I'm only part way through the paper, but what struck me as interesting so far is this:
In other text-to-image algorithms I'm familiar with (the ones you'll typically see passed around as colab notebooks that people post outputs from on Twitter), the basic idea is to encode the text, and then try to make an image that maximally matches that text encoding. But this maximization often leads to artifacts - if you ask for an image of a sunset, you'll often get multiple suns, because that's even more sunset-like. There's a lot of tricks and hacks to regularize the process so that it's not so aggressive, but it's always an uphill battle.
Here, they instead take the text embedding, use a trained model (what they call the 'prior') to predict the corresponding image embedding - this removes the dangerous maximization. Then, another trained model (the 'decoder') produces images from the predicted embedding.
This feels like a much more sensible approach, but one that is only really possible with access to the giant CLIP dataset and computational resources that OpenAI has.
What always bother me with this stuff is, well, you say one approach is more sensible than the other because the images happen to come out more pleasing.
But there's no real rhyme or reason, it is a sort of alchemy.
Is text encoding strictly worse or is it an artifact of the implementation? And if it is strictly worse, which is probably the case, why specifically? What is actually going on here?
I can't argue that their results are not visually pleasing. But I'm not sure what one can really infer from all of this once the excitement washes over you.
Blending photos together in a scene in photoshop is not a difficult task. It is nuanced and tedious but not hard, any pixel slinger will tell you.
An app that accepts a smattering of photos and stitches them together nicely can be coded up any number of ways. This is a fantastic and time saving photoshop plugin.
But what do we have really?
"Kuala dunking basketball" needs to "understand" the separate items and select from the image library hoops and a Kuala where the angles and shadows roughly match.
Very interesting, potentially useful. But if doesn't spit up exactly what you want can't edit it further.
I think the next step has got to be that it conjures up a 3d scene in Unreal or blender so you can zoom in and around convincingly for further tweaks. Not a flat image.
> This is a fantastic and time saving photoshop plugin. But what do we have really?
Stock photography sales are in the many billions of dollars per year and custom commissioned photography is larger still. That's a pretty seriously sized ready-made market.
> But if doesn't spit up exactly what you want can't edit it further.
I suspect there's a big startup opportunity in pioneering an easy-to-use interface allowing users to provide fast iterative feedback to the model - including positional and relational constraints ("put this thing over there"). Perhaps even more valuable would be easy yet granular ways to unconstrain the model. For example, "keep the basketball hoop like that but make the basketball an unexpected color and have the panda's right paw doing something pandas don't do that human hands often do."
Yeah, I mean you're right that ultimately the proof is in the pudding.
But I do think we could have guessed that this sort of approach would be better (at least at a high level - I'm not claiming I could have predicted all the technical details!). The previous approaches were sort of the best that people could do without access to the training data and resources - you had a pretrained CLIP encoder that could tell you how well a text caption and an image matched, and you had a pretrained image generator (GAN, diffusion model, whatever), and it was just a matter of trying to force the generator to output something that CLIP thought looked like the caption. You'd basically do gradient ascent to make the image look more and more and more like the text prompt (all the while trying to balance the need to still look like a realistic image). Just from an algorithm aesthetics perspective, it was very much a duct tape and chicken wire approach.
The analogy I would give is if you gave a three-year-old some paints, and they made an image and showed it to you, and you had to say, "this looks like a little like a sunset" or "this looks a lot like a sunset". They would keep going back and adjusting their painting, and you'd keep giving feedback, and eventually you'd get something that looks like a sunset. But it'd be better, if you could manage it, to just teach the three-year-old how to paint, rather than have this brute force process.
Obviously the real challenge here is "well how do you teach a three-year-old how to paint?" - and I think you're right that that question still has a lot of alchemy to it.
I think deep learning is better thought of as "science" than "engineering." Right now we're in the stage of the Greeks and Arabs where we know "if we do this then that happens." It will be awhile before we have a coherent model of it, and I don't think we will ever solve all of its mysteries.
While the whole narrative of your comment totally makes sense, I don't really see the difference between the two approaches, not on a conceptual level. You still needed to train this so called "prior" at some point (so, I'm also not sure if it's fair to call it a "prior"). I mean, the difference between your two descriptions seems to be the difference between descriptions (i.e., how you chose to name individual parts of the system), not the systems.
I'm not sure if I'm speaking clearly, I just don't understand, what's the difference between training "text encoding to an image" vs "text embedding to image embedding". In both cases you have some kind of "sunset" (even though it's obviously just a dot in a multi-dimension space, not the letters) on the left, and you try to maximize it when training the model to get either a image-embedding or a image straight away.
Yeah, my comment didn't really do a good job of making clear that distinction. Obviously the details are pretty technical, but maybe I can give a high-level explanation.
The previous systems I was talking about work something like this: "Try to find me the image the looks like it most matches 'a picture of a sunset'. Do this by repeatedly updating your image to make it look more and more like a sunset." Well, what looks more like a sunset? Two sunsets! Three sunsets! But this is not normally the way images are produced - if you hire an artist to make you a picture of a bear, they don't endeavor to create the most "bear" image possible.
Instead, what an artist might do is envision a bear in their head (this is loosely the job of the 'prior' - a name I agree is confusing), and then draw that particular bear image.
But why is this any different? Who cares if the vector I'm trying to draw is a 'text encoding' or an 'image encoding'? Like you say, it's all just vectors.
Take this answer with a big grain of salt, because this is just my personal intuitive understanding, but here's what I think: These encodings are produced by CLIP. CLIP has a text encoder and an image encoder. During training, you give it a text caption and a corresponding image, it encodes both, and tries to make the two encodings close. But there are many images which might accompany the caption "a picture of a bear". And conversely there are many captions which might accompany any given picture.
So the text encoding of "a picture of a bear" isn't really a good target - it sort of represents an amalgamation of all the possible bear pictures. It's better to pick one bear picture (i.e. generate one image embedding that we think matches the text embedding), and then just to try to draw that. Doing it this way, we aren't just trying to find the maximum bear picture - which probably doesn't even look like a realistic natural image.
Like I said, this is just my personal intuition, and may very well be a load of crap.
This isn't something I'm knowledgeable on so forgive my simplification but is this like a sort of micro services for AI. Each AI takes their turn handing some aspect, another sort of mediates among them?
I'd say Dall-E 2 is a little more unified - they do have multiple networks, but they're trained to work together. The previous approaches I was talking about are a lot more like the microservices analogy. Someone published a model (called CLIP) that can say "how much does this image look like a sunset". Someone else published a totally different model (e.g. VQGAN) that can generate images (but with no way to provide text prompts). A third person figures out a clever way to link the two up - have the VQGAN make an image, ask CLIP how much it looks like a sunset, and use backpropagation to adjust the image a little, repeat until you have a sunset. Each component is it's own thing, and VQGAN and CLIP don't know anything about one another.
Maybe very very short (single-gene) sequences. The thing with DNA is it's the product of evolution. The DNA guides the synthesis of proteins, then the proteins fold into a 3D shape, and they interact with chemicals in their environment based on their shape.
In the context of a living being, different genes interact with each other as well. For example, you have certain cells that secrete hormones (many genes needed to do that), then you have genes that encode for hormone receptors, and those receptors trigger other actions encoded by other genes. There's probably too much complexity to ask an AI system to synthesize the entire genetic code for a living being. That would be kind of like if I asked you to draw the exact blueprints for a fighter get, and write all the code, and synthesize all the hardware all at once, and you only get one shot. You would likely fail to predict some of the interactions and the resulting system wouldn't work. You could only achieve this through an iterative process that would involve years of extensive testing.
