that created art without ever having been influenced as a human being by any other human being. Otherwise I believe it's just what I said before: concluding first that AI can't create art and finding reasons second.
Does DALL-E incorporate (or even receive) feedback about which of the pictures it generated were better? It of course does not, and it currently has no function to do so. IF it incorporated this feedback and changed its weights based on it, I would agree with you that the situation could be comparable.
Until then, my point remains: DALL-E is currently like an (extraordinarily good) hat that you can put words in and extract phrases out of. A human chooses what words to put in and which of the phrases they take out are better. Unlike pulling words out of a hat, the network has some criteria by which it produces phrases, but that's not enough to call it an artist.
This is not meant to minimize how good the achievement of this network is. The level of fidelity and even understanding of the prompts is extraordinary. But its purpose is not to be creative, it is to find a point on a hyperplane that matches the input it received. It is currently at the level of a tool - though there are potential advancements that could yet turn it into an artist in its own right.
We're basically just arguing semantics at this point. You define art/artist differently than I do. We won't make further progress with this discussion, but regardless it's great to see that technology is coming so far
> On the other hand, if DALL-E were to generate 10 billion images, and choose the best 2-3 itself and give those as output, and if at least one of those 2-3 would be consistently great, then DALL-E could be indeed considered to be creating (good) art.
It's worth noting that the OpenAI samples for DALL-E 1 used CLIP to rank generated samples, and got a big boost from that. For many model architectures, you can run them in reverse to do 'image -> caption', and 'score the caption' quality: if 'the caption is bad', that indicates your image was screwed up and low-quality (introduced by Cogview). DALL-E 2 doesn't use either approach, or finetuning on user choices like InstructGPT, and I dunno if OA is going to implement any of these, but there is a wide universe of techniques applicable here to improve quality and we should keep that in mind (https://www.gwern.net/Forking-Paths) if we are going to make any assertions more sweeping than "this specific model, at this very instant, with this particular interface, is only at this level of quality".
As far as I know, the actual trained system is proprietary, and you can only use it by requesting access to their online system for generating imagery: https://openai.com/dall-e-2/
There are open-source efforts to implement it and make trained models available, but I don't imagine they are yet at the same scale of ingested data / model size as OpenAI's system: https://github.com/lucidrains/DALLE2-pytorch
Does that mean the developers put The Big Lebowski through the model during training? Where did they get all the movies and TV shows from? And does copying so directly from the source material open them and users up to copyright infringement liability?
They probably did, but even if they didn't it may have come from the terrabytes of data they scraped from the internet. OpenAI doesn't care. They claim that it's derivative enough to go under fair use. And whether it is or isn't, I guess their calculation is that the risk is worth taking to be the first to develop these algorithms, which is a huge head start if the courts decide that it does count as fair use.
Hilarious parodies like these are copyright infringement, yes, but also open-and-shut fair use defense. (You're confusing the issue of transformativeness and fair use defense.)
I feel like if they don’t have licenses for all of their source material, then the model should be required to be released into the public domain.
It’s like extremely expensive piracy that is bad for artists and bad for the environment.
I wonder if the reason OpenAI, Google, etc don’t release these things isn’t so much that they’re worried about racist/offensive output, but instead they’re worried about people using it to create images of, say, Mickey Mouse and drawing the attention of his lawyers? It’d be better for AI companies to keep all of this stuff in a legal gray area for as long as possible.
> Being bad for artists and the environment is not illegal.
Yet we have copyright laws and environmental protection laws to protect both.
> If you look at a movie poster, your brain does not need to be released in the public domain. Even if you sketch it from memory.
Because I’m not a machine. I’m contrained by physics, whereas these models are not.
Copyright laws were made to protect artists from IP theft. If I make a sculpture, it’s not trivial to copy that and steal from me, so creating copyright laws to protect sculptors would be a hard sell.
But a painting, a book, a song, etc are easy to steal, especially with technology. Copying and selling someone else’s painting is similar to copying and selling someone else’s sculpture, yet the scale of theft is obviously different (mass producing sculptures is much harder than making unlimited copies of an mp3)
These AI models are a new type of theft, and likewise need new types of legal protections for artists.
> hese AI models are a new type of theft, and likewise need new types of legal protections for artists.
Meh. This means that most artists will also struggle to make something as good as these models and will have to find something else to do for a living. Its not the first time in history that an occupation is obsoleted. Projectionnists in cinemas have mostly dissapeared.
If these models are trained on human artists, what’s going to happen if humans stop making original art? It’s obvious from the OP that even a single film is enough to add a new stylistic capability to Dall E, so even though it might have been trained on hundreds of years of human artwork, a single modern piece can still influence significantly.
> This means that most artists will also struggle to make something as good as these models
This makes me think that you’re lacking a fundamental understanding of what art is. The “demand” for art is never going to disappear, but these models are simply going to disrupt the supply by flooding the market with derivative works at a massive scale.
The internet posed a similar threat back when it first came out, but laws were changed to combat that (e.g. the DMCA). I fail to see how this situation is any different?
A country that doesn’t get on top of this issue will doom itself to cultural irrelevance in the long term, IMO.
I would hope that the spirit of the law would be considered in this. This is a clear application of fair use. Are the owners of the IP losing money on letting someone generate new characters in their movie/show?
> Are the owners of the IP losing money on letting someone generate new characters in their movie/show?
Typically, creators are very protective of this sort of thing, unless it stays in the area of fan art. If anyone tries to seriously monetize this kind of output, I'm sure we'll see a lot of cases.
