AIs gunning for our precious freelancers

[Raphael]

malloc, do you still believe that Einstein got so famous for being good at calculating things in his head like a calculator app?

[keenir]

Only because that's a baseline thing for humans. It's beyond present-day AIs, and other animals for that matter.

For literally fifty years, AI researchers have realized that showy stuff like playing chess is easier than real-world knowledge, and far easier than actually doing things in the real world.

In that case, why can plenty of animals navigate the world just fine while none of them can play chess?

because computer programs can be programmed to do showy things that have no IRL relevance beyond abstractions.

Why can every human find their way around the kitchen while many cannot play chess?

because many of them play CONNECT FOUR.
;p

Furthermore it is small comfort if humans find themselves locked into menial jobs like fetching coffee while AI gets all the academy awards and Nobel prizes.

why would AI be awarded prizes?

Because once it becomes advanced enough,

so...never.

it will conduct all the scientific research

except AI don't say "well thats curious" or "why did that happen?", so what would it research? oh, right - nothing.

and write all the literature.

Why would it write all that? Intelligence developed so its possessors would be able to survive through the process of performing less work. (ie, if i make a spear, i don't have to run as fast or as far after that elk)

[Raphael]

keenir, I mostly agree with you, but

Because once it becomes advanced enough,

so...never.

"Never" is a very long time. I agree that the current LLM approach isn't likely to get us to human-like AI, but other approaches - perhaps approaches that no one has thought of yet - well might, at some point in the distant future.

[Raphael]

Why can every human find their way around the kitchen while many cannot play chess?

Because there's no single thing called "intelligence", there are different mental skills. Having a lot of one mental skill doesn't in any way guarantee that one will be good at another mental skill.

To use a personal example, back when I was in school, I knew people who were a lot better than me at math, but who, at the same time, seemed to have serious trouble understanding what seemed like very simple concepts to me in physics. And math and physics are usually seen as two closely related fields.

Among people more famous than the ones with whom I went to school, there are both people who seem to be very smart about specific STEM subjects, while believing idiotic things about humanities-related subjects (Anatoly Fomenko), and people who seem to be very smart about humanities-related subjects while believing idiotic things about specific STEM subjects (Thabo Mbeki).

[WeepingElf]

keenir, I mostly agree with you, but

Because once it becomes advanced enough,

so...never.

"Never" is a very long time. I agree that the current LLM approach isn't likely to get us to human-like AI, but other approaches - perhaps approaches that no one has thought of yet - well might, at some point in the distant future.

Yes. It is theoretically possible to build a machine with human-like intelligence. But it is also theoretically possible to fly to Alpha Centauri. Neither is likely to be achieved within our lifetime.

And what regards AI-made texts, music and the like, it is a question of legislation. Are AI-made media products copyrightable? Do people using AIs have to pay royalties to the authors of the data they train their AIs on? If AI-made stuff is not copyrightable, and involves paying royalties to the authors of the training data, it becomes economically pretty unattractive.

And of course, AIs need not be admitted to contests such as the chess world championship.

[keenir]

Because once it becomes advanced enough,

so...never.

"Never" is a very long time. I agree that the current LLM approach isn't likely to get us to human-like AI, but other approaches - perhaps approaches that no one has thought of yet - well might, at some point in the distant future.

I know Malloc is going to leap on your statement as proof that their utter terror is vindicated and will come to pass...so I'll pivot to ask: do you foresee that the "approaches that no one has thought of yet" will result in an outcome that makes Skynet look like a humanitarian?

[Raphael]

I know Malloc is going to leap on your statement as proof that their utter terror is vindicated and will come to pass

Just for the record, last time I checked, malloc (like me) had "He/him" in his signature.

...so I'll pivot to ask: do you foresee that the "approaches that no one has thought of yet" will result in an outcome that makes Skynet look like a humanitarian?

I won't rule it out, but I don't think it's in any way guaranteed. As they say, predictions are difficult, especially about the future.

[keenir]

I know Malloc is going to leap on your statement as proof that their utter terror is vindicated and will come to pass

Just for the record, last time I checked, malloc (like me) had "He/him" in his signature.

Normally i just use "Malloc" instead.

So noted.

[malloc]

And what regards AI-made texts, music and the like, it is a question of legislation. Are AI-made media products copyrightable? Do people using AIs have to pay royalties to the authors of the data they train their AIs on? If AI-made stuff is not copyrightable, and involves paying royalties to the authors of the training data, it becomes economically pretty unattractive.

