AIs gunning for our precious freelancers

[bradrn]

The basic premise I work on is that ‘simulating intelligence’ and ‘being intelligent’ are one and the same thing. Something which can simulate intelligence sufficiently well must itself be intelligent. Trying to argue otherwise leads to incoherencies like the Chinese Room.

Ehhh... I don't know about that. Or maybe we should be extra careful in determining definitions and acceptable values for 'simulate', 'intelligence', 'well' and 'sufficient'.

Yes, we should be very careful with this. Indeed the point I’m making is that ‘simulate’ may be a meaningless word when it comes to intelligence.

(Also, always keep in mind: ‘intelligence’ is what we’re trying to define in the first place!)

Additionally, my belief is that the brain is not like a computer at all.

Strongly agreed.

• Can intelligence be measured in levels? That assumes it is quantifiable and stable enough to be compared. Can there be a meaningful comparison of the 'level of intelligence' of any two things?
• Where are the cut-offs around the concept of 'human-like things'? Can this be a meaningful category?
• When does something transition from 'like x' to 'just like x'?
• Does it matter if people perceive a/differing degree(s) of likeness between y and x when one is considering the question 'is y the same as x', or in other words 'is it true that y=x?'?

Taking these in order:

1. This is a critical point and deserves more than I can type on a phone (which I am restricted to at this moment).

2,3. Ask zompist! I wouldn’t have phrased it this way.

4. Yes, it does matter. But I think that impressions should cluster around some definite mean. (I could be wrong.)

The ideological romance of the Turing test has cast a long shadow. Perhaps in evaluating its usefulness and risks we ought to try to discuss simply what AI can and can't do as a tool, rather than constantly interrogate its capacity to appear human?

Agreed, and I feel this is how I tried to approach it originally, by framing it in terms of prediction tasks.

[malloc]

Machines are part of human culture. They do not exist outside of human culture. What you are afraid of human behaviour being superseded by is human behaviour itself – merely, another type of human behaviour.

Artificial intelligence has a level of autonomy that other machines fundamentally lack. Printing presses and cameras, however disruptive their abilities, do not generate their own content but only replicate what we deliberately put into them. Generative AI by contrast can write entire novels and draw elaborate pictures without human guidance beyond the initial prompt. Whether you consider it capable of true thought or not, it clearly acts in effectively non-deterministic ways unlike previous technologies.

Yes; we made the machines on purpose to help us survive and thrive, and you are using one to communicate your distress, alongside many other, no doubt, important uses. You sound like the UK government trying to send the Windrush generation back.

A more apt comparison is indigenous people facing colonization. There is a staggering power imbalance between creative workers and the fantastically wealthy, not to mention far right, businessmen attempting to replace them with machines. Artificial intelligence is not an impoverished immigrant facing unfair bigotry but a scab designed to break the last shred of power that workers hold and render us superfluous to capitalism.

I shall chastise my oven, boiler and coffee maker at the earliest opportunity. Their disenfranchisement of the sadly imaginary servants whom I would be able to employ in your utopia has gone too long unnoticed; said servants would no doubt love working mind-numbingly dignified shifts in my kitchen rather than doing something else meaningful over which they have autonomy and creative freedom.

That makes no sense whatsoever. Art and literature are not drudgery that creatives resent having to make for survival. Those are precisely the kind of jobs that people wish they could have instead of toiling in factories and offices. You might have a point if we were talking about industrial robots kicking people out of factories but that doesn't really apply here.

[zompist]

Should people who draw furry OCs on commission lose their jobs?

How many of these people are there, and have they lost their jobs?

There are artists whose jobs I worry about, but this seems like a weird example to pick. I've read that furry artists are well appreciated, because the clients have very specific requirements... no one wants their fursona drawn wrong.

Besides, this seems like it's one of those rare problems that can be solved by some individual advocacy. The furry community isn't huge, and if its artists need to be protected I think it'd respond. Much more than, say, website art directors.

