AIs gunning for our precious freelancers

[zompist]

I was thinking about the "but AIs will improve till they can do anything" argument. I think this is farther off than people think (though less so in terms of art than writing), and that's due to how LLMs work.

It's extremely impressive what they can do, but they really are just very fancy prediction machines, which work because they have an unimaginably large corpus. A Markov generator can similarly produce way more coherent text than it "should" based on how simple it is. But LLMs also have easily demonstratable errors: they really can't reason, they have limited learning abilities, they hallucinate "facts", they are terrible at mathematics, they are easily misled, they make logic errors that reveal that they do not actually have real-world knowledge.

The thing is, more of the same process will not necessarily remove these problems. They can be fed a bigger corpus, though "the entire Internet" is already almost as big as we can get. But a bigger corpus will not teach them not to hallucinate, will not teach them things that aren't in it, will not teach them to reason. It's even possible that they'll actually decline in quality. A Markov generator actually gets worse as its token count increases, because then all it can do its regurgitate its inputs exactly. A neural net isn't quite the same but it is prone to over-learning, or too intently reproducing its training data.

Again, AI for the last 70 years has constantly claimed to be much farther along than it was. Whatever it's just done (stacking blocks, playing chess, holding a simple conversation) has been touted as really really close to intelligence and any further obstacles pooh-poohed as moving the goalposts. And the remainder of the problem keeps biting AI in the ass. AI researchers-- and everyone else, from sf writers to journalists-- turn out to be really bad at defining what "real intelligence" is and how much work is left to do.

This isn't to deny the real advance LLMs have made. I think, as a computer science advance, it's comparable to the invention of high-level languages, or the GUI. And again, there are a lot of applications for LLMs and their equivalents in other areas: something that produces "more of the same" is a pretty useful tool. But I suspect adding more and more training data will not suddenly create "actual AI". I think some kind of lateral move will be necessary.

[malloc]

But why the interest in achieving true AI in the first place? We already have intelligent beings, namely ourselves, and we quite enjoy existing and running civilization. Yet for some reason the tech industry has spent the past 70 years trying to create something with our intelligence and thus ability to drive civilization. Many talk about automation eliminating drudgery, but you hardly need AI to eliminate boring work. Machines without intelligence like printing presses and combine harvesters can handle dull repetitive tasks quite well. By contrast, AI targets jobs that take advantage of distinctly human strengths like intelligence and creativity. It often feels like many techies genuinely dislike humanity and wish to see it forced off the stage of history or else reduced to pets of something greater.

[Ares Land]

But I suspect adding more and more training data will not suddenly create "actual AI". I think some kind of lateral move will be necessary.

Yep. And that sort of move is hard to predict.

I think there are several aspects about the AI worries.

1) From what I can see, it looks like artists are worried about being replaced, but rather about what they see (correctly, I think) as plagiarism.

2) As for the purely artistic, existential worries... I don't worry, myself. I'm not interested in the kind of derivative works current AI can produce. I don't see how this is going to change without getting quasi-sentient AI or something like that... and I don't believe this is going to happen anytime soon.

3) As for economics, yeah, there is an issue. AI can automate plenty of jobs (illustration, copywriting and so on.)
I don't think it's going to be as simple as AI destroying jobs. Is there even really some kind of correlation between automation and unemployment? We are, currently, automatically basically everything we can and we have a labor shortage. I think the effects of automation are more subtle.
Automation translate as productivity gains, and most people get no advantage from these gains. I think that's the most likely scenario: higher productivity from AI and other technological advantages, standards of living stagnating or slowly eroding.
Note that this isn't specific to AI, this has been going on for 20 to 40 years, depending on where you live and what you do for a living.
There are ways to distribute productivity gains more evenly that have worked before though. (Cheap education; shorter working hours; there are countless examples of successful policies). The problem is again, the lack of political motivation (voters aren't interested in these policies, so nothing's going to happen for the time being.)

[Ares Land]

It often feels like many techies genuinely dislike humanity and wish to see it forced off the stage of history or else reduced to pets of something greater.

On that I agree with you! You got some people like Yudkowsky (probably not even the worst example out there) with that odd sort of fascination for superintelligent AI, and they creep me out too.