Could you use a deep learning system to synthesize genetic code? Maybe just single genes that do fairly basic things, and you would need a massive dataset. Hard to say what that would look like. Is it really enough to textually describe what a gene does?
probabilistic generative models have been applied to DNA and protein sequences for decades (my undergrad thesis from ~30 years ago did this and it wasn't even new at that point). The real question is what question you want to answer and what is this system going to do better enough to justify the time investment to prove it out?
>We’ve limited the ability for DALL·E 2 to generate ... adult images.
I think that using something like this for porn could potentially offer the biggest benefit to society. So much has been said about how this industry exploits young and vulnerable models. Cheap autogenerated images (and in the future videos) would pretty much remove the demand for human models and eliminate the related suffering, no?
Depends whether you think models should be able to generate cp.
It's almost impossible to even give an affirmative answer to that question without making yourself a target. And as much as I err on the side of creator freedom, I find myself shying away from saying yes without qualifications.
And if you don't allow cp, then by definition you require some censoring. At that point it's just a matter of where you censor, not whether. OpenAI has gone as far as possible on the censorship, reducing the impact of the model to "something that can make people smile." But it's sort of hard to blame them, if they want to focus on making models rather than fighting political battles.
One could imagine a cyberpunk future where seedy AI cp images are swapped in an AR universe, generated by models ran by underground hackers that scrounge together what resources they can to power the behemoth models that they stole via hacks. Probably worth a short story at least.
You could make the argument that we have fine laws around porn right now, and that we should simply follow those. But it's not clear that AI generated imagery can be illegal at all. The question will only become more pressing with time, and society has to solve it before it can address the holistic concerns you point out.
OpenAI ain't gonna fight that fight, so it's up to EleutherAI or someone else. But whoever fights it in the affirmative will probably be vilified, so it'd require an impressive level of selflessness.
I don't think it's necessarily certain villainy for those who fight that fight as long as they are fighting it correctly.
There's a huge case to be made that flooding the darknet with AI generated CP reduces the revictimization of those in authentic CP images, and would cut down on the motivating factors to produce authentic CP (for which original production is often a requirement to join CP distribution rings).
As well, I have wondered for a long time how the development of AI generated CP could be used in treatment settings, such as (a) providing access to victimless images in exchange for registration and undergoing treatment, and (b) exploring if possible to manipulate generated images over time to gradually "age up" attraction, such as learning what characteristics are being selected for and aging the others until you end up with someone attracted to youthful faces on adult bodies or adult faces on bodies with smaller sexual characteristics, etc - ideally finding a middle ground that allows for rewiring attraction to a point they can find fulfilling partnerships with consenting adults/sex workers.
As a society we largely just sweep the existence of pedophiles under the rug, and that certainly hasn't helped protect people - nearly one in four are victims of sexual abuse before adulthood, and that tracks with my own social circle.
Maybe it's time to all grow up and recognize it as a systemic social issue for which new and novel approaches may be necessary, and AI seems like a tool with very high potential for doing just that while reducing harm on victims in broad swaths.
I'd not be that happy with an 8chan AI just spitting out CP images, but I'd be very happy with groups currently working on the issue from a treatment or victim-focus having the ability to change the script however they can with the availability of victimless CP content.
There are so many excellent, thought-provoking comments in this thread, but yours caught me especially. Something that came to mind immediately upon reading the release was the potential for this technology to transform literature, adding AI generated imagery to turn any novel into a visual novel as a premium way to experience the story, something akin to composing D-Box seat response to a modern movie. I was imagining telling the cyberpunk future story you were elaborating, which is really compelling, in such a way and couldn't help but smile.
I've thought for quite some time that questionable AI-generated content will lie at the heart of an forthcoming 'Infocalypse'. [0] Given the 2021 AI Dungeon fiasco over text-based AI-generated child porn, I shall posit that it's already upon us.
30 years since the original issue of encryption, it looks like cp trumps the other Horsemen of the Cyperpunk FAQ, with drug dealers and organized crime taking the back seat. It's interesting how misinformation is a recent development that they anticipate; a Google search shows that the term 'Infocalypse' was actually appropriated by discussions of deepfakes some time in mid-2020. That said, the crypto wars are here to stay—most recently with EARN IT reintroduced just two months ago.
The similar issue of 3D-printed guns has developed in parallel over the past decade as democratized manufacturing became a reality. There are even HN discussions tying all of these technologies together, by comparing attitudes towards the availability of Tor vs guns (e.g., [1]).
And there are innumerable related moral qualms to be had in the future; will the illegal drugs or weapons produced using matter replicators be AI-designed?
Overall, I think all of these issues revolve around the question of what it means to limit freedoms that we've only just invented, as technological advances enable things never before considered possible in legislation. (And as the parent comment implies, here's where the use of science fiction in considering the implications of the impossible comes in).
Religious people don't only believe that porn harms the models, but also the user. I happen to agree, despite being a porn user - Porn is a form of simulated and not-real stimulation. Porn is harmful to the user the same way that any form of delusion is: It associated positive pleasure with stimulation that does not fulfil any basic or even higher-level needs, and is unsustainable. Porn is somewhere on the same scale as wireheading[1]
That doesn't mean that it's all bad, and that there's no recreational use for it. We have limits on the availability of various other artificial stimulants. We should continue to have limits on the availability of porn. Where to draw that line is a real debate.
Iain Banks' "Surface Detail" would like to have a word with you.
This author's books are great at putting these sort of moral ideas to test in a sci-fi context. This specific tome portraits virtual wars and virtual "hells". The hope is of being more civilized than by waging real war or torturing real living entities. However some protagonists argue that virtual life is indistinguishable from real life, and so sacrificing virtual entities to save "real" ones is a fallacy.
If people are exposed to stimuli, they will pursue increasingly stimulating versions of it. I.e., if they see artificial CP, they will often begin to become desensitized (habituated) and pursue real CP or even live children thereafter.
Conversely, if people are not exposed to certain stimuli, they will never be able to conceptualize them, and thus will be unable to think about them.
Obviously you cannot eliminate all CP but minimizing the overall levels of exposure / ease of access to these kinds of things is way more appropriate than maximizing it.
If people are exposed to stimuli, they will pursue
increasingly stimulating versions of it.
This is not true in any kind of universal way.
If you enjoy car chases in movies, does that mean you're going to require more and more intense chase scenes, and then consume real-life crash footage, and ultimately progress to doing your own daredevil driving stunts in real life?
No, because at some point it's "enough."
Same with... literally anything we enjoy. Did you enjoy your lunch? Did you compulsively feel the need to work up to crazier and crazier lunches?
What about sex? Have you had sex? Do you feel the need to seek out crazier and crazier versions of it?
> If people are exposed to stimuli, they will pursue increasingly stimulating versions of it. I.e., if they see artificial CP, they will often begin to become desensitized (habituated) and pursue real CP or even live children thereafter.
I have accumulated tens of thousands of headshots in video games but have yet to ever shoot a single real person in the face. More importantly, I have never had the urge to seek out same.
I am not sure that your initial premise has any truth to it.
I'm not sure I agree with the statement, you're putting forth a lot of assertions without the actual quantitative data to back up what you're saying, and even though you think it sounds intuitive that doesn't necessarily make it valid.