Can you name two or three of these that you've read? Also, [citation needed] on that "fractions of a second" claim.

I avoid AI content out of principle and only know about it because it has become omnipresent. Regarding the second point, it's well-known that LLMs work instantaneously or nearly so like most computer programs.

In other words you're making up the problem and also making up the numbers.

Sure but one could easily fix these problems by adding more components to correct for then. One could incorporate modules into the AI that take care of arithmetic or check facts.

No, one can't "easily" fix these problems, or the mutli-billion-dollar AI companies would have already done so. It's not that they aren't aware of the problem.

The thing is, LLM methods simply don't produce accuracy. Even filtering out obvious errors doesn't help, because LLMs are hallucination machines. You don't get truth even if all the inputs are true: you only get truthiness.

A very minimal example: suppose you train your LLM only on Wikipedia, and let's pretend that Wikipedia is errorless. The output will sound like a Wikipedia article, but nothing about that makes its output true or trustworthy. That's just not how LLMs work.

Earlier in the thread there were links to explanations of how LLMs actually work. I don't feel like finding them for you again, but you need to understand instead of wildly guessing.

[zompist]

Informational: OpenAI's o1 tries to increase reasoning power. The tradeoff though is cost and speed. Queries cost 40 times more than earlier LLMs, and can take 10 to 30 seconds or more.

[bradrn]

[Ouch. I'm sorry, Brad, I hit edit instead of reply, and mangled your post. —z.]

[malloc]

Informational: OpenAI's o1 tries to increase reasoning power. The tradeoff though is cost and speed. Queries cost 40 times more than earlier LLMs, and can take 10 to 30 seconds or more.

So you admit that LLMs are capable of reasoning and not just stochastic parroting or hallucinating as often claimed. Unless you consider the reasoning abilities of o1 spurious and no better than its predecessors.

How many of these people are there, and have they lost their jobs?

There are artists whose jobs I worry about, but this seems like a weird example to pick. I've read that furry artists are well appreciated, because the clients have very specific requirements... no one wants their fursona drawn wrong.

Besides, this seems like it's one of those rare problems that can be solved by some individual advocacy. The furry community isn't huge, and if its artists need to be protected I think it'd respond. Much more than, say, website art directors.

My point was that nobody considers drawing cartoon animals a form of fine art that requires an avant-garde approach to creativity. It simply requires knowledge of anatomy and how to stretch it to create fictional beings like anthropomorphic animals. Sure the current wave of generative AI may not invent whole new styles of art like cubism or dadaism, but they can certainly generate cartoon animals for far less than furry artists would demand.

No, one can't "easily" fix these problems, or the mutli-billion-dollar AI companies would have already done so. It's not that they aren't aware of the problem.

Perhaps not easily but feasibly. The first airplanes struggled to stay airborne for more than a minute but after one decade aerospace engineers were turning out machines capable of dogfighting. Today even the worst airplane can outfly even the fastest and strongest birds as easily as humans outwit bugs.

[zompist]

1. No, intelligence is not a unitary concept, and it would be useless in philosophy— or neuroscience— if it were. To understand intelligence means to break it into smaller pieces and look for continua, not binary states.

Fair enough. Any thoughts on what those ‘smaller pieces’ could be?

Sure, let's start with neurons. One neuron is a tiny computer... exactly how powerful it is we don't really know, but it's connected to 7000 other neurons on average... the top neurons go up to 200,000. Each one generates a signal— the message is not a single pulse, it's a frequency of pulses. Neurons excite or inhibit one another; there are also neurotransmitters which affect them (how and why, I don't think we know for sure).

Right there, that tells us things about cognition that are not intuitive from our POV as conscious beings. E.g., we have processors to burn. This means that the brain can often solve a problem by breaking it up into tiny pieces— e.g. neurons that respond to one tiny chunk of the visual field. Spatial processing is often handled by spatial arrangements of neurons. If you were doing language processing, this is a very different thing from procedural code— e.g. you don't have to "look up a word in the lexicon", with the implication that you have lookup code. You could just send an input to every neuron that encodes a word and see which one responds.

