Ask HN: What creative field will be safe from AI?
I am a visual artist and Dall-E 2 has made me realize that my skills will soon be almost worthless.
My role, if it continues to exist, will consist of prompting the software and selecting among its outputs. This is unsatisfying work that I cannot take pride in. My artistic skills, my craft, has been automated.
Quoting Wikipedia: On 19 November 2019, Go master Lee Sedol announced his retirement from professional play, stating that he could never be the top overall player of Go due to the increasing dominance of AI. Lee referred to them as being "an entity that cannot be defeated".
I am young enough to switch tracks and focus on a new creative field.
What creative skills will be safe from AI?
54 comments
[ 2.6 ms ] story [ 120 ms ] threadAt least until a true AI/HCI is created.
Most "AI" things are pretty superficial pattern recognition that are just better at unbiased processing of large "shallow" datasets.
However AI can definitely play a role in the process, especially around editing. Professional writers have the advantage of professional editors and expert beta readers. I would love a fiction writing companion similar to GitHub copilot that could suggest improvements and refine prose.
I think visual artists could benefit from similar tooling. Imagine a comic art tool that can take a pencil drawing and ink it in, or even take a rough sketch and refine the details.
First off, GPT-3 is terrible at plots and all long form writing. That doesn't mean that AI/computers are going to be bad at it. There's an algorithm. Many screenwriters today write based off Aristotle's Poetics or some variation of Hero's Journey. The West Wing is brilliant and deep, but it's also written around an algorithm.
We'll probably see some hybrid engines around this. Just like computers have GPUs and CPUs and RAM and HDDs. There will likely be an engine for plot, and another GPT-3 style one for "rendering" that plot.
A downside is that it does have a certain "voice". I would say its natural voice is similar to 16 year olds writing fan fiction. If you don't guide it well (and most don't), the outputs will all be similar, and they will sound tacky.
And so guiding will matter a lot. I see AI as possibly taking the role of scriptwriter and actor, but a human needs to take the role of a director. Most interpretations today are quite lazy and expect AI to do all the work. I think a form of creative coordination is needed. You still have to put your AI in a situation where they can be inspired. You can't just tell them to write out a full scene of a man confront his adulterous wife. You give them the saucy plotwriting from Patterson, snippets of over the top dialogue from Conan the Cimmerian, and you get some form of artwork.
But if you're not a good writer in the first place, you can't actually train AI to write well. Most of the parties interested in AI writing are screenwriters. It's sort of like accountants and engineers jumping after calculators. The machine doesn't obsolete them; it allows them to do more of the thing they wanted to do, and less of the tedious stuff they didn't want to.
Some of the cherry-picked examples were impressive, but the others, not so much. The output lacks the deliberation & thought that an artist’s work would exhibit given the same prompt.
Not to mention that most of the results are not not completely cohesive (e.g. a rabbit missing an ear) so not sure about their usability.
As people start to generate crap at tremendous volume with AI, more value and emphasis will be placed on creative works made by humans. I very well could be wrong, but either way it's not happening anytime soon.
---
Games like go and chess are different, of course computers are better for calculating and optimizing moves and branching paths - no creativity is needed. This portion is completely irrelevant to the discussion of computers trying to generate art.
None of ‘em. None of them will be safe. Humans ascribe meaning to works where there is none, and intent by the artist or an AI can be wholly immaterial to what people appreciate about it. Great works may come from AI but the greatness of a work is virtue of the work itself.
That said, just because a computer can do it doesn’t mean that a human can’t, or that people will prefer the AI work over the human’s. I think AI threatens jobs in a lot of places, but not necessarily in creative spaces.
Isn't this the majority of the industry? Sure, maybe there are a few artists that can do profound work with context, meaning, etc. But when you're doing book covers for airport novels, you don't need any of that. Or when doing a website with a few images? You need a result that's good enough at the cheapest price. AI generated content is great for that.
You can apply an AI-powered “Starry Night” filter to any picture, but do you really think Van Gogh would draw something like that?
Erm. Not really. I used to pay artists very frequently for art to be used in web pages as illustration. Not many people do this. But I did.
But when I saw the quality of pictures genereated by CLIP guided diffusion, I knew they were production-ready.
I used some of those instead.
I briefly worked with a company that served ecommerce sites with automating of various tasks- one of them was generating description of products.
In a meeting I asked a client how did they do it before. They said that they hired struggling writers. We replaced them for a fraction of cost.
We were not only doing bland product description, we were also generating beautiful prose around products.