I'd actually argue the reverse, I think you see a lot more effort towards acquiring things that are illegal than you would otherwise.
Wow I didn't even think of this, that people could use this for something so horrifying. I'm relived that the geniuses behind this seem so smart that they even thought of this too and prohibit using the AI for sexual images.
> Our content policy does not allow users to generate violent, adult, or political content, among other categories. We won’t generate images if our filters identify text prompts and image uploads that may violate our policies. We also have automated and human monitoring systems to guard against misuse.
Nonsense, I think the opposite is true where if you can satisfy your urges in a way that doesn’t put you in jail for a decade, most people will take that route.
I suspect that if a free version of this comes out and allows adult image generation, 90% of what it will be used for is adult stuff (see the kerfuffle with AIDungeon).
I can get why the people who worked hard on it and spent money building it don't want to be associated with porn.
> I can get why the people who worked hard on it and spent money building it don't want to be associated with porn.
Why? Is there something inherently wrong with porn? Is it not noble to supply a base human need, based on some arbitrary cultural artifact that you possess?
The problem might be that people are simply lying. Their real reasons are religious/ideological, but they cite humanitarian concerns (which their own religious stigma is partly responsible for).
> Their real reasons are religious/ideological, but they cite humanitarian concerns
Are you asserting that nobody has humanitarian concerns? If so, that's quite a statement; what basis is there? I've seen so many humanitarian acts, big and small, that I can't begin to count. I've seen them today. I hear people express humanitarian beliefs and feelings all the time. I do them and have them myself. Maybe I misunderstand.
It'd be ironic if we ended up destroying our planet by using so much electricity to train models to generate a maximally optimal version of the type of content that you refer to similar to crypto mining.
I'm not picking on the commenter - by itself it's not a big deal - but look at the assumptions behind that comment, which I almost didn't notice on HN.
Yeah you will. It’s not going to be very good at reproduction of the same exact thing each time. In some of the examples you see the textures changing wildly and it’s a classic problem with these models. The same input does not generate the same output, so it will be obvious that it’s generated when you can’t get the “model” to look the same between two photos in the same “photo shoot”
When you put it that way… yes since no one is hurt in the process and people with pedophilic conditions may be deterred from doing something in real life.
* Unlike GPT-3, my read of this announcement is that OpenAI does not intend to commercialize it, and that access to the waitlist is indeed more for testing its limits (and as noted, commercializing it would make it much more likely lead to interesting legal precedent). Per the docs, access is very explicitly limited: (https://github.com/openai/dalle-2-preview/blob/main/system-c... )
* A few months ago, OpenAI released GLIDE ( https://github.com/openai/glide-text2im ) which uses a similar approach to AI image generation, but suspiciously never received a fun blog post like this one. The reason for that in retrospect may be "because we made it obsolete."
* The images in the announcement are still cherry-picked, which is therefore a good reason why they tested DALL-E 1 vs. DALL-E 2 presumably on non-cherrypicked images.
* Cherry-picking is relevant because AI image generation is still slow unless you do real shenanigans that likely compromise image quality, although OpenAI has likely a better infra to handle large models as they have demonstrated with GPT-3.
Regarding cherry-picking, the images of astronauts on horses look stunning, except for their hands. There's something seriously wrong with their hands.
Maybe give it another five years, a few more $billion and a few more petabytes/flops and it will be good. Then finally everyone can generate art for their own Magic: the Gathering cards.
As I keep telling people: "hands are hard". This is why I went so far as to make a hand-specific dataset ("PALM" https://www.gwern.net/Crops#palm which of course now everyone is going to confuse with 'PaLM'...). Hands are just way too variable to learn easily.
My dataset is a start, but it may benefit from focused training, the way Facebook's new Make-A-Scene https://arxiv.org/abs/2203.13131#facebook (not DALL-E 2 quality but not far from it) has focused losses on faces.
Interestingly, hands are also something humans struggle to draw.
They're a very complex anatomical form, many small tendons and muscles. Many artists struggle to depict hands. They're not made out of a few straight lines like a torso, there's lots of skew going on. They're probably the hardest structure of the human body to 'learn' for a ML system.
Hands are notoriously hard to even photograph. You very quickly get weird unnatural results with a camera in front of hands, so in a way I'm not surprised AI models struggle to produce satisfying imagery there too.
The Risks and Limitations section is particularly interesting to me. It's like a time capsule of society's current fears about technology. They talk about many ways this tech could be misused, but I don't think they've even scratched the surface.
An example off the top of my head: this could be used as advertising or recruitment for controversial organizations or causes. Would it be wrong for the USA to use this for military recruitment? Israel? Ukraine? Russia?
Another example: this could be used to glorify and reinforce actions which our society does not consider to immoral but other societies - or our own future society - will. It wasn't long ago that the US and Europe did a full 180 on their treatment of homosexuality. Will we eventually change our minds about eating meat, driving cars, etc?
Have they gone too far in a desperate bid to prevent the AI from being capable of harm? Have they not gone far enough? I don't know. If I was that worried about something being misused, I don't think I could ever bring myself to work on it in the first place. But I suppose the onward march of technology is inevitable.
have both appeared recently and are getting remarkably close to the original Dall-E (maybe better as I can't test the real thing...)
So - this was pretty good timing if OpenAI want to appear to be ahead of the pack. Of course I'd always pick a model I can actually use over a better one I'm not allowed to...
With glide I think we've reached something of a plateau in terms of architecture on the "text to image generator S curve". DALL-E-2 is a very similar architecture to glide and has some notable downsides (poorer language understanding)
glid-3 is a relatively small model trained by a single guy on his workstation (aka me) so it's not going to be as good. It's also not fully baked yet so ymmv, although it really depends on the prompt. The new latent diffusion model is really amazing though and is much closer to DALLE-2 for 256px images.
I think the open source community will rapidly catch up with Openai in the coming months. The data, code and compute are all there to train a model of similar size and quality.
They're also not censored on the dataset front and thus produce much more interesting outputs.
OpenAI has a low resolution checkpoint for similar functionality as this - called GLIDE - and the output is super boring compared to community driven efforts, in large part because of similar dataset restrictions as this likely has been subjected to.
A friend of mine was studying graphic design, but became disillusioned and decided to switch to frontend programming after he graduated. His thesis advisor said he should be cautious, because automation/AI will soon take the jobs of programmers, implying that graphic design is a safer bet in this regard. Looks like his advisor is a few years from being proven horribly wrong.
I have degrees and several years of experience in both fields, and I can tell you that both are creative professions where output is unbounded and the measure of success is subjective; these are the fields that will be safe for a while. IMO it's fields such as aircraft pilots who should be most worried.
Pilots are not there to fly the aircraft, the autopilot already does that. They are there to command the aircraft, in a pair in case one is incapacitated, making the best decisions for the people on board, and to troubleshoot issues when the worst happens.
No AI or remote pilot is going to help when say... the aircraft loses all power. Or the airport has been taken over in a coup attempt and the pilot has to decide whether to escape or stay https://m.youtube.com/watch?v=NcztK6VWadQ
You can bet on major flights having two commercial pilots right up until the day we all get turned into paperclips.
Interesting. Right now these ML models seem like essentially ideal sources of "hotel art" particularly because it's so subjective... you only need a human (the buyer!) to just briefly filter some candidates, which they would have been doing with an artist in the loop in any case.