I always recommend Hardin's Color for Philosophers for an introduction to why vision is such a hard problem, and why the brain doesn't work like a TV. (Start with the fact that the eye doesn't focus on one spot— it is constantly moving. That makes the problem of 'what you're looking at' much harder.) At a low level we can look for things like edge detection, which builds up into object detection and movement detection.

The problem of course is that neuroscience is bottom-up and doesn't mesh too well with top-down intuition (and metaphors based on computers). I've read a bunch of books by neuroscientists— they are always interesting but they always theorize way way beyond the data.

OK, learning and accuracy are critical, but I see intelligence on its own as being fundamentally amoral. To me morality depends more on consciousness, as stated already.

OK, for "morality" substitute "meaning". Any animal, even a bacterium, needs to know what is important, what is a threat, what is beneficial, what can be ignored.

(And ‘common sense’ as a criterion could well result in myself being considered unintelligent!)

Sounds like you've been internalizing neurotypical propaganda. Differences among humans are interesting, but tell us almost nothing about "human intelligence" except, again, that its not unitary.

And yet I note that, at the end of this; you anchor this in how shrews’ external behaviour differs from humans’. It all ultimately comes down to observations of behaviour.

I don't think that's a fair characterization of human ethologists' methods or theories. They are not Skinnerians, and they are not afraid to maintain that animals have mental states. Recall that we do have direct observation of one animal's inner consciousness. We can't overapply empathy, but we can use it.

Conversely: if some animal could navigate as well as humans but using an inhuman cognition to do so, would we be able to tell? Would it matter that there are underlying differences?

Think two minutes about how a dog functions in the world. It should be obvious that yes, we can tell, and yes, it matters. Also note that we are not limited to watching animal behavior; we also have anatomy and can figure out, e.g., senses or eye architectures that humans don't have.

[zompist]

Informational: OpenAI's o1 tries to increase reasoning power. The tradeoff though is cost and speed. Queries cost 40 times more than earlier LLMs, and can take 10 to 30 seconds or more.

So you admit that LLMs are capable of reasoning and not just stochastic parroting or hallucinating as often claimed.

Try to keep up: the present discussion is whether LLMs display intelligence. You want to assume your conclusion again, which is not interesting.

[Ketsuban]

Artificial intelligence has a level of autonomy that other machines fundamentally lack. Printing presses and cameras, however disruptive their abilities, do not generate their own content but only replicate what we deliberately put into them. Generative AI by contrast can write entire novels and draw elaborate pictures without human guidance beyond the initial prompt. Whether you consider it capable of true thought or not, it clearly acts in effectively non-deterministic ways unlike previous technologies.

This is straight-up untrue, for the record: latent diffusion models have training data, the same as everything else. Things which are well-represented in the training data are robustly represented in the output, things represented more sparsely in the training data are represented more weakly in the output, and things not represented in the training data at all are not represented in the output.

Latent diffusion started off as a method for noise reduction: you train it on pairs of images where one of the pair is the result of adding noise to the other accompanied by a description of what the original contains, and the end result is a model which can remove noise from an image to produce (something very much like) the original image. The realisation was that since the description is already tokenised you can start with literal randomness and tell the model to remove the noise to uncover an original image that doesn't actually exist which somewhat matches the description (which is now a prompt for what the image should contain). That random initial state is what you're interpreting as "nondeterminism".

My point was that nobody considers drawing cartoon animals a form of fine art that requires an avant-garde approach to creativity. It simply requires knowledge of anatomy and how to stretch it to create fictional beings like anthropomorphic animals. Sure the current wave of generative AI may not invent whole new styles of art like cubism or dadaism, but they can certainly generate cartoon animals for far less than furry artists would demand.

The output of something like Stable Diffusion has no consistency, and as Zomp says people want their fursona represented consistently between pictures. It's neat for brainstorming character ideas based on mashing together ideas in its training data since unlike a human it has no preconceived notions of what a fusion of two concepts looks like, but not much beyond that.

