AIs gunning for our precious freelancers

[rotting bones]

You know what would be really funny? If upscaling AIs only hallucinated Obama into a white man, not black men in general.

[Raphael]

Imagine robots mining and refining the ore, more robots manufacturing silicon wafers, and so forth all controlled by artificial intelligence.

[..]

Imagine something like an LLM but instead of composing text in response to prompts, it chooses a particular AI and drafts a prompt based on stimuli.

[...]

Imagine something like the Burj Khalifa but filled with billions and billions of GPUs and so forth.

Imagine all the robots
Livin' for today...


OK, more seriously, we might get a machines-do-everything civilization in a hundred years or a thousand years, but I don't see it happening in the near future.

And I don't see any reason to believe that AI is racing ahead of us.

[keenir]

show me an AI that can survive outside of the captive enviroment it depends upon.

Sufficiently powerful AI could control all the industry necessary to maintain itself, though.

sufficiently powerful AI can't even make toast.

Imagine robots mining and refining the ore, more robots manufacturing silicon wafers, and so forth all controlled by artificial intelligence. Nothing in principle prevents an entire civilization of intelligent machines with nary a human in sight.

not sure if you've been watching too many TERMINATOR movies, or History Channel programs about what machines might accomplish in interstellar space,

we have explained it to you, MANY MANY TIMES.
AI is specialized, even beyond things like cheetahs. the protein-folding AI can't win at JEOPARDY, for example, nor play chess.

Sure but

"I don't want to listen to you people; i want to fearmonger!"

considered in the aggregate, artificial intelligence has mastered so much of human activity and continues to advance everyday.

human programming continues to advance. you have yet to offer up an actual AI program that advances itself.

Presumably the next step is some kind of master AI that specializes in choosing the appropriate model for a given task.

the next step...for whom? the next step by whom? every AI thus far, has been designed by a human, and to perform a specialized task; why would a human design that master AI? (other than to turn your ravings into prophetic words)

Well, for one, AI software runs on computer hardware, and there are physical limits to how fast computer hardware can get. That's why "the Singularity" as originally imagined can't happen; it would require the processing power of microchips to keep on increasing for the foreseeable future, which can't work. At some point, you get to individual electronic elements the size of a few atoms, and then you can't make them smaller any more.

The tech industry already solved or at least mitigated that problem by using data centers and server farms full of hardware instead of trying to pack all the processing power necessary for AI into one small machine. They have massive warehouses with millions of chips all linked together to process far more data than any individual chip could manage. If necessary they could simply stuff more and more chips with maxed out processing power into the building

clearly you don't know microchips generate geat - you stuff chips together, they melt.

Imagine something like the Burj Khalifa but filled with billions and billions of GPUs and so forth.

so now you're just building castles in the sky, almost literally.

Meanwhile we humans already reached the limits of our processing power hundreds of thousands years ago when our brains reached more or less their present size

and that processing power can create true art, probes to Pluto, and computer programs..

Artificial intelligence is racing past us while we remain stuck in quicksand.

I've seen limestone race faster in its growth, than AI has been.

[zompist]

The original picture is here. No one made it up; an AI really really turned a picture of Obama into a white man. Now you could point out that that was 5 years ago and image processing has improved, but you don't, you just cheat.

I already did. I answered in the past about how engineers try to debias models. Raphael's point was about my post. I only said it's amazing that AI can interpolate details to form crisp images from blurry images. I shrank the image to a quarter size. That's enough for a comparison to Star Trek's Enhance, right?

I agree that AI has significant hallucinations. I reported some of them here when ChatGPT first came out. IIRC It was unable to compare two dates and figure out which item is older. This was before any of the massive media blow up. I think recent models would be better at the date test I posted here.

I have repeatedly stressed that the models are unreliable, should not be used for profit and businesses should hire humans. What do you want from me?

To stop misrepresenting other people for propaganda purposes.

You say "AI can interpolate details to form crisp images from blurry images". And that's true, I agree! See, it's not that hard to say true things instead of making up things people never said.

In the earlier conversation, somewhere in this thread, you were maintaining that the AI was somehow finding out what was really there, which it isn't. What it does is, as you say above, make the image crisp. It does this by putting in details that are likely to be there, based on its training data.

[alice]

Today's fun AI-related link: AI system resorts to blackmail if told it will be removed. Found when looking for something online to complement the report on today's BBC news about Google's AI-generated movies now incorporating speech.

[rotting bones]

To stop misrepresenting other people for propaganda purposes.

You say "AI can interpolate details to form crisp images from blurry images". And that's true, I agree! See, it's not that hard to say true things instead of making up things people never said.

In the earlier conversation, somewhere in this thread, you were maintaining that the AI was somehow finding out what was really there, which it isn't. What it does is, as you say above, make the image crisp. It does this by putting in details that are likely to be there, based on its training data.

It's not just based on the training data. There is a probabilistic factor, but it's based on the mathematical properties of folds in 3D surfaces. That's what the model is trying to infer, and it does work in many cases. You can see how much success can be achieved so far if you look at the papers.

I don't thinking blowing up microscopic images is an accurate description of Star Trek's Enhance. The example I presented is much closer to what I said in the original post. At any rate, it's the kind of thing I was thinking of.