Imagine what Disney would do if you used DALL-E to create an animated feature film in the style of Mickey Mouse, but with cats instead of mice, and they found out you used actual footage from, say, Fantasia to train an AI model. No idea if they would win, but I'm certain they'd sue.
I was assuming it was just movie posters and screenshots from online articles/reviews/imdb etc, and not any analysis of the video itself (I might be wrong though - using video would make the number of available input frames grow by orders of magnitude, not that there is a lack of pictures online).
E.g. The Shining picture of Jack Nicholson with the door isn't representative of the "look" of the film, but very much an iconic still frame and basically what you see in a Google image search for "The Shining".
Wouldn't the images be a transformative work? But then again there were those recent music cases where the infringement was only a small number of notes in sequence. This will almost inevitably end up in court because I'm not sure there is a comparable case.
It's funny that 10 years ago "a human did it" would be, by far, the most plausible explanation. Today, it's "AI did it." Put another way, if a human made these then it'd be as shocking as seeing DALL-E 2 in 2012.
I think "legit" in this context might mean more of the Twitter post author's representation of their use of DALL-E 2:
- Were the prompts shown the ones fed to DALL-E 2 or were there more complex details described in the prompt?
- Were these the first images generated for the prompt, or did the author generate many images and cherry-pick the best example, and if so from how many?
>> I suspect at some point in the next 5-15 years we will begin to see AI generated entertainment perfectly tailored to a person's preferences.
Kermit in Debby Does Dallas. Kermit in the Graduate. Kermit with 2 broken flippers. Oh the depravity. I'm not sure getting high quality visualizations of any random passing thought is a good idea ;-)
Maybe the loneliness is a good thing? It seems like everyone (in America at least) hates each other, so extreme isolation may be just what we need in this country.
I think it's probable that everyone hates eachother because we're so isolated that we fail to form empathetic bonds with the people around us. More isolation doesn't seem like the solution.
We are kind of already there (not having entertainment in common). It is hard to find a TV series or film that everyone in my friendship circle knows equally well -- the last one was Breaking Bad, and that was over 10 years ago. Everyone finds their niche interest I suppose. Its almost the same with music -- the last decade where our musical interests converged were the 90ies.
Interesting point. I would love to see my personal perfect TV show.
But I would also love to see the perfect show for a combination of people, like my loved ones, so we can watch together.
And lastly, I would love to see the perfect show for a large majority of the population, because there is something special about a shared experience like Game of Thrones or Friends. To this day I can shout “pivooooot” and most everyone gets it.
On the one hand, the ability for Star Trek's Holodeck to create large amounts of content from a few terse natural language instructions is look less and less implausible.
On the other hand, I feel like this will ultimately be kinda like traditional procgen algorithms: once you've seen enough of what it produces it all starts feeling very bland and same-y. Sure, the AI may be able to produce a feature-length movie based on the input "What if Nicolas cage had played The Terminator and Aaron Sorkin wrote the script?", but somehow none of it would be surprising or interesting to you, it would lack the novelty and playfulness of a good human creative work, and it likely would be very shallow in its themes.
On the gripping hand though, perhaps in achieving that level of sophistication we inadvertently create something more alive and aware than we intended and instead of merely trying to produce satisfactory results it actually attempts to express itself in ways that resonate with us.
> There’s nothing inherent in human creativity that can’t be replicated by an AI.
I didn't say otherwise, I'm just skeptical the hypothetical system in question would be sophisticated enough to produce it in faster-than-human time scales.
Loved the David Lynch ones. I'm disappointed the image it generated for Eraserhead[1] didn't have Kermit as the baby[2]. I'm curious as to what it would generate for "Kermit the Frog in David Lynch's Dune".
These are honestly not very impressive (no sarcasm here) and further convince me that the next AI Winter will come with this coming recession.
Don't get me wrong, they are still impressive in the quality of the visual they produce, but just like Markov Chain demos of old, they're neat but way miss the mark.
None of these capture the "feel" of Kermit the Frog. Most of them look like weird designs for the Ninja Turtles movie in the 90s.
There are several distinctive features of Kermit that a missing from nearly all of these.
- For any of the "live action" ones, Kermit should still always be a puppet.
- Kermit notoriously has lanky arms,
- Kermit never has eye lids
- His eyes sit way on top of his head.
- He often has his weird neck decoration.
- His eyes have a very distinctive pupil shape.
None of these get Kermit correct, they all just look like frogs (maybe Dalle2 isn't trained on copyrighted/trademarked material?)
There are fan made versions of some of these which show just how different Dalle2 is from human imagination:
Again if this was done on someone's laptop it would be really impressive. However the fact that so much talent and resources were poured into pushing AI to it's limits and this is what we get tells me we've hit another brick wall as far as research goes.
I thought I was taking crazy pills, none of them look like kermit bur rather they look like a generic frog. They don't even have the same pattern around his collar.
It is odd, isn't it? It captures "essential" characteristics of all those films in a honestly brilliant way - but it doesn't capture any of the iconic characteristics of Kermit himself!
Agreed. They are self-evident, high-quality representations of KtF. When faced with disagreement on that, it's hard to know what to say in response to convince someone other than to point to another canonical representation of KtF and say, "These are the same."
> - For any of the "live action" ones, Kermit should still always be a puppet.
- Kermit notoriously has lanky arms,
- Kermit never has eye lids
- His eyes sit way on top of his head.
- He often has his weird neck decoration.
- His eyes have a very distinctive pupil shape.
This is something akin to isolated demands for rigor. Apparently these features are not essential to KtF-ness, because most people look at these pictures and see unambiguous KtF.