[sasasha]

How many of these people are there, and have they lost their jobs?

There are artists whose jobs I worry about, but this seems like a weird example to pick. I've read that furry artists are well appreciated, because the clients have very specific requirements... no one wants their fursona drawn wrong.

Besides, this seems like it's one of those rare problems that can be solved by some individual advocacy. The furry community isn't huge, and if its artists need to be protected I think it'd respond. Much more than, say, website art directors.

My point was that nobody considers drawing cartoon animals a form of fine art that requires an avant-garde approach to creativity. It simply requires knowledge of anatomy and how to stretch it to create fictional beings like anthropomorphic animals. Sure the current wave of generative AI may not invent whole new styles of art like cubism or dadaism, but they can certainly generate cartoon animals for far less than furry artists would demand.

Let's run with this scenario a bit further.

As zomp pointed out, the client wants something personal, something they identify with (or even, as). They also want something consistent: from one image to the next, they neither want their fursona growing extra arms, or, say, its facial features radically altering. It is a representation of their being, self-imagined and defined, and the details matter. These details, however, are – like intelligence – non-unitary: they are somewhat malleable, a fluid river of results from a creative discussion, if you like, between artist and client.

GAI can generate as many images as you've got time and money and the server has bandwidth for. (FYI these can be very present limitations: my last 10 attempts to use Craiyon to make illustrations for this discussion have been met with the error message "Too much traffic", and whilst I got an alternative AI image generator to do the job, I would have had to pay not an insignificant amount to get what I wanted out of it, as it took all my 'credits' to get it only nearly right.) You can have a sort of "creative discussion" with it – but that discussion is slow, frustrating, expensive, and incredibly low-quality: what human artist would ruin an otherwise acceptable image with the random addition of an extra arm?

Practically every image I have ever asked AI to generate has had some bizarre flaw like this. It's not a problem with parameters. You can tell it to avoid doing x, y and z but (a) it won't listen reliably, and (b) you can't tell it to avoid everything that would be, to you, inappropriate for the task at hand. That would be an infinite task. This point gets back to discussions about the differences between cognition and procedural prediction.

Lets say we got one image of our fursona that we were really happy with. What then? How do you get the AI to keep drawing variations of that character? It won't, even if you upload the image and tell it to keep all the details the same; it will draw the analogy of zompist's excellent designation: truthiness. It can't care that you want to see a stylised representation of yourself, or intuit the steps necessary to satisfy you, or think outside the box to come up with something you didn't realise you wanted (a trick many artists make their bread and butter from, which comes from human cognition, human intuition, human empathy, etc). It just spaffs out the easiest possible interpretations of your input data based on its training data with absolutely no regard for your humanity.

This is, frankly, not artistry. The people threatened by it are not artists, but clients, who may think they are paying for something artistically valuable in purchasing AI services.

However, artists can use AI artistically. It's an art to get it to produce worthwhile stuff (and then edit out extra arms and other random reality-like gunk). As this tool develops, I have no doubt that artists specialising in its use will develop, and it will be interesting to see whether this new canvas will one day hold truly great works of art.

Many people had similar objections to photography when it was new: it would kill artistry. It has changed artistry, but painting is not dead because photography exists.


To return quickly to our previous discussion:

That makes no sense whatsoever.

What made no sense whatsoever was your bizarre assertion that machines are somehow mine, to which this was obviously a response. With humans as with AI: if you don't want nonsense in the output, don't put nonsense in the input.

Art and literature are not drudgery that creatives resent having to make for survival. Those are precisely the kind of jobs that people wish they could have instead of toiling in factories and offices. You might have a point if we were talking about industrial robots kicking people out of factories but that doesn't really apply here.

I'm a professional artist, and I'm afraid work is still drudgery. I maintain a day's work a week in a physically gruelling, rather dangerous role, for contrast, because I find being an artist full time too draining. You'll find that many artists feel the same: Cage and Feldman maintained a removals company together for much of their lives, even as their works were played and studied around the world.