[WeepingElf]

As to the question whether AIs are "cheap" or "expensive", this depends on which part of the process you look at. Training an AI is expensive - it involves an awful lot computing power and data input, and that costs money. But once it has been trained, it will work for you at little extra cost. The economics of AI are thus surprisingly similar to the economics of books or music recordings - it is very costly to write the manuscript or make a recording, but it is dirt cheap to run off a few more copies once the creative work is "in the box".

[Linguoboy]

As to the question whether AIs are "cheap" or "expensive", this depends on which part of the process you look at. Training an AI is expensive - it involves an awful lot computing power and data input, and that costs money. But once it has been trained, it will work for you at little extra cost.

This isn't a one-time cost either. If you don't continue to train a LLM on more recent corpora, the results will become progressively outdated. Already users of ChatGPT are complaining about obsolete information making its way into texts.

[zompist]

As to the question whether AIs are "cheap" or "expensive", this depends on which part of the process you look at. Training an AI is expensive - it involves an awful lot computing power and data input, and that costs money. But once it has been trained, it will work for you at little extra cost. The economics of AI are thus surprisingly similar to the economics of books or music recordings - it is very costly to write the manuscript or make a recording, but it is dirt cheap to run off a few more copies once the creative work is "in the box".

No, not really. LLMs cost money to operate. E.g. see this article on the big search engines. If Google used the equivalent of ChatGPT on every query, it would cost the company $36 billion— halving its profit— and require half a million servers.

Or as this article says, "Training a large language model like that used by ChatGPT is expensive — likely in the tens of millions of dollars — but running it is the true expense." Running (not developing) ChatGPT costs millions of dollars a day.

Or there's carbon footprints. Ironically, it's not training that's the problem; it's running them. This article estimates that just 100 Bing-style LLMs in the world would take up 40% of the carbon budget for the entire computing industry, and make it impossible to keep global warming under control.

(As the article notes, there won't be 100 search engines. But the demand for LLMs is enormous. Again, the problem isn't one fanfic writer generating one illo with Midjourney. It's companies all around the world generating their huge piles of text with LLMs, or analyzing their piles of data on consumers, or labs doing cutting-edge graphics work.)

[WeepingElf]

So I was wrong about the economics; thank you for enlightening me. I am not an expert on these matters, and should better shut up

[KathTheDragon]

Do what I do and reframe anything you don't know for sure as a question.

[WeepingElf]

Yes, that's good advice. Thank you.

[Dune]

So ... I hire writers, both precious freelancers and staff writers. I'm one of the editors for a modest-sized network of sites—we publish attempts at Actual Serious Journalism, as well as SEO bullshit to keep the lights on. I thought I'd create an account and chip in with my perspective (I had an account on the old board but pretty much just lurked, so I don't think anyone will remember me).

When ChatGPT appeared last year, I was very concerned, both for the writers I work with and for myself. It was obviously an immense leap forward, and it was easy to picture all of our jobs vanishing in a year or two once OpenAI worked out the kinks in their model.

Now ... I'm a lot less concerned. Like a lot of publications, we've experimented with using AI to write articles (no, I'm not proud of it, but we all work for somebody). The results have been uniformly poor. I've used both ChatGPT-3 and GPT-4 and neither has produced anything I would ever sign off on publishing, at least without reworking it so extensively that we might as well have just written the damn thing from scratch. We've also had a staggering number of would-be freelancers submit writing samples to us that they created with GPT, and it's almost always INSTANTLY apparent that they've done so (it's an automatic disqualifier, of course). ChatGPT has a very distinctive writing style, and getting it to produce something in a different voice actually requires a fair amount of effort and skill.

The thing that's struck me lately is that I don't think AI content has gotten any better in the past few months. Actually, to the contrary, I think it's getting worse—flatter and more generic, more recognizable. Six months ago I think you might have been able to fool me by submitting a GPT'd sample, but today, I don't think you could. I suspect this is partly because OpenAI has added more limitations to keep the bot from producing objectionable content, but I think it's also because they've fed it more training data, and as zompist said, that has a homogenizing effect.