Yet, it did produce things which "look like weird designs for the Ninja Turtles movie in the 90s."
In other words, it has done as good a job of costume design, lighting and photographing a live action Kermit as New Line Cinema paid $13.5m to accomplish for TMNT in 1990.
And you know who they got to do those creature designs?
> Again if this was done on someone's laptop it would be really impressive. However the fact that so much talent and resources were poured into pushing AI to it's limits and this is what we get tells me we've hit another brick wall as far as research goes.
You might be missing the point of what OpenAI is doing. The point is to show off the capability of their models in a way that's likely to go viral and lead to more business for OpenAI. Some people laughed at GPT-3's silly demos, but when they launched GitHub Copilot...
If people say Dalle can improve the workflow of digital artists, sure, but Copilot hasn't revolutionized programming either, you still have to be a good programmer to finish whatever you are doing:
> A paper accepted for publication in the IEEE Symposium on Security and Privacy in 2022 assessed the security of code generated by Copilot [...] The study found that across these axes in multiple languages, 39.33% of top suggestions and 40.73% of total suggestions lead to code vulnerabilities. Additionally, they found that small, non-semantic (i.e., comments) changes made to code could impact code safety.[14]
What happened next? Is anyone using copilot for serious work? Has it changed programming in a fundamental way?
I personally have zero use for copilot since the for type of code I write the actual code writing is not a bottle neck, so automating that process is of no value to me. On top of that getting the details exactly right is essential so the ratio of boiler plate to real code is very, very low for me.
I completely disagree. I think these are taking it a step further than your examples. Dalle2 is not just using the existing Kermit and pasting it in different environments, it's modifying Kermit to fit in that world.
It's clearly just an existing photo of Kermit pasted over an image from the film. There are even two sets of arms. I could Photoshop that in a few minutes.
I think it's impressive. It looks like Kermit is a character in the Star Wars universe. There are a few issues with the eyes and feet, and it's also hard to tell if it's a creature or a person in a frog suit. However, it gets 90% of the way there, and the pose is great for a frog/human hybrid.
The most exciting thing is how this could be used as a starting point for design. I could take the Dalle2 Kermit image above, fix the eyes/feet, add a few distinctive Kermit features, and have a great piece of concept art in an hour, rather than taking a day or two to create something from scratch. Obviously it can't be applied to all workflows, but for those it's suited for, it'll save vast amounts of time and costs. For that reason, it's already something of real value in its current state. The same can't be said about the Star Wars examples you provided.
I disagree that this is unimpressive, but do largely agree about AI winter. Dall-E 2 is probably the most impressive AI implementation I can recall in the past 5-10 years and it's still highly specialized problem being solved, and it's unclear what market it really can go after other than freelancing digital artists online who work for tiny commissions. I guess it's also gonna be great for NFTs but I consider that market illusory and will disappear within a few years.
I am definitely going to bet against the "AI winter is returning" idea by investing huge amounts of my time into understanding these algorithms. History doesn't have to repeat, sometimes that's the foolish prediction (the apple newton was made fun of by the simpsons but when the ipad came out the timing was perfect). I don't like overinflating OpenAIs already enormous ego but these are incredible images.
Agree there's no reason it has to come again, just things got so hyped the past 5-10 years I wouldn't be surprised if a reality check on the horizon involves a "winter" of sorts and paring down of investment, as it becomes clear that current techniques are great at highly specialized problems where there is a lot of data. There's lots of firms that have both though so maybe it just keeps chugging along as ROI is found.
Well the investment landscape in general is kinda drying up (innovation is getting harder) and AI is one of the few areas that could or is bearing fruit so that's a big part of it too (even hardware and software is kinda stagnating). If investment is going to go anywhere AI is still one of the few places in the world that could have 100x returns (despite many frauds, yes).
I agree with what you say in re: Kermit. Most of these images look to me like a frog that looks like Kermit the Frog but isn't. Metaphorically (and literally) Jim Henson isn't in these images.
However, I don't think you're correct in your assessment of the import of this sort of thing: it's an imagination machine. This isn't a brick wall, it's a foundation on which to build.
While I think many are over-interpreting the quality of these results, yours is sounding like a clear case of a No True Kermit fallacy.
There are many ways to define what "Kermit the Frog in $MOVIE" means, and the choice the AI made is absolutely valid. There are of course various other valid choices, but this doesn't invalidate the ones presented.
Furthermore, judging by some other examples in this HN thread, it seems that the fact most of the pictures are not puppets is more of a choice of the human choosing the photos, as in other cases DALL-E was indeed adding puppet-like characters in movie-like decors.
> These are honestly not very impressive (no sarcasm here) and further convince me that the next AI Winter will come with this coming recession.
"Sure, this AI can produce high-resolution realistic images leaps and bounds above anything that's been shown before... but there's an aspect which could use improvement. Obviously, this proves that the current AI technology will never amount to anything and we should just give up on it now."
I don’t understand how this isn’t infringing some copyright. Anyone do any research on this?
If we accept that a model trained on copyright material does not infringe on the materials rights, then circumventing all copyright can be as simple as creating a sufficiently close derivative and giving it away.
I shouldn’t clarify - I’m not talking about the Kermit images, I’m talking about Dalle2 itself. If you had it render Neo from the Matrix, albeit imperfectly, would that be transformative use?
If you drew it yourself there’s precedent under fair use. If you made something that drew it when prompted for “The Matrix” presumably it knows what that is and therefore is more ambiguous
That remains to be seen. If I take a photograph of a still from the Matrix and print it, that's not the same as me photo-realistically drawing the same still from memory, which is itself not the same as me photo-realistically drawing it while looking at the still itself.