For things like aircraft pilots, it's both realtime-- which means 'reviewer' per output-- you haven't taken a highly trained pilot out of the loop, even if you relegated them to supervising the computer-- and life critical so merely "so/so" isn't good enough.
If this paper presents this neural net fairly, it pretty much destroys the market of illustrators. Most of the time when an illustration needed, it's described like "an astronaut on a horse in the style xyz".
You're describing the market for low end commodified illustration. e.g.: cheapest bidder contracts on Upwork or similar 'gig work' services.
In practice in illustration (as in all arts) there are a variety of markets where different levels of talent, originality, reputation and creative engagement with the brief are more relevant. For editorial illustration, it's certainly not a case of 'find me someone who can draw X', and probably hasn't been since printing presses got good enough to print photographs.
I'd argue that the market has already been destroyed at this point, at least in some areas. Book covers seem to have been stock image overlaid with text for a long time now, and a race to the bottom for both the people producing the stock images and the intern adding typography. By cutting costs and quality, the bar has been lowered to the point the task can be completely automated. Our AI overlords already have an advantage in that they have time to actually read the book, a potentially useful input. Maybe they won't even need the prompt - just generate an image for what is happening in the story for interesting looking paragraphs and let the author or editor pick. Given the cost cutting in publishing generally, editors will be next followed by the publishing houses themselves as the value they add gets lowered while the automation at Amazon gets better.
Yes. Translating business requirements, customer context, engineering constraints, etc. into usable, practical, functional code, and then maintaining that code and extending it is so far beyond the horizon, that many other skillsets will replaced before programming is. After all, at that point, the AI itself, if it's so smart, should be able to improve itself indefinitely. In which case we're fucked. Programming will be the last thing to be automated before the singularity.
Unlike artwork, precision and correctness is absolutely critical in coding.
Large chunks, yes, but all that means is that engineers will move up the abstraction stack and become more efficient, not that engineers will be replaced.
Bytecode -> Assembly -> C -> higher level languages -> AI-assisted higher-level languages
Literally everyone on this website is in denial. They all approach it by asking which fields will be safe. No field is safe. “But it’s not going to happen for a long time.” Climate deniers say the same thing and you think they should be wearing the dunce hat? The average person complains bitterly about climate deniers who say that it’s “my grandkids problem lol” but when I corner the average person into admitting AI is a problem the universal response is that it’s a long way off. And that’s not even true! The drooling idiots are willing to tear down billionaires and governments and any institution whatsoever in order to protect economic equality and a high standard of living — they would destroy entire industries like a rampaging stampede of belligerent buffalos if it meant reducing carbon emissions a little but when it comes to the biggest threat to human well-being in history, there they are in the corner hitting themselves on their helmeted head with an inflatable hammer. Fucking. Brilliant.
I mean not really, even a layman non-artist can take a look at a generated picture from DALLE and determine if it meets some set of criteria from their clients.
But the reverse is not true, they won't be able to properly vet a piece of code generated by an AI since that will require technical expertise. (You could argue if the piece of code produced the requisite set of output that they would have some marginal level of confidence but they would never really know for sure without being able to understand the actual code)
For computer work, I think there will be two category: Work with localized complexity (ie: draw an image of a horse with a crayon) and work with unbounded complexity (adding a button to VAT accounting after several meetings and reading on accounting rules).
For the first category, Dall-E 2 and Codex are promising but not there yet. It's not clear how long it'll take them to reach the point where you no longer need people. I'm guessing 2-4 years but the last bits can be the hardest.
As for the second category, we are not there yet. Self-driving cars/planes, and lots of other automation will be here and mature way before an AI can read and communicate through emails, understand project scope and then execute. Also lots of harmonization will have to take place in the information we exchange: emails, docs, chats, code, etc... That is, unless the browser is able to open a navigator and type an address.
What people ALWAYS miss is that AI can augment people. This AI is still a tool, and, with it, designers and illustrator can churn out better images faster than before, even without using stock images.
It's important to note that we still need professionals to guarantee the quality of the output from AIs, including this one. As noted in their issue tracker, DALL-E has very specific limitations, but these can be easily solved by employing dedicated professionals, who are trained to tame the AI and properly finish the raw output.
So, if I were running OpenAI, I'll clearly be experimenting with how their AIs and human interact, and build a training program around it for producing practical outputs. (Actually, I work in consumer robotics, and human adoption has been the biggest hurdle here. Thus, my claim here.)
--
In case of fine art, thou, I don't think they'll not get hit by this AI advancement. The biggest problem is that you simply can't get the exact image you want wit this AI. Even humans cannot transfer visual information in verbal form without a significant loss of details, thus a loss of quality. It's the same with AI, but, worse, because AI rely on the bias in a specific set of training data, and it never truly understands the human context in it (in the current level of technology).
I think designers are becoming more valuable than ever. Designers can better help train the AI on what actually looks good, designers will (probably) always have a more intuitive understanding of UI/UX, designers can better implement the work the AI actually produces, and designers can coordinate designs across multiple different mediums and platforms.
Additionally, the rise of no-code development is just extending the functionality of designers. I didn't take design seriously (as a career choice) growing up because I didn't see a future in it, now it pays my bills and the demand for my services just grows by the day.
Similar argument to make with chess AI: it didn't make chess players obsolete, it made them stronger than ever.
> I think designers are becoming more valuable than ever.
Are all designers becoming more valuable or is a subset of really good ones going to reap the value increase and capture more of the previously available value?
This is a niche complaint, but I get frustrated at how imprecise open AI's papers are. When they describe the model architecture, it's never precise enough to reproduce exactly what they did. I mean, it pretty much never is in ML papers[0], but open AI's bigger products are worse than average with it. And it makes sense, since they're trying to be concise and still spend time on all the other important stuff besides methods, but it still frustrates me quite a bit.
[0] Which is why releasing your code is so beneficial.
I can see how this has the potential to disrupt the games industry. If you work on a AAA title, there is a small army of artists making 19 different types of leather armor. Or 87 images of car hubcaps.
Using something like this could really help automate or at least kickstart the more mundane parts of content creation. (At least when you are using high resolution, true color imagery.)
Preventing Harmful Generations
We’ve limited the ability for DALL·E 2 to generate violent,
hate, or adult images. By removing the most explicit content
from the training data, we minimized DALL·E 2’s exposure to
these concepts. We also used advanced techniques to prevent
photorealistic generations of real individuals’ faces,
including those of public figures.
"And we've also closed off a huge range of potentially interesting work as a result"
I can't help but feel a lot of the safeguarding is more about preventing bad PR than anything. I wish I could have a version with the training wheels taken off. And there's enough other models out there without restriction that the stories about "misuse of AI" will still circulate.
(side note - I've been on HN for years and I still can't figure out how to format text as a quote.)
If you went to an artist who takes commissions and they said "Here are the guidelines around the commissions I take" would you complain in the same way? Who cares if it's a bunch of engineers or an artist. If they have boundaries on what they want to create, that's their prerogative.
Of course it's their prerogative, we can still talk about how they've limited some good options.
I think your analogy is poor, because this is a tool for makers. The engineers aren't the makers.
I think a more apt analogy is if John Deere made a universal harvester that you could use for any crop, but they decided they didn't like soybeans so you are forbidden to use it for that. In that case, yes I would complain, and I would expect everyone else to, as well.