If I had to describe most of the AI content that's crossed my desk, I'd say it reads like something produced by a reasonably talented sophomore writer without a whole lot of life experience. You know how, if you Google something like "what to do on a trip to Madrid," you'll mainly get bullshit SEO articles that seem like they were incestuously copied from each other, and you can't tell if any of the writers have ever set foot in Madrid? And then you get fed up and you add "Reddit" to your search and you read a thread full of comments that maybe aren't written as well on a technical level, but you can somehow tell that the anecdotes are actually true, and the posts just feel more genuine and "alive"? Yeah. You've probably noticed that's what's missing from AI content, that subtle spark that comes from lived experience and original thinking. It's not just superstition, there is actually a difference, and if you break down the writing and analyze it on a sentence-by-sentence, word-by-word level, you can identify exactly what the AI is doing wrong. And I have done this in meetings with executives, who naturally would love to stop paying our freelancers at all if they could. To their credit, most of them have gotten it ... eventually.

This holds for many different types of writing. If I want a journalistic piece, an LLM cannot get out there and interview people. If I want a review of a product, it cannot use the product. If I want an opinion piece, well, a LLM has no opinions of its own. It can produce a facsimile of all three, but it just won't be as good as the real thing. Publishers and editors are noticing this, and realizing that people can tell the difference. Not all of them, but enough.

It's possible that there's a way for AI researchers to overcome this. But I just haven't seen any evidence of that yet.

I don't mean to downplay things—some writing can be produced by AI with no loss of quality. Some writers have already lost their jobs to AI. The job market for writing will contract as people compete for the jobs that require a human touch. But now that the dust has settled a little, I'm pretty confident that the job market is not going away entirely.

Caveat: Most of the AI-gen writing I've seen has come from ChatGPT, since it's the most famous and accessible LLM. There are others, of course, and it's possible one of them is producing stuff that would really shock me. But I doubt it.

Caveat 2: All of this applies to written content. I feel bad for illustrators, because to my eye, high-quality AI images are essentially indistinguishable from human-produced ones. But that's my impression as a layperson—maybe that's not actually true and people who professionally evaluate visual art can clock the difference instantly.

[linguistcat]

Caveat 2: All of this applies to written content. I feel bad for illustrators, because to my eye, high-quality AI images are essentially indistinguishable from human-produced ones. But that's my impression as a layperson—maybe that's not actually true and people who professionally evaluate visual art can clock the difference instantly.

I can help with this one: AI isn't great at stylized art because it doesn't know how a human might actually simplify things in order to stylize them. It can do highly "realistic" rendered/detailed pieces, but it puts too much details where no human would while having objects start and stop weirdly. Like hair coming from nowhere in a way that wouldn't make sense for a human artist, or decorative "patterns" that aren't really patterns. They suffer from the same thing other AIs do, no real world experience.

[WeepingElf]

This issue is also currently being discussed in a German conworlders' forum where the following was posted: apparently, the Authors' Guild has filed a lawsuit against OpenAI for copyright violation.

[malloc]

Funny how so many artists and writers are filing lawsuits against it. Almost as if it poses some threat to their livelihoods... But more seriously, it's good that people are fighting back against this destructive juggernaut. While it seems doubtful that humble artists can prevail against an ultra-wealthy industry at war with the humanities (if not Homo sapiens itself), we must at least try.

[zompist]

Thanks for an actual editor's perspective, Dune!

If I had to describe most of the AI content that's crossed my desk, I'd say it reads like something produced by a reasonably talented sophomore writer without a whole lot of life experience.

That's about my impression too. For AI, that's an impressive achievement! But ironically, it'd be a greater achievement if LLMs were able to simulate a smart sixth grader. My impression is that kids that age are far more creative and original— they don't have much life experience yet, but they are quirky and write things that adults might not think of. By high school they are basically producing pastiches, which is pretty much what LLMs do.

[linguistcat]

Funny how so many artists and writers are filing lawsuits against it. Almost as if it poses some threat to their livelihoods...

I mean, more threat than fanfic writers selling their derivative* fiction without the permission of the original author. And people get slapped with C&D letters (or these days, emails) from people's lawyers even when they DON'T make money off of fanfiction sometimes, if the author is particularly protective and think it doesn't count as fair use of their work. So why wouldn't authors who have the means to do so sue companies that used their works as training fodder without permission, let alone an offer of compensation? They're way more likely to get something out of it than going after Susie Such-and-Such who wrote a commissioned fanfic for an extra $20.