Copyright law is way more nuanced than you think, especially around fair use.
It's established that you can't assert copyright over a picture generated by an ML algorithm. I don't think it's established whether a picture generated by an ML algorithm can infringe someone else's copyright (e.g. if you ask DALL-E to produce a picture of Mickey Mouse, and it does a perfect rendition of him, is this picture in the public domain, or does it infringe Disney's copyright?).
Hang on, I don't see where that said that you can't assert copyright over a ML generated picture. I think the decision says that the ML algorithm itself cannot assert copyright?
The decision says that no one can be considered the author of that picture, as copyright only protects human creative works. It is possible that the work infringes someone else's copyright (the decision doesn't address this point at all), but unless that happens, the image is in the public domain and you can freely copy it.
Note that, unlike patents and trademarks, copyright doesn't require registration. The court could have found that, while the AI can't assert copyright, the copyright still exists and belongs to the owner of the AI or the designer of the AI etc. Instead, they found that the work can't be protected by copyright, since it is not the product of human creativity.
Edit: here is a citation from the article that I'm basing the previous assertions on:
> Both in its 2019 decision and its decision this February, the USCO found the “human authorship” element was lacking and was wholly necessary to obtain a copyright, Engadget’s K. Holt wrote. Current copyright law only provides protections to “the fruits of intellectual labor” that “are founded in the creative powers of the [human] mind,” the USCO states.
Everything that you, a person with a paintbrush, could paint "in the style of" something else is informed by what your model (your brain) has been exposed to. There's no getting around that, and commissioning you to paint something "in the style of postwar authoritarian England" does not infringe the copyright of V for Vendetta (even if I told you "make it look just like the movie"); it's an original painting.
Stylistic inspiration is not an infringement of copyright, in either that case or the "do it on a computer" case here.
The Kermit the Frog aspect though is interesting - it applies equally for both the human and machine made works - if an argument could be made that the subject of the work sufficiently resembles the character, maybe there's a trademark issue at hand?
But in any scenario, nothing legally novel about the work being created by machine.
> …except for the fact that it was created by a machine.
We've already had this for years. The photos you take on any modern smart phone are partially the invention of AI (doubly so if you use something like portrait or night mode). It's not just raw CCD output, and yet, the photographer retains the copyright.
DALL-E's terms could require users to assign copyright. If not, I don't see any reason it wouldn't go to the person who came up with the prompt and picked one particular generation.
If I take a picture of a mountain with a camera, neither the mountain nor the camera hold copyright. DALL-E's just another tool in the toolbox.
That’s obviously very different. The contributions of the AI processing my phone camera does is not significant, and can be substituted relatively easily with Photoshop or similar software. It’s just a convenience feature.
Dall-E is the complete opposite. My contribution is insignificant (a prompt), whereas Dall-Es is 99% of the work.
> If I take a picture of a mountain with a camera, neither the mountain nor the camera hold copyright. DALL-E's just another tool in the toolbox.
Not sure how this contradicts my point. Nobody “owns” that mountain, so nobody could call your photo of it a derivative work.
If you take a photo of a frame of a movie, then who owns the rights to that photo? It’s not a raw frame, it’s a photo of a screen. Does that mean you didn’t just steal someone’s IP?
These questions need answers, and copyright laws try to answer them in as fair a way as possible. Dall E raises new questions, and thus we need to update copyright laws with new answers.
My personal view is that obtaining licensing for every single work used to train the model is impossible. So instead of simply making this type of AI research illegal, the fairest solution is to put that model into the public domain.
There are still financial incentives to perform this research, as the discoveries could be used later to train a commercial model with proper licensing.
> My contribution is insignificant (a prompt), whereas Dall-Es is 99% of the work.
I disagree entirely with this. Picking a prompt and potentially going through dozens, hundreds, or thousands of variations and re-generations has artistic value.
Just like the photograph of a mountain. Nature did most of the work, but a human selecting an angle, lighting, etc. from the nearly infinite combinations available matters.
How is this so good? If you told me this was done by a very talented human, I’d still be surprised at how good they are. That must count as a sort of Turing test, no?
Well, most of the pictures (if not all), while astoundingly good as an idea, have the tell-tale signs of being AI-generated, the kinds of mistakes a human would never make.
For example, looking at the WALL-E one [0], you can clearly see that the hands and feet aren't actually separated properly. There is also plenty of missing "logic" around the armpits. These are the kinds of mistakes a human can't make - especially one that is so adept at drawing the other parts so perfectly.
How long does it take to generate these pics? No human is that good at producing art of this quality.
There may still be a few anatomic mistakes, but many artists also make some. The picture quality, lightning, and the way it capture the graphic essence and mood of these movies is just amazing and beyond what even the most talented artists can pull out in the same time frame.
I'd guess these are curated, but I have no idea how many tries were discarded, as I'm still waiting for the invitation. Even in the best examples you can find fused fingers, deformities, melting, and slight errors that a human artist would need to touch up before it truly passes as a human work.
Not to detract from the accomplishment, but none of these are Kermit. The "Kermit" part of the query seems mostly to have accomplished querying for "humanoid frog"
If I hadn't been prompted with the words Kermit, no way would I have guessed that they were supposed to be Kermit. For example, Kermit has a characteristic shape of his pupils. None of the examples have that.
Some of them look a lot like Kermit! But many do not, depending on how far away the target aesthetic is from "muppet." A smarter Dalle2 could do a better job at preserving the "essential Kermit-ness" perhaps. But maybe it's more impressive to adapt to a completely new aesthetic?