I am still on the waitlist of Dall-E 2, but some friends have access and I played with it for hours, and it is not a trivial technology. It is production ready that can replace much of stock photos with generated content. So, not only artists, but some photographers are at risk, too.
It's not like AI will replace all creative workers, but it will certainly replace some of them, starting today. Hell, it has already started.
> and don't forget you're not seeing ALL of the pics that have been generated, but some of them
This is not some corporate trick. This is standard practice. When I worked on an image generation project, I learned that it was standard to generate 250 or so images, and use CLIP to rank them, i.e. let CLIP score the images based on input prompt. Then the top-k images with highest scores, i.e. most similarity to prompt according to CLIP will be displayed.
> for calculating and optimizing moves and branching paths - no creativity is needed.
In the AlphaGo documentary, it was shown that one move was very original. It was unprecedented in Go history. Maybe it wasn’t creative, but it can mimic creativity.
If you want to stay on the visual creativity wicket, consider getting in to VR educational experiences (massive opportunity: far superior to traditional teaching methods in many cases) and computer games (huge money in the industry, if you can get it).
The field comes with many challenges though. Accessibility is not yet figured out. Reading text in 3D with current tech is painful. And the technology is far from wide spread among target population.
I'd appreciate being pointed to any community or project wanting to advance the field. Especially ones trying to democratize the education in an open-source fashion.
as intelligence requires self awareness
This is pretty much exactly what I see here. The human must still come up with the idea that an astronaut riding a horse is a good idea, and must selectively adjust the output until his/her vision is realized, the AI takes care of the majority of the individual brushstrokes/digital operations, and there's still likely some significant human-level work at the end. But yes, artists with skills but little vision, or stuck in jobs where the vision is not theirs, such as making advertisement illustrations, will likely be largely replaced.
The AI companies are re-encoding large corpora of human creative work as a kind of compressed, conceptually-indexed representation. This allows that work to be mixed and the patterns regurgitated in very flexible ways.
Everybody's output (our creative work) is assimilated by the AI, and becomes the AI's output.
For example, OpenAI's Codex is trained on about 54 million public GitHub repositories.
This allows Microsoft's CoPilot to regurgitate pieces of that code without attribution. Code from many sources is blended together and regurgitated for Microsoft's customers (without any acknowledgement of the source).
It is perhaps the greatest theft of intellectual property in the history of Man.
How did you come to this realization? We will always need artists. Art, painting was always a zero-sum game with very few on top making millions and millions making peanuts. Dall-E 2 will further this divide. Art will venture further in the zero-sum area.
But we will always need artists, and people will continue to create art no matter what. Some of those created art will garner attention and many of them will speak to many of us. We will pay. I regularly pay for art.
> My role, if it continues to exist, will consist of prompting the software and selecting among its outputs.
No. Non-artists can do it comfortably. I have used CLIP-guided diffusion generated art in many web pages, that I picked myself, that replaced some art jobs.
> My artistic skills, my craft, has been automated.
But not your identity or you soul. Art always had a personal component.
> Go master Lee Sedol announced his retirement from professional play
I don't know why he did it. Chess was solved before AI with IBM Deep Blue, and during AI with Alpha Zero. Yet Chess continues to be extremely famous. Chess is more popular than ever. People are going gaga over Magnus Carlsen, Vishy Anand, and matches in general. Chess and Go will be famous even after a century or five.
> What creative skills will be safe from AI?
People will like to draw even after Dall-E version N. Teachers will be needed. Art teachers' jobs are safe.
People will love to listen to a live concert no matter what. Musicians' jobs are safe.
I can go on. Art, of any kind, was a field of lack of symmetry. Barbers, plumbers, accountants, tailors, tellers make similar amount of money all over the world (considering the PPP). But artists don't. Art was always a field of huge disparity.
One barber can do one person's hair at one time. But a musician can stream to millions and play live to thousands. Art always scaled and was thus always a field of wild inequality.
What AI will do immediately is further the divide. It will wipe out the vocation of many at the bottom.
Not traditionally considered a "creative work", but teaching jobs will be safe for quite a long time.
Have you considered illustration jobs, design jobs? Generating one image is far from a finished designed, well-thought layout. What about game art or animation studio?
Maybe, the only creative skills that will matter in the future (from a future art historian perspective) will be the ones to program or use AI.