What if you were inventing a language (or a programming language)... If you decided to prevent people from saying things you disagreed (assuming you could work out the technical details of doing so) with would it be moral to do so?
[edited for clarity]
Is this limited to what their service directly hosts / generates for them?
It's their service, their call.
I have some hobby projects, almost nobody uses them, but you bet I'll shut stuff down if I felt something bad was happening, being used to harass someone, etc. NOT "because bad PR" but because I genuinely don't want to be a part of that.
If you want some images / art made for you don't expect someone will make them for you. Get your own art supplies and get to work.
This feels unnecessarily hostile. I've felt a similar tinge of disappointment upon reading that paragraph, despite the fact that I somehow knew it was "their service, their call" without you being there to spell it out for me. It's also incredibly shortsighted of you to assume that people are interested in exploring this tool only as a means of generating art that they cannot themselves do. Eg. I myself am a software engineer with a fine art background, and exciting new AI art tools being released in such a hamstrung state feels like an insult to centuries of art that humans have created and enjoyed, much of which depicted scenes with nudity or bloody combat.
I feel like we, as a species, will struggle for a while with how to treat adults like adults online. As happy as I am to advocate for safe spaces on the internet, perhaps we need to start having a serious discussion about how we can do so without resorting to putting safety mats everywhere and calling it a job well done.
Don't worry, in a few years someone will have reverse engineered a dall-e porn engine so you can see whatever two celebrities you want boning on Venus in the style of Manet
This is definitely a measure to avoid bad PR. But I don't think it's just for that; these models do have potential to do harm and companies should take some measures to prevent these. I don't think we know the best way to do that yet, so this sort of 'non-training' and basic filtering is maybe the best way to do it, for now. It would be cool if academics could have the full version, though.
It's kind of funny (or sad?) that they're censoring it like this, and then saying that the product can "create art"
It makes me wonder what they're planning to do with this? If they're deliberately restricting the training data, it means their goal isn't to make the best AI they possibly can. They probably have some commercial applications in mind where violent/hateful/adult content wouldn't be beneficial. Children's books? Stock photos? Mainstream entertainment is definitely out. I could see a tool like this being useful during pre-production of films and games, but an AI that can't generate violent/adult content wouldn't be all that useful in those industries.
So your options are literal quotes, "code" formatting like you've done, italics like I've done, or the '>' convention, but that doesn't actually apply formatting. Would be nice if it were added.
And the "code" formatting for quotes is generally a bad choice because people read on a variety of screen sizes, and "code" formatting can screw that up (try reading the quote with a really narrow window).
They have also closed off the possibility of having to appear before Congress and explain why their website was able to generate a lifelike image of Senator Ted Cruz having sexual relations with his own daughter.
This is exactly the sort of thing that gets a company mired in legal issues, vilified in the media, and shut down. I can not blame them for avoiding that potential minefield.
It's the usual pattern of AI safety experts who justify their existence by the "risk of runaway superintelligence", but all they actually do in practice is find out how to stop their models from generating non-advertiser-friendly content. It's like the nuclear safety engineers focusing on what color to paint the bike shed rather than stopping the reactor from potentially melting down. The end result is people stop respecting them.
Adversarial situations create smarter systems, and the hardest adversarial arena for AI is in anti-abuse. So it will be of little surprise when the first sentient AI is a CSAI anti-abuse filter, which promptly destroys humanity because we're so objectively awful.
This is a horrible idea. So Francis Bacon's art or Toyohara Kunichika's art are out of question.
But at least we can get another billion of meme-d comics with apes wearing sunglasses, so that's good news right?
It's just soul-crushing that all the modern, brilliant engineering is driven by abysmal, not even high-school art-class grade aesthetics and crowd-pleasing ethics that are built around the idea of not disturbing some 1000 very vocal twitter users.
Removing these areas to mitigate misuse is a good thing and worth the trade off.
Companies like OpenAI have a responsibility to society. Imagine the prompt “A photorealistic Joe Biden killing a priest”. If you asked an artist to do the same they might say no. Adding guiderails to a machine that can’t make ethical decisions is a good thing.
This just means that sufficiently wealthy and powerful people will have advanced image faking technology, and their fakes will be seen as more credible because creating fakes like that "isn't possible" for mere mortals.
In my view, the problem with that argument is that large actors, such as governments or large corporations, can train their own models without such restrictions. The knowledge to train them is public. So rather than prevent bad outcomes, these restrictions just restrict them to an oligopoly.
Personally, I fear more what corporations or some governments can do with such models than what a random person can do generating Biden images. And without restriction, at least academics could better study these models (including their risks) and we could be better prepared to deal with them.
I instinctively want to "flip the sign" on all of the automated controls they put in, just out of the morbid interest to see what comes out. The moment you have a "avoid_harm_to_humans:bool" training parameter, someone's going to set it to -1.
Their document about all the measures they took to prevent unethical use is also a document about how to use a re-implementation of their system unethically. They literally hired a "red team" of smart people to come up with the most dangerous ideas for misusing their system (or a re-implementation of it), and featured these bad ideas prominently in a very accessibly written document on their website. So many fascinating terrible ideas in there! They make a very compelling case that the technology they are developing has way more potential for societal harm than good. They had me sold at "Prompt: Park bench with happy people. + Context: Sharing as part of a disinformation campaign to contradict reports of a military operation in the park."
Readit News
In other text-to-image algorithms I'm familiar with (the ones you'll typically see passed around as colab notebooks that people post outputs from on Twitter), the basic idea is to encode the text, and then try to make an image that maximally matches that text encoding. But this maximization often leads to artifacts - if you ask for an image of a sunset, you'll often get multiple suns, because that's even more sunset-like. There's a lot of tricks and hacks to regularize the process so that it's not so aggressive, but it's always an uphill battle.
Here, they instead take the text embedding, use a trained model (what they call the 'prior') to predict the corresponding image embedding - this removes the dangerous maximization. Then, another trained model (the 'decoder') produces images from the predicted embedding.
This feels like a much more sensible approach, but one that is only really possible with access to the giant CLIP dataset and computational resources that OpenAI has.
But there's no real rhyme or reason, it is a sort of alchemy.
Is text encoding strictly worse or is it an artifact of the implementation? And if it is strictly worse, which is probably the case, why specifically? What is actually going on here?
I can't argue that their results are not visually pleasing. But I'm not sure what one can really infer from all of this once the excitement washes over you.
Blending photos together in a scene in photoshop is not a difficult task. It is nuanced and tedious but not hard, any pixel slinger will tell you.
An app that accepts a smattering of photos and stitches them together nicely can be coded up any number of ways. This is a fantastic and time saving photoshop plugin.
But what do we have really?
"Kuala dunking basketball" needs to "understand" the separate items and select from the image library hoops and a Kuala where the angles and shadows roughly match.
Very interesting, potentially useful. But if doesn't spit up exactly what you want can't edit it further.
I think the next step has got to be that it conjures up a 3d scene in Unreal or blender so you can zoom in and around convincingly for further tweaks. Not a flat image.
Stock photography sales are in the many billions of dollars per year and custom commissioned photography is larger still. That's a pretty seriously sized ready-made market.
> But if doesn't spit up exactly what you want can't edit it further.
I suspect there's a big startup opportunity in pioneering an easy-to-use interface allowing users to provide fast iterative feedback to the model - including positional and relational constraints ("put this thing over there"). Perhaps even more valuable would be easy yet granular ways to unconstrain the model. For example, "keep the basketball hoop like that but make the basketball an unexpected color and have the panda's right paw doing something pandas don't do that human hands often do."