That's actually one of the more interesting and impressive things: the AI has substituted in "humanoid frog" and in some cases given it movie-appropriate features like Matrix sunglasses ot Pixar-style eyes which are probably unique to the image rather than simply pasting photographs of the original muppet into movie poster settings which would be an acceptable lower level response to the brief.
197 comments
[ 4.3 ms ] story [ 230 ms ] threadDo you feel that the human mind is more than an "appropriately" trained "biological" neural network?
What do you consider the limits of a DALL-E like system compared to a "true" mind?
My personal opinion is that the Chinese Room argument is fancy handwaving that crucially relies on never being explicit about what it means by "understanding", combined with an appeal to intuition.
I strongly believe that there is nothing "magical" about the human mind or brain (that could not be replicated artificially), and thus that a comparably trained, appropriately designed system ("DALL-E successor") OR a copy OR a simulation of a human brain would be all just as capable and "understanding"/"conscious" as another human...
bingo
GPT-3 is far, far better at generating descriptions of loving humanity than the average dog. But it's pretty obvious that the dog's goal formation and excitement hormones aren't particularly similar to letter-by-letter ASCII output probability calculations on a database, and that GPT-3 has no more grasp of the dog's love of humanity in general and this human in particular than the dog has of Shakespeare. "Thought" and intelligible results are essentially orthogonal, which limits the feasibility of training the former...
It isn't hand-waving. It's against it, really. It's a thought experiment that encourages a sceptic attitude towards jumps in understanding mental processes. The operator in the Chinese Room doesn't understand Chinese. While the translations are excellent, he or she wouldn't be able to go out in the street and ask for a glass of water if their life depended on it. Hence, a computer that mechanically translates Chinese cannot be automatically assumed to understand Chinese.
The argument doesn't need to explain exactly what understanding means. We all (sort of) know what it means. The same goes for e.g. attention. That's what makes it so hard to define what strong AI is and how to verify it. The Turing Test famously tries to decide this (without defining anything, I might add), and the Chinese Room is a good argument against it being the proper test.
Then what does "understanding" a language mean? Your "asking for a glass of water on the street" example implies to me that you demand a system with: a) capability for internal intent b) ability to express intent in target language to grant it "understanding".
Basically, first you deliberately construct a system not capable of intent (pure stateless query-response algorithm), then you deny that it "understands" Chinese on the basis of being unable to express intent. That does not hold.
Give me a precise definition for what you mean by "understanding" and I'll dismantle the Chinese Room for you.
You probably can't define "chair" in an exact enough sense. It would merely show the shortcoming of the definition rather than refuting the Chinese Room.
> a system not capable of intent
If I now would require a definition of "intent" I would be just as childish.
> That does not hold.
First, the system isn't stateless. It has the operator who has to remember what has happened before in order to come up with a correct translation.
Second, your argument implies that understanding requires intent. That alone is a tall order to prove. But it's the point of the Chinese Room: the mechanism doesn't understand, doesn't have intent, nothing. Yet it performs exactly as an intelligent, understanding, intentional translator. So would you not agree with the conclusion: you can't judge intent by only looking at I/O behavior?
Note that Searle (probably, this is just my understanding) doesn't mean that the Room is useless or "dumb" (in most senses of the word).
It's just like Dada poets did 100 years ago: you use a mechanical process to generate quasi-random output, and then you choose some of this output to present to other humans. The way you provide input to the mechanical process (e.g. what words you choose to put in the hat) and the curation are the real creative part of the art, not the mechanical process generating the text itself.
Van Gogh invented Starry Night without any prompting despite it not being a real scene (much less anything he had ever seen and such abstraction was very rare in the 1880s). Picasso made Les Demoiselles d’Avignon in 1907; it was so radical even his fellow artists were unable to comprehend it.
It doesn't change the fact that DALL-E is pretty amazing tech, but it's still as far behind human ability as any AI is today. It is way way better than what came before, but that's true of most technologies.
For this one in particular, here are a few more results for Battlestar and The Office:
https://twitter.com/Miles_Brundage/status/153247388947686195...
If DALL-E just spits out 1000 images and then a human goes through them and picks the best 2-3, and those are good - it's impressive, but the human was still a crucial part of the process. On the other hand, if DALL-E were to generate 10 billion images, and choose the best 2-3 itself and give those as output, and if at least one of those 2-3 would be consistently great, then DALL-E could be indeed considered to be creating (good) art.
When DALL-E x.0 does that, and when it also generates similar quality from much higher-level prompts ("paint a sad picture", or "social commentary on BLM" or something like this, instead of a description of what the picture should show and in what style), then I for one will be in complete agreement that it's indeed an artist in itself.
Personally, I don't expect this to happen in the next few decades, as I don't think the current approaches are very promising for the type of intelligence that you would need to actually do this type of reasoning, but that remains to be seen, and I am fully confident that it will happen some day.
Fairly certain you can get results using this prompt with no issues.
Dalle 1.0 -> 2.0 was a shocking improvement, I expect the jump to 3.0 to be equally jarring if not more so
I disagree. The human artist's tastes at least partially originate in other people, both individuals and general societies/cultures, and oftentimes the artist directly incorporates feedback into future work. Are you aware that students in art school, music conservatories, etc constantly get feedback from instructors and peers? I reject your premise entirely unless you can give me an example of a human that created art without ever having been influenced as a human being by any other human being. Otherwise I believe it's just what I said before: concluding first that AI can't create art and finding reasons second.