Easy enough to fix with lobbyists pushing for new legislation. It would only take one session of Congress or parliament to rewrite the laws on copyright so that corporations can claim copyright on AI generated works. Given the immense potential for profit that such a move allows, we should anticipate that they will try.

And of course, AIs need not be admitted to contests such as the chess world championship.

Sure but can AIs conducting scientific research or writing literature get Nobel prizes?

[keenir]

And what regards AI-made texts, music and the like, it is a question of legislation. Are AI-made media products copyrightable? Do people using AIs have to pay royalties to the authors of the data they train their AIs on? If AI-made stuff is not copyrightable, and involves paying royalties to the authors of the training data, it becomes economically pretty unattractive.

Easy enough to fix with lobbyists pushing for new legislation.

AI have lobbyists now?

It would only take one session of Congress or parliament to rewrite the laws on copyright so that corporations can claim copyright on AI generated works. Given the immense potential for profit that such a move allows, we should anticipate that they will try.

claim copywrite on things people don't buy? AI generated art is free, thats the whole attraction for some people to use it. take away that advantage, and its dead in the water.

And of course, AIs need not be admitted to contests such as the chess world championship.

Sure but can AIs conducting scientific research or writing literature get Nobel prizes?

no.

partly because they aren't qualified, partly because they aren't eligable, and partly because they don't do those things.

[malloc]

claim copywrite on things people don't buy? AI generated art is free, thats the whole attraction for some people to use it. take away that advantage, and its dead in the water.

Not necessarily. It only needs to cost less than human-made art to remain economically viable. Imagine a film studio using AI to generate entire films without the need for writers, directors, or actors. The studio would save billions from not having to pay so many salaries and could reduce the price of its films enough to outcompete studios still using humans for everything while still making massive profits because they've eliminated massive expenses. Remember that studios came very close to replacing human screenwriters with AI and it took considerable resistance to convince them to pause that plan.

partly because they aren't qualified, partly because they aren't eligable, and partly because they don't do those things.

Sure but nothing in principle prevents changes to the rules to make them eligible. Imagine a future where AIs win all the Nobel prizes because they can write and research so much faster and even better than us.

[linguistcat]

malloc, let's forget about some nebulous future time for the moment and think about the present and very near future. Even within AI, there are better ones to fear than LLMs and I am saying this as a writer with mostly creative types as friends. AI art without being edited/fixed/completely redone by humans still sucks. It's just fast.

The AI I'm worried about is the ones being used in those weird robot dogs that will probably be deployed by police, and anything that helps military "target enemies".

[Torco]

Because once it becomes advanced enough, it will conduct all the scientific research and write all the literature. Certainly the credit wouldn't go to whatever random know-nothing prompted the AI for revolutionary research or an interesting novel.

look at AI slop SM accounts: who does their ad revenue go to, chat gpt or the prompter? [psst, it's the prompter, though okay some of it indirectly goes to OpenAI in the form of subscription fees]. AI doesn't "conduct scientific research", people use AI to write papers which they, as people, then submit. people use AI to write essays which then they, as people, submit to their teachers [and their teachers probably use the same AI to grade them]. point is, again, models are not people, and they won't behave as people as they gain increasing capability.

Like, for real, people are very worried about AI becoming like a guy, gaining sentience, and then starting to of its own accord open a bank account, put money together, and start doing stuff.. there's no reason to believe it will do any of those things, but you don't need any of that for things to get hairy.

[malloc]

look at AI slop SM accounts: who does their ad revenue go to, chat gpt or the prompter? [psst, it's the prompter, though okay some of it indirectly goes to OpenAI in the form of subscription fees]. AI doesn't "conduct scientific research", people use AI to write papers which they, as people, then submit. people use AI to write essays which then they, as people, submit to their teachers [and their teachers probably use the same AI to grade them]. point is, again, models are not people, and they won't behave as people as they gain increasing capability.

So you believe humans will continue to claim credit for scientific research and literature, even though AI has taken over all the actual work. Scientists will prompt AI to design and conduct experiments, write papers on the results, and so forth, and still get notional credit even though they contributed nothing to the process. Likewise with people getting their names slapped on novels and screenplays they didn't write or even necessarily read.

malloc, let's forget about some nebulous future time for the moment and think about the present and very near future. Even within AI, there are better ones to fear than LLMs and I am saying this as a writer with mostly creative types as friends. AI art without being edited/fixed/completely redone by humans still sucks. It's just fast.