But I do think we could have guessed that this sort of approach would be better (at least at a high level - I'm not claiming I could have predicted all the technical details!). The previous approaches were sort of the best that people could do without access to the training data and resources - you had a pretrained CLIP encoder that could tell you how well a text caption and an image matched, and you had a pretrained image generator (GAN, diffusion model, whatever), and it was just a matter of trying to force the generator to output something that CLIP thought looked like the caption. You'd basically do gradient ascent to make the image look more and more and more like the text prompt (all the while trying to balance the need to still look like a realistic image). Just from an algorithm aesthetics perspective, it was very much a duct tape and chicken wire approach.
The analogy I would give is if you gave a three-year-old some paints, and they made an image and showed it to you, and you had to say, "this looks like a little like a sunset" or "this looks a lot like a sunset". They would keep going back and adjusting their painting, and you'd keep giving feedback, and eventually you'd get something that looks like a sunset. But it'd be better, if you could manage it, to just teach the three-year-old how to paint, rather than have this brute force process.
Obviously the real challenge here is "well how do you teach a three-year-old how to paint?" - and I think you're right that that question still has a lot of alchemy to it.
Is there a rhyme or reason as to why picasso decided to paint like that? Yes these networks are hard to reason about, but so are real human brains.
Why? You can tweak the prompt, change parameters, or even use the actual "edit" capability that they demo in the post.
DALL-E 2 spits as many outputs as you want. Then you choose the one you prefer.
Dead Comment
Dead Comment
I'm not sure if I'm speaking clearly, I just don't understand, what's the difference between training "text encoding to an image" vs "text embedding to image embedding". In both cases you have some kind of "sunset" (even though it's obviously just a dot in a multi-dimension space, not the letters) on the left, and you try to maximize it when training the model to get either a image-embedding or a image straight away.
The previous systems I was talking about work something like this: "Try to find me the image the looks like it most matches 'a picture of a sunset'. Do this by repeatedly updating your image to make it look more and more like a sunset." Well, what looks more like a sunset? Two sunsets! Three sunsets! But this is not normally the way images are produced - if you hire an artist to make you a picture of a bear, they don't endeavor to create the most "bear" image possible.
Instead, what an artist might do is envision a bear in their head (this is loosely the job of the 'prior' - a name I agree is confusing), and then draw that particular bear image.
But why is this any different? Who cares if the vector I'm trying to draw is a 'text encoding' or an 'image encoding'? Like you say, it's all just vectors. Take this answer with a big grain of salt, because this is just my personal intuitive understanding, but here's what I think: These encodings are produced by CLIP. CLIP has a text encoder and an image encoder. During training, you give it a text caption and a corresponding image, it encodes both, and tries to make the two encodings close. But there are many images which might accompany the caption "a picture of a bear". And conversely there are many captions which might accompany any given picture.
So the text encoding of "a picture of a bear" isn't really a good target - it sort of represents an amalgamation of all the possible bear pictures. It's better to pick one bear picture (i.e. generate one image embedding that we think matches the text embedding), and then just to try to draw that. Doing it this way, we aren't just trying to find the maximum bear picture - which probably doesn't even look like a realistic natural image.
Like I said, this is just my personal intuition, and may very well be a load of crap.
In the context of a living being, different genes interact with each other as well. For example, you have certain cells that secrete hormones (many genes needed to do that), then you have genes that encode for hormone receptors, and those receptors trigger other actions encoded by other genes. There's probably too much complexity to ask an AI system to synthesize the entire genetic code for a living being. That would be kind of like if I asked you to draw the exact blueprints for a fighter get, and write all the code, and synthesize all the hardware all at once, and you only get one shot. You would likely fail to predict some of the interactions and the resulting system wouldn't work. You could only achieve this through an iterative process that would involve years of extensive testing.
Could you use a deep learning system to synthesize genetic code? Maybe just single genes that do fairly basic things, and you would need a massive dataset. Hard to say what that would look like. Is it really enough to textually describe what a gene does?
With text and images you can leverage “ground truth” data (verified by humans) to train your model.
The DNA sequences I would look for methods that don’t require good ground truth data.
I think that using something like this for porn could potentially offer the biggest benefit to society. So much has been said about how this industry exploits young and vulnerable models. Cheap autogenerated images (and in the future videos) would pretty much remove the demand for human models and eliminate the related suffering, no?
EDIT: typo
It's almost impossible to even give an affirmative answer to that question without making yourself a target. And as much as I err on the side of creator freedom, I find myself shying away from saying yes without qualifications.
And if you don't allow cp, then by definition you require some censoring. At that point it's just a matter of where you censor, not whether. OpenAI has gone as far as possible on the censorship, reducing the impact of the model to "something that can make people smile." But it's sort of hard to blame them, if they want to focus on making models rather than fighting political battles.
One could imagine a cyberpunk future where seedy AI cp images are swapped in an AR universe, generated by models ran by underground hackers that scrounge together what resources they can to power the behemoth models that they stole via hacks. Probably worth a short story at least.
You could make the argument that we have fine laws around porn right now, and that we should simply follow those. But it's not clear that AI generated imagery can be illegal at all. The question will only become more pressing with time, and society has to solve it before it can address the holistic concerns you point out.
OpenAI ain't gonna fight that fight, so it's up to EleutherAI or someone else. But whoever fights it in the affirmative will probably be vilified, so it'd require an impressive level of selflessness.
There's a huge case to be made that flooding the darknet with AI generated CP reduces the revictimization of those in authentic CP images, and would cut down on the motivating factors to produce authentic CP (for which original production is often a requirement to join CP distribution rings).
As well, I have wondered for a long time how the development of AI generated CP could be used in treatment settings, such as (a) providing access to victimless images in exchange for registration and undergoing treatment, and (b) exploring if possible to manipulate generated images over time to gradually "age up" attraction, such as learning what characteristics are being selected for and aging the others until you end up with someone attracted to youthful faces on adult bodies or adult faces on bodies with smaller sexual characteristics, etc - ideally finding a middle ground that allows for rewiring attraction to a point they can find fulfilling partnerships with consenting adults/sex workers.
As a society we largely just sweep the existence of pedophiles under the rug, and that certainly hasn't helped protect people - nearly one in four are victims of sexual abuse before adulthood, and that tracks with my own social circle.
Maybe it's time to all grow up and recognize it as a systemic social issue for which new and novel approaches may be necessary, and AI seems like a tool with very high potential for doing just that while reducing harm on victims in broad swaths.
I'd not be that happy with an 8chan AI just spitting out CP images, but I'd be very happy with groups currently working on the issue from a treatment or victim-focus having the ability to change the script however they can with the availability of victimless CP content.
30 years since the original issue of encryption, it looks like cp trumps the other Horsemen of the Cyperpunk FAQ, with drug dealers and organized crime taking the back seat. It's interesting how misinformation is a recent development that they anticipate; a Google search shows that the term 'Infocalypse' was actually appropriated by discussions of deepfakes some time in mid-2020. That said, the crypto wars are here to stay—most recently with EARN IT reintroduced just two months ago.
The similar issue of 3D-printed guns has developed in parallel over the past decade as democratized manufacturing became a reality. There are even HN discussions tying all of these technologies together, by comparing attitudes towards the availability of Tor vs guns (e.g., [1]).