[malloc]

@sasasha: Then how do you explain the sheer glut of AI images on the internet today? The artistic websites I frequent are choked with them unless I choose to filter them. Perhaps furry art is not the best example given its idiosyncratic requirements, but the general point remains. Remember that last year, Hollywood writers had to go on strike to avoid studios replacing them with LLMs, demonstrating that generative AI can easily replace even the best human writers and only deliberate resistance prevents it from taking over screenwriting.

@ketsuban: Let me clarify that I am not claiming generative AI is literally outside the bounds of causality, only that it acts in ways that we cannot predict or directly control. Printing presses only print what type you put into the press, cameras only produce images of whatever you point them at, but you have no idea what text or images generative AI will produce beyond the broad limitations of its training data. It effectively makes all the decisions (even if through pseudorandom means) about your drawing or novel beyond the initial prompt.

[keenir]

"Is anyone else amused by the fact that AI just lost its job -- to AI?" -John Stewart; The Daily Show.

[keenir]

How many of these people are there, and have they lost their jobs?
There are artists whose jobs I worry about, but this seems like a weird example to pick. I've read that furry artists are well appreciated, because the clients have very specific requirements... no one wants their fursona drawn wrong.

Besides, this seems like it's one of those rare problems that can be solved by some individual advocacy. The furry community isn't huge, and if its artists need to be protected I think it'd respond. Much more than, say, website art directors.

My point was that nobody considers drawing cartoon animals a form of fine art that requires an avant-garde approach to creativity.

I've been on both sides of a request to draw a fictional character...and I've still never seen an AI draw a flying echidna. your saying that its all cartoon animals and poo-pooing the sort of art required, tells me that you've never tried it.

It simply requires knowledge of anatomy and how to stretch it to create fictional beings like anthropomorphic animals. Sure the current wave of generative AI may not invent whole new styles of art like cubism or dadaism, but they can certainly generate cartoon animals for far less than furry artists would demand.

and yet, as Zompist pointed out, its not entirely a matter of $$...because the clients want accuracy and their mental image of the character.

No, one can't "easily" fix these problems, or the mutli-billion-dollar AI companies would have already done so. It's not that they aren't aware of the problem.

Perhaps not easily but feasibly. The first airplanes struggled to stay airborne for more than a minute but after one decade aerospace engineers were turning out machines capable of dogfighting. Today even the worst airplane can outfly even the fastest and strongest birds as easily as humans outwit bugs.

not sure what relevance that has.

[keenir]

@sasasha: Then how do you explain the sheer glut of AI images on the internet today?

nobody said that nobody would spam the internet with AI images.

Remember that last year, Hollywood writers had to go on strike to avoid studios replacing them with LLMs, demonstrating that generative AI can easily replace even the best human writers and only deliberate resistance prevents it from taking over screenwriting.

that doesn't mean the scripts are any good - only that there were people at the studios who were willing to save money by using AI scripts that were worse than some of the human scripts that critics say are cliche and not as original as in the Golden Age.

[malloc]

that doesn't mean the scripts are any good - only that there were people at the studios who were willing to save money by using AI scripts that were worse than some of the human scripts that critics say are cliche and not as original as in the Golden Age.

The scripts must have been good enough for studio executives to believe they would lead to hit or at least watchable films, otherwise they wouldn't have considered them in the first place. Certainly the film industry has its problems, but I cannot imagine everyone agreeing to put out unwatchable nonsense just to save money. Everyone would simply stop watching new movies and the whole film industry would collapse.

I've been on both sides of a request to draw a fictional character...and I've still never seen an AI draw a flying echidna. your saying that its all cartoon animals and poo-pooing the sort of art required, tells me that you've never tried it.

That seems quite easy to remedy. You could just ask the generative AI to draw a flying echidna and it would easily draw one just like it draws anything else. Presumably the AI has echidna reference images and information on how to depict flight in its database.