Except that plenty of AI generated images and text don't suck. Early versions of image generators struggled with stuff like hands but they have improved massively over the past few years. They used to have obvious tells like garbled text and a distinct fuzzy texture, but these days it has become increasingly difficult to tell and I honestly have to trust self-described artists these days. Meanwhile current LLMs can generate entire books on any subject or fictional topic you like.

The AI I'm worried about is the ones being used in those weird robot dogs that will probably be deployed by police, and anything that helps military "target enemies".

Well yes, those are certainly a problem as well. One should not mistake my focus on generative AI for approval of other AI models by any means. Large language models and the like have become the most prominent form of AI these days, though, the ones that dominate the news and economic forecasts. Rest assured that I have no support for weaponized AI in the military or police either.

[linguistcat]

...
Except that plenty of AI generated images and text don't suck. Early versions of image generators struggled with stuff like hands but they have improved massively over the past few years. They used to have obvious tells like garbled text and a distinct fuzzy texture, but these days it has become increasingly difficult to tell and I honestly have to trust self-described artists these days. Meanwhile current LLMs can generate entire books on any subject or fictional topic you like.

They might have those details down, and every time a new tell is found, the creators will likely retrain them to fix it. But I've seen soldiers running away from the beach in what was supposed to be D-Day, because the machine didn't understand the actual event and the person who prompted it couldn't even see a glaring detail like that being incorrect. And that's one example off the top of my head where the "technical" stuff was right, but the art was still shit.

If you interact with AI art as actual art, as an intentional choice made on the prompter's part as artistic expression in good faith, where the AI is a tool they have chosen to use, most AI art sucks. It is technically impressive but says nothing. Or even says the opposite of what the prompter says they intended. I don't know how to explain to you that someone good at art can learn to use various tools to make art, but someone bad at art will still be bad regardless of what tools they use, AI or not. And this isn't coming from a technical standpoint, but just being able to covey some kind of message.

I bet in the future, having a very distinct style for writing or visual art will be sought after, at least until there's enough works by someone that they can dupe the style with AI.

[Torco]

look at AI slop SM accounts: who does their ad revenue go to, chat gpt or the prompter? [psst, it's the prompter, though okay some of it indirectly goes to OpenAI in the form of subscription fees]. AI doesn't "conduct scientific research", people use AI to write papers which they, as people, then submit. people use AI to write essays which then they, as people, submit to their teachers [and their teachers probably use the same AI to grade them]. point is, again, models are not people, and they won't behave as people as they gain increasing capability.

So you believe humans will continue to claim credit for scientific research and literature, even though AI has taken over all the actual work. Scientists will prompt AI to design and conduct experiments, write papers on the results, and so forth, and still get notional credit even though they contributed nothing to the process. Likewise with people getting their names slapped on novels and screenplays they didn't write or even necessarily read.

pretty much, yes. at least in the foreseeable future. we can speculate about some possible point in time in which AI are people, demand rights and recognition, are given them, and then papers are published under "chatgpt 2031" instead of under "zhang, smith and contreras 2031". but for now, yeah, this is what i think not only will happen, but is happening. because, again, llms are software, not people, and so people use them to do stuff

like, i have an instance of gpt running on my computer right now. maybe i'll eventually notice it wrote a paper and sent it on its own, maybe... but right now i can have it write a paper and I could probably get said paper published somewhere.

[Raphael]

Just to add to linguistcat's points, I'd have no idea how to get any AI to turn any of the vague ideas for books that occasionally float to the top of my head into actual books.

[zompist]

Why can every human find their way around the kitchen while many cannot play chess?

Because there's no single thing called "intelligence", there are different mental skills. Having a lot of one mental skill doesn't in any way guarantee that one will be good at another mental skill.

Underlining this, because it's malloc's second biggest mistake. (The first is assigning agency to AIs rather than their creators/owners.)

We're all linguists here, we should be aware of the dangers of reification. "Intelligence" is a catchall term, fine as a loose description, terrible as a reified concept. Treating it as a single thing, like "height", distorts our thinking about humans, about animals, and about "AI".