And there are innumerable related moral qualms to be had in the future; will the illegal drugs or weapons produced using matter replicators be AI-designed?
Overall, I think all of these issues revolve around the question of what it means to limit freedoms that we've only just invented, as technological advances enable things never before considered possible in legislation. (And as the parent comment implies, here's where the use of science fiction in considering the implications of the impossible comes in).
[0] https://en.wikipedia.org/wiki/Four_Horsemen_of_the_Infocalyp...
[1] https://news.ycombinator.com/item?id=8816013
Deleted Comment
That doesn't mean that it's all bad, and that there's no recreational use for it. We have limits on the availability of various other artificial stimulants. We should continue to have limits on the availability of porn. Where to draw that line is a real debate.
[1] https://en.wikipedia.org/wiki/Wirehead_(science_fiction)
Deleted Comment
This author's books are great at putting these sort of moral ideas to test in a sci-fi context. This specific tome portraits virtual wars and virtual "hells". The hope is of being more civilized than by waging real war or torturing real living entities. However some protagonists argue that virtual life is indistinguishable from real life, and so sacrificing virtual entities to save "real" ones is a fallacy.
Or some such, it's been a while.
If people are exposed to stimuli, they will pursue increasingly stimulating versions of it. I.e., if they see artificial CP, they will often begin to become desensitized (habituated) and pursue real CP or even live children thereafter.
Conversely, if people are not exposed to certain stimuli, they will never be able to conceptualize them, and thus will be unable to think about them.
Obviously you cannot eliminate all CP but minimizing the overall levels of exposure / ease of access to these kinds of things is way more appropriate than maximizing it.
If you enjoy car chases in movies, does that mean you're going to require more and more intense chase scenes, and then consume real-life crash footage, and ultimately progress to doing your own daredevil driving stunts in real life?
No, because at some point it's "enough."
Same with... literally anything we enjoy. Did you enjoy your lunch? Did you compulsively feel the need to work up to crazier and crazier lunches?
What about sex? Have you had sex? Do you feel the need to seek out crazier and crazier versions of it?
I have accumulated tens of thousands of headshots in video games but have yet to ever shoot a single real person in the face. More importantly, I have never had the urge to seek out same.
I am not sure that your initial premise has any truth to it.
I'd actually argue the reverse, I think you see a lot more effort towards acquiring things that are illegal than you would otherwise.
> Our content policy does not allow users to generate violent, adult, or political content, among other categories. We won’t generate images if our filters identify text prompts and image uploads that may violate our policies. We also have automated and human monitoring systems to guard against misuse.
I can get why the people who worked hard on it and spent money building it don't want to be associated with porn.
Why? Is there something inherently wrong with porn? Is it not noble to supply a base human need, based on some arbitrary cultural artifact that you possess?
Are you asserting that nobody has humanitarian concerns? If so, that's quite a statement; what basis is there? I've seen so many humanitarian acts, big and small, that I can't begin to count. I've seen them today. I hear people express humanitarian beliefs and feelings all the time. I do them and have them myself. Maybe I misunderstand.
I'm not saying that AI will pass all Turing tests. But as far as having a virtual girlfriend/prostitute.
I'm not picking on the commenter - by itself it's not a big deal - but look at the assumptions behind that comment, which I almost didn't notice on HN.
* I recommend reading the Risks and Limitations section that came with it because it's very through: https://github.com/openai/dalle-2-preview/blob/main/system-c...
* Unlike GPT-3, my read of this announcement is that OpenAI does not intend to commercialize it, and that access to the waitlist is indeed more for testing its limits (and as noted, commercializing it would make it much more likely lead to interesting legal precedent). Per the docs, access is very explicitly limited: (https://github.com/openai/dalle-2-preview/blob/main/system-c... )
* A few months ago, OpenAI released GLIDE ( https://github.com/openai/glide-text2im ) which uses a similar approach to AI image generation, but suspiciously never received a fun blog post like this one. The reason for that in retrospect may be "because we made it obsolete."
* The images in the announcement are still cherry-picked, which is therefore a good reason why they tested DALL-E 1 vs. DALL-E 2 presumably on non-cherrypicked images.
* Cherry-picking is relevant because AI image generation is still slow unless you do real shenanigans that likely compromise image quality, although OpenAI has likely a better infra to handle large models as they have demonstrated with GPT-3.
* It appears DALL-E 2 has a fun endpoint that links back to the site for examples with attribution: https://labs.openai.com/s/Zq9SB6vyUid9FGcoJ8slucTu
Maybe give it another five years, a few more $billion and a few more petabytes/flops and it will be good. Then finally everyone can generate art for their own Magic: the Gathering cards.
(That's the end goal, right?)
My dataset is a start, but it may benefit from focused training, the way Facebook's new Make-A-Scene https://arxiv.org/abs/2203.13131#facebook (not DALL-E 2 quality but not far from it) has focused losses on faces.
They're a very complex anatomical form, many small tendons and muscles. Many artists struggle to depict hands. They're not made out of a few straight lines like a torso, there's lots of skew going on. They're probably the hardest structure of the human body to 'learn' for a ML system.
> an astronaut riding a horse, and the astronaut has five fingers on each hand
An example off the top of my head: this could be used as advertising or recruitment for controversial organizations or causes. Would it be wrong for the USA to use this for military recruitment? Israel? Ukraine? Russia?
Another example: this could be used to glorify and reinforce actions which our society does not consider to immoral but other societies - or our own future society - will. It wasn't long ago that the US and Europe did a full 180 on their treatment of homosexuality. Will we eventually change our minds about eating meat, driving cars, etc?
Have they gone too far in a desperate bid to prevent the AI from being capable of harm? Have they not gone far enough? I don't know. If I was that worried about something being misused, I don't think I could ever bring myself to work on it in the first place. But I suppose the onward march of technology is inevitable.
GLID-3: https://colab.research.google.com/drive/1x4p2PokZ3XznBn35Q5B...
and a new Latent Diffusion notebook: https://colab.research.google.com/github/multimodalart/laten...
have both appeared recently and are getting remarkably close to the original Dall-E (maybe better as I can't test the real thing...)
So - this was pretty good timing if OpenAI want to appear to be ahead of the pack. Of course I'd always pick a model I can actually use over a better one I'm not allowed to...
glid-3 is a relatively small model trained by a single guy on his workstation (aka me) so it's not going to be as good. It's also not fully baked yet so ymmv, although it really depends on the prompt. The new latent diffusion model is really amazing though and is much closer to DALLE-2 for 256px images.
I think the open source community will rapidly catch up with Openai in the coming months. The data, code and compute are all there to train a model of similar size and quality.
What kind of prompts is GLID-3 especially good for? I remember getting lucky when I was playing around a few times but I didn't do it systematically.
Do you happen to know how much GPU RAM I need to run glid-3 and/or the latent diffusion model, if I don't want to run on colab?
OpenAI has a low resolution checkpoint for similar functionality as this - called GLIDE - and the output is super boring compared to community driven efforts, in large part because of similar dataset restrictions as this likely has been subjected to.
I don't see a run button?
On.. maybe "Runtime -> Run All" from the menu ...
Shows me a spinning circle around "Download model" ...
26% ...
Fascinating, that Google offers you a computer in the cloud for free ..
Now it is running the model. Wow, I'm curious ..
Ha, it worked!