Until then, my point remains: DALL-E is currently like an (extraordinarily good) hat that you can put words in and extract phrases out of. A human chooses what words to put in and which of the phrases they take out are better. Unlike pulling words out of a hat, the network has some criteria by which it produces phrases, but that's not enough to call it an artist.
This is not meant to minimize how good the achievement of this network is. The level of fidelity and even understanding of the prompts is extraordinary. But its purpose is not to be creative, it is to find a point on a hyperplane that matches the input it received. It is currently at the level of a tool - though there are potential advancements that could yet turn it into an artist in its own right.
It's worth noting that the OpenAI samples for DALL-E 1 used CLIP to rank generated samples, and got a big boost from that. For many model architectures, you can run them in reverse to do 'image -> caption', and 'score the caption' quality: if 'the caption is bad', that indicates your image was screwed up and low-quality (introduced by Cogview). DALL-E 2 doesn't use either approach, or finetuning on user choices like InstructGPT, and I dunno if OA is going to implement any of these, but there is a wide universe of techniques applicable here to improve quality and we should keep that in mind (https://www.gwern.net/Forking-Paths) if we are going to make any assertions more sweeping than "this specific model, at this very instant, with this particular interface, is only at this level of quality".
There are open-source efforts to implement it and make trained models available, but I don't imagine they are yet at the same scale of ingested data / model size as OpenAI's system: https://github.com/lucidrains/DALLE2-pytorch
Especially if the movie(s) that are eventually generated this way are ripping whole scenes or sequences out of other films, a la copilot.
It’s like extremely expensive piracy that is bad for artists and bad for the environment.
I wonder if the reason OpenAI, Google, etc don’t release these things isn’t so much that they’re worried about racist/offensive output, but instead they’re worried about people using it to create images of, say, Mickey Mouse and drawing the attention of his lawyers? It’d be better for AI companies to keep all of this stuff in a legal gray area for as long as possible.
If you look at a movie poster, your brain does not need to be released in the public domain. Even if you sketch it from memory.
Yet we have copyright laws and environmental protection laws to protect both.
> If you look at a movie poster, your brain does not need to be released in the public domain. Even if you sketch it from memory.
Because I’m not a machine. I’m contrained by physics, whereas these models are not.
Copyright laws were made to protect artists from IP theft. If I make a sculpture, it’s not trivial to copy that and steal from me, so creating copyright laws to protect sculptors would be a hard sell.
But a painting, a book, a song, etc are easy to steal, especially with technology. Copying and selling someone else’s painting is similar to copying and selling someone else’s sculpture, yet the scale of theft is obviously different (mass producing sculptures is much harder than making unlimited copies of an mp3)
These AI models are a new type of theft, and likewise need new types of legal protections for artists.
Meh. This means that most artists will also struggle to make something as good as these models and will have to find something else to do for a living. Its not the first time in history that an occupation is obsoleted. Projectionnists in cinemas have mostly dissapeared.
> This means that most artists will also struggle to make something as good as these models
This makes me think that you’re lacking a fundamental understanding of what art is. The “demand” for art is never going to disappear, but these models are simply going to disrupt the supply by flooding the market with derivative works at a massive scale.
The internet posed a similar threat back when it first came out, but laws were changed to combat that (e.g. the DMCA). I fail to see how this situation is any different?
A country that doesn’t get on top of this issue will doom itself to cultural irrelevance in the long term, IMO.
Yes, but we don't have copyright laws to protect the environment. So harm to the environment is not a copyright argument.
> Because I’m not a machine. I’m contrained by physics, whereas these models are not.
I don't even know what that means.
> But a painting, a book, a song, etc are easy to steal, especially with technology.
The AI does not merely duplicate training samples. That said, effort is also unrelated to copyright.
Typically, creators are very protective of this sort of thing, unless it stays in the area of fan art. If anyone tries to seriously monetize this kind of output, I'm sure we'll see a lot of cases.
Imagine what Disney would do if you used DALL-E to create an animated feature film in the style of Mickey Mouse, but with cats instead of mice, and they found out you used actual footage from, say, Fantasia to train an AI model. No idea if they would win, but I'm certain they'd sue.
E.g. The Shining picture of Jack Nicholson with the door isn't representative of the "look" of the film, but very much an iconic still frame and basically what you see in a Google image search for "The Shining".
Although if an individual created all of these then that's about the same amount of impressive
- Were the prompts shown the ones fed to DALL-E 2 or were there more complex details described in the prompt?
- Were these the first images generated for the prompt, or did the author generate many images and cherry-pick the best example, and if so from how many?
The first one, I’d have a huge problem with. Lying about prompts is a no no. Thankfully there’s not much incentive to lie.
Here's what I got for "A still of Kermit The Frog in Blade Runner 2049 (2017)":
https://imgur.com/a/y7t3RKx
We are at the precipice of someone releasing a $100M blockbuster movie just based on the language in the script with zero cost beyond compute.
What will this mean for the future of entertainment…
Kermit in Debby Does Dallas. Kermit in the Graduate. Kermit with 2 broken flippers. Oh the depravity. I'm not sure getting high quality visualizations of any random passing thought is a good idea ;-)
Just imagine how much lonelier the world is going to feel when people don't even have entertainment in common anymore.
But I would also love to see the perfect show for a combination of people, like my loved ones, so we can watch together.
And lastly, I would love to see the perfect show for a large majority of the population, because there is something special about a shared experience like Game of Thrones or Friends. To this day I can shout “pivooooot” and most everyone gets it.