*I mean this in the literal sense that fanfiction is derived from a source fiction, and not in any snobby, anti-fanfiction manner. I'm very pro-fanfic myself, but I understand that not making money off it goes a long ways toward arguing it's fair use if someone tries to take you to court.

[malloc]

How many people here would want to compete against something that can perform their job without error or exhaustion at far lower cost? It may seem easy to dismiss this issue when people from other fields, particularly those you disdain, are struggling to maintain their jobs. But you must also consider that the people in charge will not stop with writers or artists. Once they've reaped the massive savings of laying off those professions, they will attempt to repeat their success with any profession they can find. You may think your job is safe from automation, but remember that writers and artists once thought the same thing.

What even is the endgame of all this automation? People used to argue that once we had enough automation, we could put aside boring dangerous work and focus on fun jobs like art and literature. Yet that clearly is not the plan anymore, nor even a possibility given current and future AI technology. What happens when technology advances to the point that most humans are superfluous to running the economy and why doesn't anyone care?

[linguistcat]

How many people here would want to compete against something that can perform their job without error or exhaustion at far lower cost?

So first off, none of this is performing the jobs of writers and artists without error. In fact one of the biggest concerns is that LLMs introduce A LOT of error into what they produce. And it isn't free either. The companies that offer free use of LLMs and generative art either make people pay for any advanced features or limit the amount they can be used, because they are banking on people buying (or rather renting since it's usually a subscription) and making them money later.

As a side note, I still think that you had to have been looking for AI art on DA to get overwhelming amounts of it, since I have never turned on the setting to remove AI art, and I do not get any suggested to me on a normal day. Zero. Because I religiously told DA to remove it from my suggestions when this all started and never went looking for any.

What even is the endgame of all this automation? ...What happens when technology advances to the point that most humans are superfluous to running the economy and why doesn't anyone care?

You know the economy doesn't exist without humans, right? Or sapient beings at least, but LLMs and other supposed AI are far from being sapient. Elephants, octopuses and ravens are closer to joining the world economy than neural networks are. Acting like the economy can cut out most humans without being replaced is just not reasonable.

I would say that the sooner all the billionaires ship themselves to Mars and die of bad planning and safety standards the better.

[malloc]

So first off, none of this is performing the jobs of writers and artists without error. In fact one of the biggest concerns is that LLMs introduce A LOT of error into what they produce. And it isn't free either. The companies that offer free use of LLMs and generative art either make people pay for any advanced features or limit the amount they can be used, because they are banking on people buying (or rather renting since it's usually a subscription) and making them money later.

But eventually the technology will improve to the point that it never or rarely faults, just as airplanes went from barely flight worthy to flying passengers across the ocean. Likewise the cost will drop just every other piece of technology until it falls below the wages need to sustain a human. What happens when generative AI costs less than even the cheapest human writer?

You know the economy doesn't exist without humans, right? Or sapient beings at least, but LLMs and other supposed AI are far from being sapient. Elephants, octopuses and ravens are closer to joining the world economy than neural networks are. Acting like the economy can cut out most humans without being replaced is just not reasonable.

From the standpoint of corporations, the economy is simply the line going up. If they can find a way to keep increasing profits without having to pay a single human, they will. Homo sapiens is merely one resource among many available to them in their quest for endless growth. And while current LLMs are not sentient, future waves of AI will undoubtedly reach that goal. For reasons I cannot quite understand, the tech industry is obsessed with brining about the Singularity and has trillions at its disposal to spend on research.

[Raphael]

For reasons I cannot quite understand, the tech industry is obsessed with brining about the Singularity and has trillions at its disposal to spend on research.

While we might get actual AI some day, the Singularity is extremely unlikely to ever happen; the whole idea was always based on a number of faulty assumption, such as that computing speed would keep getting faster for ever.

As for reasons, well, understanding the motivations of people who find things fascinating that you yourself don't care about is often difficult and might even be impossible. It's simply a fact of life that many people have been fascinated by the idea of artificial intelligent beings since long before computers in the modern sense existed.

I dunno about your "trillions" number. Has anyone done the math on that?