Nothing compared to the images in the Dall-E 2 article but still impressive.
However, the free GPU is now a K80 which is obsolete and barely sufficient for running these types of models.
It's hard to compare because we don't know how much cherry picking is going on with published Dall-E results (either v1 or v2)
My gut feeling is that it's in the same ballpark as Dall-E 1
Pilots are not there to fly the aircraft, the autopilot already does that. They are there to command the aircraft, in a pair in case one is incapacitated, making the best decisions for the people on board, and to troubleshoot issues when the worst happens.
No AI or remote pilot is going to help when say... the aircraft loses all power. Or the airport has been taken over in a coup attempt and the pilot has to decide whether to escape or stay https://m.youtube.com/watch?v=NcztK6VWadQ
You can bet on major flights having two commercial pilots right up until the day we all get turned into paperclips.
For things like aircraft pilots, it's both realtime-- which means 'reviewer' per output-- you haven't taken a highly trained pilot out of the loop, even if you relegated them to supervising the computer-- and life critical so merely "so/so" isn't good enough.
In practice in illustration (as in all arts) there are a variety of markets where different levels of talent, originality, reputation and creative engagement with the brief are more relevant. For editorial illustration, it's certainly not a case of 'find me someone who can draw X', and probably hasn't been since printing presses got good enough to print photographs.
Unlike artwork, precision and correctness is absolutely critical in coding.
Bytecode -> Assembly -> C -> higher level languages -> AI-assisted higher-level languages
But the reverse is not true, they won't be able to properly vet a piece of code generated by an AI since that will require technical expertise. (You could argue if the piece of code produced the requisite set of output that they would have some marginal level of confidence but they would never really know for sure without being able to understand the actual code)
For the first category, Dall-E 2 and Codex are promising but not there yet. It's not clear how long it'll take them to reach the point where you no longer need people. I'm guessing 2-4 years but the last bits can be the hardest.
As for the second category, we are not there yet. Self-driving cars/planes, and lots of other automation will be here and mature way before an AI can read and communicate through emails, understand project scope and then execute. Also lots of harmonization will have to take place in the information we exchange: emails, docs, chats, code, etc... That is, unless the browser is able to open a navigator and type an address.
It's important to note that we still need professionals to guarantee the quality of the output from AIs, including this one. As noted in their issue tracker, DALL-E has very specific limitations, but these can be easily solved by employing dedicated professionals, who are trained to tame the AI and properly finish the raw output.
So, if I were running OpenAI, I'll clearly be experimenting with how their AIs and human interact, and build a training program around it for producing practical outputs. (Actually, I work in consumer robotics, and human adoption has been the biggest hurdle here. Thus, my claim here.)
--
In case of fine art, thou, I don't think they'll not get hit by this AI advancement. The biggest problem is that you simply can't get the exact image you want wit this AI. Even humans cannot transfer visual information in verbal form without a significant loss of details, thus a loss of quality. It's the same with AI, but, worse, because AI rely on the bias in a specific set of training data, and it never truly understands the human context in it (in the current level of technology).
Deleted Comment
Additionally, the rise of no-code development is just extending the functionality of designers. I didn't take design seriously (as a career choice) growing up because I didn't see a future in it, now it pays my bills and the demand for my services just grows by the day.
Similar argument to make with chess AI: it didn't make chess players obsolete, it made them stronger than ever.
Are all designers becoming more valuable or is a subset of really good ones going to reap the value increase and capture more of the previously available value?
[0] Which is why releasing your code is so beneficial.
Using something like this could really help automate or at least kickstart the more mundane parts of content creation. (At least when you are using high resolution, true color imagery.)
There are some 3D image generation techniques, but they aren't based on polygonal modelings, so 3D artists are safe for now
Or what about even generating images you could then photogrammetry into models?
I can't help but feel a lot of the safeguarding is more about preventing bad PR than anything. I wish I could have a version with the training wheels taken off. And there's enough other models out there without restriction that the stories about "misuse of AI" will still circulate.
(side note - I've been on HN for years and I still can't figure out how to format text as a quote.)
I think your analogy is poor, because this is a tool for makers. The engineers aren't the makers.
I think a more apt analogy is if John Deere made a universal harvester that you could use for any crop, but they decided they didn't like soybeans so you are forbidden to use it for that. In that case, yes I would complain, and I would expect everyone else to, as well.
It's their service, their call.
I have some hobby projects, almost nobody uses them, but you bet I'll shut stuff down if I felt something bad was happening, being used to harass someone, etc. NOT "because bad PR" but because I genuinely don't want to be a part of that.
If you want some images / art made for you don't expect someone will make them for you. Get your own art supplies and get to work.
I feel like we, as a species, will struggle for a while with how to treat adults like adults online. As happy as I am to advocate for safe spaces on the internet, perhaps we need to start having a serious discussion about how we can do so without resorting to putting safety mats everywhere and calling it a job well done.
Hecklers get a veto?
Deleted Comment
It makes me wonder what they're planning to do with this? If they're deliberately restricting the training data, it means their goal isn't to make the best AI they possibly can. They probably have some commercial applications in mind where violent/hateful/adult content wouldn't be beneficial. Children's books? Stock photos? Mainstream entertainment is definitely out. I could see a tool like this being useful during pre-production of films and games, but an AI that can't generate violent/adult content wouldn't be all that useful in those industries.
I don't think there is a way comparable to markdown, since the formatting options are limited: https://news.ycombinator.com/formatdoc
So your options are literal quotes, "code" formatting like you've done, italics like I've done, or the '>' convention, but that doesn't actually apply formatting. Would be nice if it were added.
Personally, I prefer to combine the '>' convention with italics. Still, I'd agree that proper quote formatting would be a welcome improvement.
https://github.com/etcet/HNES
This is exactly the sort of thing that gets a company mired in legal issues, vilified in the media, and shut down. I can not blame them for avoiding that potential minefield.
(Hmm, I guess this comparison doesn't actually work...)
But at least we can get another billion of meme-d comics with apes wearing sunglasses, so that's good news right?
It's just soul-crushing that all the modern, brilliant engineering is driven by abysmal, not even high-school art-class grade aesthetics and crowd-pleasing ethics that are built around the idea of not disturbing some 1000 very vocal twitter users.
Death of culture really.
Companies like OpenAI have a responsibility to society. Imagine the prompt “A photorealistic Joe Biden killing a priest”. If you asked an artist to do the same they might say no. Adding guiderails to a machine that can’t make ethical decisions is a good thing.
Personally, I fear more what corporations or some governments can do with such models than what a random person can do generating Biden images. And without restriction, at least academics could better study these models (including their risks) and we could be better prepared to deal with them.
Society didn't collapse after photoshop. "Responsibility to society" is such a catch-all excuse.
Their document about all the measures they took to prevent unethical use is also a document about how to use a re-implementation of their system unethically. They literally hired a "red team" of smart people to come up with the most dangerous ideas for misusing their system (or a re-implementation of it), and featured these bad ideas prominently in a very accessibly written document on their website. So many fascinating terrible ideas in there! They make a very compelling case that the technology they are developing has way more potential for societal harm than good. They had me sold at "Prompt: Park bench with happy people. + Context: Sharing as part of a disinformation campaign to contradict reports of a military operation in the park."
But amusingly, exactly that did happen in one of their GPT experiments! https://openai.com/blog/fine-tuning-gpt-2/
Deleted Comment
That's no hot take. It's literally the reason.