On the other hand, I feel like this will ultimately be kinda like traditional procgen algorithms: once you've seen enough of what it produces it all starts feeling very bland and same-y. Sure, the AI may be able to produce a feature-length movie based on the input "What if Nicolas cage had played The Terminator and Aaron Sorkin wrote the script?", but somehow none of it would be surprising or interesting to you, it would lack the novelty and playfulness of a good human creative work, and it likely would be very shallow in its themes.
On the gripping hand though, perhaps in achieving that level of sophistication we inadvertently create something more alive and aware than we intended and instead of merely trying to produce satisfactory results it actually attempts to express itself in ways that resonate with us.
Simply tune the parameters associated with novelty and playfulness and you’ll get the desired result.
There’s nothing inherent in human creativity that can’t be replicated by an AI. Most creative work is derivative and remix’s prior art.
This is a good short video on the phenomenon of remixing https://m.youtube.com/watch?v=MZ2GuvUWaP8
I didn't say otherwise, I'm just skeptical the hypothetical system in question would be sophisticated enough to produce it in faster-than-human time scales.
Nah, you should be able to add “…but with some novelty and playfulness” to the prompt.
You can't handle the future! We live in a world that has time machines, and those time machines have to be manned by robots! Who's gonna do it? You?
"Like Facebook but like make it not suck"
AI: "Here you go!"
https://user-images.githubusercontent.com/1332366/171921054-...
[1]: https://twitter.com/HvnsLstAngel/status/1531774195234791424?...
[2]: https://duckduckgo.com/?t=ffab&q=eraserhead+baby&iax=images&...
DALL-E would just shoot back a still with Kyle McLachlan in it. He's already so Kermit like!
Don't get me wrong, they are still impressive in the quality of the visual they produce, but just like Markov Chain demos of old, they're neat but way miss the mark.
None of these capture the "feel" of Kermit the Frog. Most of them look like weird designs for the Ninja Turtles movie in the 90s.
There are several distinctive features of Kermit that a missing from nearly all of these.
- For any of the "live action" ones, Kermit should still always be a puppet. - Kermit notoriously has lanky arms, - Kermit never has eye lids - His eyes sit way on top of his head. - He often has his weird neck decoration. - His eyes have a very distinctive pupil shape.
None of these get Kermit correct, they all just look like frogs (maybe Dalle2 isn't trained on copyrighted/trademarked material?)
There are fan made versions of some of these which show just how different Dalle2 is from human imagination:
Kermit actually has been on family guy: https://static.wikia.nocookie.net/muppet/images/7/71/Famguys...
There are several "Kermit in Star Wars Examples" here are two: https://i.kym-cdn.com/entries/icons/original/000/021/668/ker..., https://i.ytimg.com/vi/6MebZx-4950/maxresdefault.jpg
Again if this was done on someone's laptop it would be really impressive. However the fact that so much talent and resources were poured into pushing AI to it's limits and this is what we get tells me we've hit another brick wall as far as research goes.
> - For any of the "live action" ones, Kermit should still always be a puppet. - Kermit notoriously has lanky arms, - Kermit never has eye lids - His eyes sit way on top of his head. - He often has his weird neck decoration. - His eyes have a very distinctive pupil shape.
In other words, it has done as good a job of costume design, lighting and photographing a live action Kermit as New Line Cinema paid $13.5m to accomplish for TMNT in 1990.
And you know who they got to do those creature designs?
Jim Henson
So maybe we shouldn't be so dismissive.
You might be missing the point of what OpenAI is doing. The point is to show off the capability of their models in a way that's likely to go viral and lead to more business for OpenAI. Some people laughed at GPT-3's silly demos, but when they launched GitHub Copilot...
If people say Dalle can improve the workflow of digital artists, sure, but Copilot hasn't revolutionized programming either, you still have to be a good programmer to finish whatever you are doing:
> A paper accepted for publication in the IEEE Symposium on Security and Privacy in 2022 assessed the security of code generated by Copilot [...] The study found that across these axes in multiple languages, 39.33% of top suggestions and 40.73% of total suggestions lead to code vulnerabilities. Additionally, they found that small, non-semantic (i.e., comments) changes made to code could impact code safety.[14]
What happened next? Is anyone using copilot for serious work? Has it changed programming in a fundamental way?
I personally have zero use for copilot since the for type of code I write the actual code writing is not a bottle neck, so automating that process is of no value to me. On top of that getting the details exactly right is essential so the ratio of boiler plate to real code is very, very low for me.
For example, your Star Wars example...
https://i.ytimg.com/vi/6MebZx-4950/maxresdefault.jpg
It's clearly just an existing photo of Kermit pasted over an image from the film. There are even two sets of arms. I could Photoshop that in a few minutes.
Then, the Dalle2 image...
https://pbs.twimg.com/media/FUEDDm2UEAAO8yb?format=jpg&name=...
I think it's impressive. It looks like Kermit is a character in the Star Wars universe. There are a few issues with the eyes and feet, and it's also hard to tell if it's a creature or a person in a frog suit. However, it gets 90% of the way there, and the pose is great for a frog/human hybrid.
The most exciting thing is how this could be used as a starting point for design. I could take the Dalle2 Kermit image above, fix the eyes/feet, add a few distinctive Kermit features, and have a great piece of concept art in an hour, rather than taking a day or two to create something from scratch. Obviously it can't be applied to all workflows, but for those it's suited for, it'll save vast amounts of time and costs. For that reason, it's already something of real value in its current state. The same can't be said about the Star Wars examples you provided.
However, I don't think you're correct in your assessment of the import of this sort of thing: it's an imagination machine. This isn't a brick wall, it's a foundation on which to build.
There are many ways to define what "Kermit the Frog in $MOVIE" means, and the choice the AI made is absolutely valid. There are of course various other valid choices, but this doesn't invalidate the ones presented.