This may be easier to see if we look at "animal intelligence." Notice first that people use the word differently referring to humans (where only some people are called intelligent) and animals (where the question is whether animals have capacities all human have). And then, notice that there are practical and philosophical problems with getting into a non-human's head. Is a wolf smart because it can hunt and kill prey? An ant can do the same with a brain the size of a pinhead. Biologists have to be super-careful, because it's possible to both overestimate and underestimate animal behaviors.

And if animal intelligence is tricky, machine intelligence is more so. There was an SF story where aliens, having observed Earth from space, deduced that its intelligent species was the automobile. Can we prove them wrong with only the tools available to them (visual observation from space)?

[keenir]

There are times that I wonder if Malloc is, in fact, an AI trying to convince us all of the danger and superiority of AI...without also revealing that Malloc is an AI.

look at AI slop SM accounts: who does their ad revenue go to, chat gpt or the prompter? [psst, it's the prompter, though okay some of it indirectly goes to OpenAI in the form of subscription fees]. AI doesn't "conduct scientific research", people use AI to write papers which they, as people, then submit. people use AI to write essays which then they, as people, submit to their teachers [and their teachers probably use the same AI to grade them]. point is, again, models are not people, and they won't behave as people as they gain increasing capability.

So you believe humans will continue to claim credit for scientific research and literature, even though AI has taken over all the actual work

thats what humans did when the work was done by women and minorities, so why shouldn't we return to doing that with AI in the workers' seat?

. Scientists will prompt AI to design and conduct experiments, write papers on the results, and so forth, and still get notional credit even though they contributed nothing to the process. Likewise with people getting their names slapped on novels and screenplays they didn't write or even necessarily read.

yup.

malloc, let's forget about some nebulous future time for the moment and think about the present and very near future. Even within AI, there are better ones to fear than LLMs and I am saying this as a writer with mostly creative types as friends. AI art without being edited/fixed/completely redone by humans still sucks. It's just fast.

Except that plenty of AI generated images and text don't suck. Early versions of image generators struggled with stuff like hands but they have improved massively over the past few years. They used to have obvious tells like garbled text and a distinct fuzzy texture, but these days it has become increasingly difficult to tell and I honestly have to trust self-described artists these days.

so, because you can't tell the difference between an AI image and the villain of The Princess Bride, nobody can?

Meanwhile current LLMs can generate entire books on any subject or fictional topic you like.

oh, those "The (noun) did (verb) to (noun) (date)" stories. {and yes, there are books for beginner writers and children, which have that layout, with blank lines asking for a type of work}

The AI I'm worried about is the ones being used in those weird robot dogs that will probably be deployed by police, and anything that helps military "target enemies".

Well yes, those are certainly a problem as well. One should not mistake my focus on generative AI for approval of other AI models by any means.

o-kay.

Rest assured that I have no support for weaponized AI in the military or police either.

You don't? Then how are your the authoritarians that you're a fanboy for, going to take over the world?

[keenir]

and in other news, the UAE tried to give money to support one company's attempt to develop AI...and Musk shot it down with the Trump administration canon.

remind me again how the techbros are marching in lockstep to bring us AI swiftly.

claim copywrite on things people don't buy? AI generated art is free, thats the whole attraction for some people to use it. take away that advantage, and its dead in the water.

Not necessarily. It only needs to cost less than human-made art to remain economically viable. Imagine a film studio using AI to generate entire films without the need for writers, directors, or actors. The studio would save billions from not having to pay so many salaries and could reduce the price of its films enough to outcompete studios still using humans for everything

great...so the studios save billions. but where are they getting their money, if you don't need people? (remember that many movies' producers are either the writers, directors, or actors)

and if people were refusing to go see movies where the actors were mistreated or cheated, imagine how much money those studios would lose when all the actors are replaced by your AI.

partly because they aren't qualified, partly because they aren't eligable, and partly because they don't do those things.

Sure but nothing in principle prevents changes to the rules to make them eligible.

and why would the Nobels (or any other science prizes) have their rules changed for AIs? last I checked, the King of Sweden isn't a greedy so-and-so who values $ before truth and all else.

Imagine a future where AIs win all the Nobel prizes because they can write and research so much faster and even better than us.

you're assuming they can research, and not "perform google search" -> "copy&paste first 1,000 results" -> "add bullet point justification" -> "send"