Bu I don’t think the game is over:
- if you find AI art tools too demeaning and they make you feel that you’re only a dumb prompt writer, become or convince a programmer to build a better tool that connects with your own creative mind
- Or, come up with a new form of art that rebels against the current technological things. Your predecessors were able to raise a middle finger to photography by inventing new abstract forms of art
My tentative guess is that creators whose individual personality and identity are an important part of their appeal will be protected from obsolescence for the time being. Some commercial artists, like many “serious” artists, have public identities and distinctive styles that are part of their appeal; they should be able to continue charging decent rates for their work. But a lot of commercial art is anonymous, without the creators even being credited by name. Content producers will be happy to replace their work with cheaper, faster, easily customizable AI-generated work when they can.
A person starting out now as an artist, musician, or writer, in order to succeed, probably needs to put as much effort into creating a distinctive and attractive public personality as into honing their skills and producing superior work. Some creative people will be able and willing to do that, but it will be difficult for a lot of others.
As Machine Learning Models are intellectual property we will likely see the unionification of artists as they work towards protecting the industry. Honestly the majority of these algorithms are going in the complete wrong direction and companies that use this will likely go bankrupt if google follows anything similar to amazon's business model.
Sure.
> that will never be fixed.
Less sure.
> Due to the inherit way that these GANs
Not a GAN. DALLE-2 uses a diffusion model for generation and a transformer/diffusion to convert CLIP text embeds to CLIP image embeds. DALLE-1 used a variational autoencoder, not a GAN. I mention this simply because both diffusion and VAE's seem to be more robust than GAN's (at the cost of slower inference times).
> It will probably take hundred of thousands of images at the least to create a decent looking image.
A little confused about this. DALLE-2 took hundreds of _millions_ of images to train, but it doesn't need even one to generate an image now. Having said that, finetuning requires small datasets like this and is perhaps what you meant?
> Each of which mind you is copyrighted.
I hear you on this - as it's a little devastating not to be even attributed for your work if you're a hard-working artist. However, training on datasets that aren't explicitly public domain or an open license is not a violation of copyright. Nor, at least presently, is releasing the weights of that model to the public. It would only be a violation to distribute the media itself - a problem which has been trivially circumvented by just releasing the URL's of the dataset for other researchers to download. In the case of smaller datasets, releases come with disclaimers that the data is not to be used for anything other than research. OpenAI of course has not even released the URL's for their web datasets.
> As Machine Learning Models are intellectual property
Hm, I think I agree that this could become an issue. You can see this in the licensing terms in StyleGAN2/3 already, for instance (typical nvidia, they basically own your outputs last time I read the thing).
At least in terms of what I've been seeing lately however, this won't be a very effective regulatory moat for AI companies. Research tends to be replicated rapidly by lots of brilliant open source developers. Sometimes, the paper even comes out long before the model is released giving plenty of time for a clean-room implementation. There are also of course great institutions like in Heidelberg doing bleeding edge, state of the art work and releasing _everything_. See e.g. latent-diffusion.
Machine learning requires lots of hardware - that's the best moat these companies can have. NVIDIA GPU's and Google TPU's are what's to be concerned about here.
With each concurrent generation the cost to produce the tools necessary with decrease exponentially as hardware also become more freely available. I don't see hardware as a limiting factor. Though the two monopolies have been known to work together to slow down technological progress. Honestly I think that our progress would likely have gone a lot faster without them. But for some strange reason they would rather follow moore's law. I don't really believe moore's law exists, there's limits to everything and the rules they mention are unsustainable and unrealistic. It makes me wonder where the real end of it is or if we've made the wrong decisions along the way solely for profit motive. How many times has funding been cut for research programs simply because nothing happened for a single quarter?
Like most Generative Artificial Networks, DALLE-2 is not really anything new. In the computer graphics field several research groups have been discussing similar techniques for decades. I've even seen a few demos by various colleges of their own implementations. The idea that DALLE is the future next step for several fields such as art is misleading at best and straight up delusional at worst. The main reason I state the issues within DALLE-2 not being fixed is mainly due to these factors. It's the wrong direction for this type of problem especially so for the billions of parameters that they use. The hyperfixation that google is attempting to do to this industry for it's own financial interests will pigeonhole several companies towards suboptimal solutions that will at best give them a somewhat workable solution at an expensive price. This of which will likely bankrupt them in the process as even with improvements to techniques will still cost them in the millions to produce even a few of these samples simply due to the labor cost involved.