Furthermore, judging by some other examples in this HN thread, it seems that the fact most of the pictures are not puppets is more of a choice of the human choosing the photos, as in other cases DALL-E was indeed adding puppet-like characters in movie-like decors.
"Sure, this AI can produce high-resolution realistic images leaps and bounds above anything that's been shown before... but there's an aspect which could use improvement. Obviously, this proves that the current AI technology will never amount to anything and we should just give up on it now."
If we accept that a model trained on copyright material does not infringe on the materials rights, then circumventing all copyright can be as simple as creating a sufficiently close derivative and giving it away.
Not to say that copyright is good to begin with.
That's correct. People do this all the time, sans the giving it away part.
Also, there is no way that you can argue these images are not transformative.
Exactly. Kermit has been very much transformed, and "in the style of" is not copyright infringement AFAIK.
Copyright law is way more nuanced than you think, especially around fair use.
Note that, unlike patents and trademarks, copyright doesn't require registration. The court could have found that, while the AI can't assert copyright, the copyright still exists and belongs to the owner of the AI or the designer of the AI etc. Instead, they found that the work can't be protected by copyright, since it is not the product of human creativity.
Edit: here is a citation from the article that I'm basing the previous assertions on:
> Both in its 2019 decision and its decision this February, the USCO found the “human authorship” element was lacking and was wholly necessary to obtain a copyright, Engadget’s K. Holt wrote. Current copyright law only provides protections to “the fruits of intellectual labor” that “are founded in the creative powers of the [human] mind,” the USCO states.
Stylistic inspiration is not an infringement of copyright, in either that case or the "do it on a computer" case here.
The Kermit the Frog aspect though is interesting - it applies equally for both the human and machine made works - if an argument could be made that the subject of the work sufficiently resembles the character, maybe there's a trademark issue at hand?
But in any scenario, nothing legally novel about the work being created by machine.
…except for the fact that it was created by a machine.
Just like copyright law had to be revised to deal with software and the internet, it will need to be revised to deal with AI.
We've already had this for years. The photos you take on any modern smart phone are partially the invention of AI (doubly so if you use something like portrait or night mode). It's not just raw CCD output, and yet, the photographer retains the copyright.
DALL-E's terms could require users to assign copyright. If not, I don't see any reason it wouldn't go to the person who came up with the prompt and picked one particular generation.
If I take a picture of a mountain with a camera, neither the mountain nor the camera hold copyright. DALL-E's just another tool in the toolbox.
Dall-E is the complete opposite. My contribution is insignificant (a prompt), whereas Dall-Es is 99% of the work.
> If I take a picture of a mountain with a camera, neither the mountain nor the camera hold copyright. DALL-E's just another tool in the toolbox.
Not sure how this contradicts my point. Nobody “owns” that mountain, so nobody could call your photo of it a derivative work.
If you take a photo of a frame of a movie, then who owns the rights to that photo? It’s not a raw frame, it’s a photo of a screen. Does that mean you didn’t just steal someone’s IP?
These questions need answers, and copyright laws try to answer them in as fair a way as possible. Dall E raises new questions, and thus we need to update copyright laws with new answers.
My personal view is that obtaining licensing for every single work used to train the model is impossible. So instead of simply making this type of AI research illegal, the fairest solution is to put that model into the public domain.
There are still financial incentives to perform this research, as the discoveries could be used later to train a commercial model with proper licensing.
I disagree entirely with this. Picking a prompt and potentially going through dozens, hundreds, or thousands of variations and re-generations has artistic value.
Just like the photograph of a mountain. Nature did most of the work, but a human selecting an angle, lighting, etc. from the nearly infinite combinations available matters.
We've also established in court that monkeys can't hold copyright (https://en.wikipedia.org/wiki/Monkey_selfie_copyright_disput...), and they're quite a bit more sentient than DALL-E.
https://archive.ph/89lqw
This shows the limits of the Turing test. To pass it a program must not only be smart enough, it must be dumb enough too.
Pulling what DALL-E does is a tell-tale sign it’s most likely not human, and would make it fail the test.
For example, looking at the WALL-E one [0], you can clearly see that the hands and feet aren't actually separated properly. There is also plenty of missing "logic" around the armpits. These are the kinds of mistakes a human can't make - especially one that is so adept at drawing the other parts so perfectly.
[0] https://twitter.com/HvnsLstAngel/status/1531512163738669057/...
There may still be a few anatomic mistakes, but many artists also make some. The picture quality, lightning, and the way it capture the graphic essence and mood of these movies is just amazing and beyond what even the most talented artists can pull out in the same time frame.
2022: This machine fails the Turing test, it's way too smart! No human could be this good at creating art.
"To me those are clearly Kermit the Frogs."
"To me those are clearly not Kermit the Frogs."
Then there's nothing really to argue about. Instead we can discuss what we see and how that affects our subjective perceptions.
For example, Kermit the Frog doesn't have eyelids, but most of these images show a frog with eyelids.
I don't have access to DALL-E 2, but I wonder if a prompt like "A cameo from Kermit the Frog in ..." would give more literal Kermits.
But people who do viral news and posts don't...read. So, their impact will continue to go unnoticed in comparison to DeEpFaKeS and Dall-E.
I understand computers and I understand back-propagation but this... it feels like magic to me.
Can someone indulge me in a short explanation of how this works and how is it this good?
- https://www.assemblyai.com/blog/how-dall-e-2-actually-works/
- https://www.youtube.com/watch?v=F1X4fHzF4mQ