Regarding early airplanes, my point is that they went from completely useless to winning military battles in the space of one decade. Technology can advance amazingly fast and what seems like a worthless novelty today can radically change the world sooner than we expect.

[Ketsuban]

The scripts must have been good enough for studio executives to believe they would lead to hit or at least watchable films, otherwise they wouldn't have considered them in the first place.

You have a quantity of faith in the ability of studio executives to make good creative decisions which I suspect basically any creative professional would find excessive—I doubt there's a director credited who doesn't have a story about needing to deal with bozotic suggestions from executives. Mel Brooks talks in his memoir about being handed a legal pad and told to take a batch of notes on what he couldn't do in Blazing Saddles, dutifully noting down every veto and eagerly yessing everything the suit was saying, then once he was free immediately throwing the page in the bin. It's such a well-recognised pattern that the field of game development gives us the story of a creative adding an entirely superfluous duck for the executive to remove so they feel like they had input without damaging any of the actual work.

[zompist]

That seems quite easy to remedy. You could just ask the generative AI to draw a flying echidna and it would easily draw one just like it draws anything else. Presumably the AI has echidna reference images and information on how to depict flight in its database.

Just for fun, I tried asking for pictures of flying echidnas using Gemini, one of Google's free products. It produced impressively rendered pics, but I'd put them at the level of "badly photoshopped." It simply threw in either butterfly or bird wings behind the animal. And I couldn't get it to draw in a furry style (despite adding prompts like "in a furry style" or "anthropomorphic").

Regarding early airplanes, my point is that they went from completely useless to winning military battles in the space of one decade. Technology can advance amazingly fast and what seems like a worthless novelty today can radically change the world sooner than we expect.

You really have to stop scaring yourself with fallacies. You've got two going on here, extrapolation and induction. Neither prove anything.

We've had spaceships for 68 years, and guess what, we've never got a human to Mars. We've never gotten a spaceship to the nearest star. We don't have a cure for cancer. We don't commute with personal jetpacks. Sorry, it's just not true that "if you can imagine it, it will happen", much less in "one decade."

[malloc]

Just for fun, I tried asking for pictures of flying echidnas using Gemini, one of Google's free products. It produced impressively rendered pics, but I'd put them at the level of "badly photoshopped." It simply threw in either butterfly or bird wings behind the animal. And I couldn't get it to draw in a furry style (despite adding prompts like "in a furry style" or "anthropomorphic").

It makes sense that the free version of this technology would yield less impressive results. They probably reserve the full power of generative AI for paying customers. Even so, the AI demonstrates some remarkable intuition, that animals can only fly with wings, that the most common wing types are avian or insectoid*, and so forth. You can hardly criticize a free-to-use prototype AI for failing to consider the speculative evolution of volant monotremes in a biologically rigorous manner.

You really have to stop scaring yourself with fallacies. You've got two going on here, extrapolation and induction. Neither prove anything.

We've had spaceships for 68 years, and guess what, we've never got a human to Mars. We've never gotten a spaceship to the nearest star. We don't have a cure for cancer. We don't commute with personal jetpacks. Sorry, it's just not true that "if you can imagine it, it will happen", much less in "one decade."

So you believe that instead of continuing its rapid advance like aerospace, AI will soon hit a wall like space exploration or cancer research? My understanding has always been that space travel and possibly cancer research have been stymied by lack of funding more than anything. Space travel received enormous funds in the sixties for the moon landing but once that was accomplished, the funding shrank considerably. If we spent nearly a trillion dollars a year on space research like the military, we probably could have gotten to Mars by now.

Artificial intelligence by contrast receives an enormous amount of funding from the world's richest people and political support from world leaders. Unless the money printer suddenly shuts off tomorrow, AI researchers have all the funding they need to push the technology to its physical limits. There is no reason to think those physical limits preclude massive improvements in AI including self-awareness and human-level intelligence.

* Though perhaps true creativity and insight would be giving the echidna the bat wings fitting for a flying mammal...