While the art industry has been transformed by the technology industry through the gig economy intentionally for this purpose. Even at those rates you still end up with the cost of producing the models being outcompeted by the value provided. This is especially true as art improves overtime, the level of detail and fidelity along with varying art styles. Not to mention the fact it is an unsustainable model that prey on artists by systemically abusing them by turning them into slaves. But who am I say that it won't just collapse. After all companies such as walt disney, pixar, and dreamworks have been known for doing just this for decades. As other countries build "incentive" plans to systemically enslave their workers as well to outsource the art work towards studios abroad as well this likely won't change. Even the best animation studios such as Studio Ghibli only pay just above the poverty line. While it's never been my dream to draw or even write the idea that people of any kind being subject to abuse is something that I can't bear to see. All I can say is my deep hate for this never ending rat race. I always find it interesting when I read about these algorithms how the people who read, learn and create them seemingly have no sense of empathy for those that they seem to see as beneath them. Though I'm not claiming to say that you, yourself are to blame or are even contributing towards it. It's just something that bothers me.
Anyways, continuing on towards copyright. It's interesting to me how these datasets are generated is through the theft of artwork. I've already mentioned the labor cost involved in generating their own datasets and to mention it's unlikely anyone besides google will be able to afford to get away with this simply due to the magnitude of the lawsuit they would face. (It would likely be class action.) It always makes me wonder just how detrimental such as case would be and the ef...
Personally, I love hyperrealism when it comes to painting. Of course, a child could basically take a photograph now and be on par with a master hyperrealist compared to hundreds of years ago but the output in isolation is not what matters. The display of human skill is what is interesting about hyperrealism.
Most art people love is a display of human skill and/or the personality of the artist that can't be separated from the output.
Almost anyone can get a quality piano sampler, midi sequencer and then output classical piano virtuoso music at a tempo only Yuja Wang could currently play. You could even speed it up 10X if you wanted to complete inhuman tempos. No one cares about that output though. Exactly what makes this interesting is Yuja Wang's playing that seems inhuman.
This idea that AI is going to replace all human activity is transhumanist nonsense and a simpleton view of the world IMO.
https://passo.uno/posts/computer-aided-technical-writing-is-...
Unfortunately, this is subject to fraud, and can be manipulated for commercial purposes. I can easily see some “Milli Vanilli“ type of visual artists coming out of this.
Reality has given us the exact opposite situation. AI has proven to be basically a bounded RANDOM function that can instantly throw countless weird and unique ideas our way in seconds. Fighting against this is fighting against a great tool. The power of a concept artist designing a new landscape to just describe a few things, have AI output a few ideas, and then off you go, sampling interesting color palettes, taking forms or ideas from here or there. The power of a composer to have gone through 50 AI generated melodies to find a couple of hooks that they probably wouldn't've thought of, only to inspire a new variant, or to pull them in and build around them.
TL;DR don't fight it, use it as an idea generator to give you some Lego blocks for a particular creative Enterprise. Your role will shift more to the taste maker, where you'll be expected to have good artistic taste for curating and assembling and guiding AI outputs.
So creating interconnected or intersectional works, and weaving them together repeatedly and meaningfully over long periods. I'd say that sort of thing is quite far off, and possibly not even coming without another tech breakthrough.
It cant write songs, stories, long jokes, create brands, create videos, create businesses...
We might start to see commercialized products having some success at figuring out static things like style though in the next 5 to 10 years. Fashion and interior designers have "things are a bit harder in 3D" as a shield at the moment, but it might just be a matter of time for them... it will likely depend on how much manufacturability constraints can be trained.
Another thing with Art is that the current low to mid level art work could be outsourced to a much cheaper labour country. In terms of Web Design, Stock footage etc. They are a much bigger threat now than AI.
So, likely just because you can make AI works of art, it doesn't mean people will value them equally.
I personally wouldn't worry about it. Yes, as the saying goes in investing, "people can remain irrational longer than you can be solvent", but if it happens that "AI" starts to replace entire professions wholesale, then we'll have bigger problems. Do you really want to spend your life running from niche statistical machinations?
> Our hope is that DALL·E 2 will empower people to express themselves creatively. DALL·E 2 also helps us understand how advanced AI systems see and understand our world, which is critical to our mission of creating AI that benefits humanity.
This stuff is so far up its own ass, it's comical to me. I get the feeling that OpenAI has a similar problem to the critiques of the MIT Media Lab, in that they're both demo factories in order to generate more funding.
On the other hand, we still don't know how to teach AI to distinguish between good and bad art. So perhaps a lot more humans will become professional art critics.
They will no more replace artists than photoshop or illustrator did.