I don't think anyone should be surprised that the smaller the input image, the less information it contains. So the less likely the upscaled image would be to resemble the depicted object. I would also train a separate model for the much harder task of upscaling microscopic images. I expect the underlying statistical inference tasks are not the same. Applying a normal upscaling model to that task is an interesting experiment, but a misuse from an end user standpoint.

[malloc]

Today's fun AI-related link: AI system resorts to blackmail if told it will be removed. Found when looking for something online to complement the report on today's BBC news about Google's AI-generated movies now incorporating speech.

So now AI can produce films complete with dialog and blackmail humans? When will we start taking it seriously instead of dismissing it all as smoke and mirrors?

the next step...for whom? the next step by whom? every AI thus far, has been designed by a human, and to perform a specialized task; why would a human design that master AI? (other than to turn your ravings into prophetic words)

You should ask the people trying to create AGI and bring about the Singularity why they have devoted so much effort and resources to such things. Whether you believe it or not, there are numerous people obsessed with such things, many of them with billions of dollars at their disposal.

clearly you don't know microchips generate geat - you stuff chips together, they melt.

so now you're just building castles in the sky, almost literally.

Everyone familiar with computers knows they generate waste heat. Computer engineers already take that into account and incorporate cooling systems into data centers, hence the controversy over AI using so much water. Likewise the existence of massive skyscrapers like the Burj Khalifa demonstrates that huge buildings for billions of chips are certainly possible. Now it might turn out that AGI requires so many chips in one place that it becomes impractical to power and cool. Perhaps a machine with the cognitive abilities of Trump requires multiple city blocks and the power output of a medium sized country. Nonetheless we have not reached the point where we can reach that conclusion.

I've seen limestone race faster in its growth, than AI has been.

Over the past decade, AI has gone from science fiction speculation for the most part to folding proteins, writing novels, drawing pictures, composing music, creating highly convincing video complete with speech, and much more. You can criticize AI for many things, but sluggish growth is certainly not one of them. The average LLM can write better than almost anyone on this forum on pretty much any subject you care to name or even pure fiction while recent AI models can produce full video faster than any human with a camera, yet you dismiss it as limestone growth.

[zompist]

In the earlier conversation, somewhere in this thread, you were maintaining that the AI was somehow finding out what was really there, which it isn't. What it does is, as you say above, make the image crisp. It does this by putting in details that are likely to be there, based on its training data.

It's not just based on the training data. There is a probabilistic factor, but it's based on the mathematical properties of folds in 3D surfaces. That's what the model is trying to infer, and it does work in many cases. You can see how much success can be achieved so far if you look at the papers.

I was talking about image generation; are you talking about protein folding again? If not I don't get what you're referring to.

I don't think anyone should be surprised that the smaller the input image, the less information it contains. So the less likely the upscaled image would be to resemble the depicted object. I would also train a separate model for the much harder task of upscaling microscopic images. I expect the underlying statistical inference tasks are not the same. Applying a normal upscaling model to that task is an interesting experiment, but a misuse from an end user standpoint.

If you're understanding the resolution problem now, that's great. However, maybe it's worth showing mathematically that a Star Trek or Blade Runner "enhance" function cannot interpolate correct information indefinitely.

That image of Obama is 32x32. Let's assume it's using 24-bit color. Then there are 16.7 billion possible images that size. That is every possible 32x32 image: not just pictures of people, pictures of anything.

Take a headshot of everyone in the world in 32x32— that's 8 billion people right there. There is not room in those images to distinguish everyone— some huge number are collapsed together into each picture-of-a-person. (Recall that only a fraction of the 16.7 billiion possible images are headshots of people.)

Now compare these with 128x128 images. There are 16 times as many— 267 billion.

An enhance function can only map each 32x32 image into a new 128x128 image. It's a one-to-one mapping, so it cannot access the entire 267-billion image space— it can only produce one of 16.7 billion images, though they will be nicely crisp due to statistical interpolation. No actual information has been added; all those images that collapsed together still do so. Or to put it another way, if you and your Uncle Anand are both 32x32 image Z, the algorithm cannot separate you out, even though a fresh 128x128 picture would do so. Z is just one image, not two, there is no way to recover the missing information.

(There's a complication in that you can introduce randomness so that the mapping is not one-to-one. So one seed might produce expansion X while another produces expansion Y. This produces variation but it does not produce more accuracy— the algorithm is, in effect, randomly choosing between X and Y, it is not choosing and cannot choose based on the information it has, the bits in the 32x32 image.)

(BTW this all applies to any image sizes; there is not something special about 32x32. All you get with bigger images is a larger starting image space. If you started with 128x128 you'd have 275 trillion possible images; if you then AI-upscaled to 1024x1024, you'd also have 275 trillion images, though the image space for 1024x1024 is 64 times larger.)

[keenir]

Today's fun AI-related link: AI system resorts to blackmail if told it will be removed. Found when looking for something online to complement the report on today's BBC news about Google's AI-generated movies now incorporating speech.

So now AI can produce films complete with dialog and blackmail humans?

wow, you can read link titles. *gasp*

and what conclusions did you draw from the actual article itself?

When will we start taking it seriously instead of dismissing it all as smoke and mirrors?

when you stop cheerleading for AI and Trump in neighboring breaths.

the next step...for whom? the next step by whom? every AI thus far, has been designed by a human, and to perform a specialized task; why would a human design that master AI? (other than to turn your ravings into prophetic words)

You should ask the people trying to create AGI and bring about the Singularity why they have devoted so much effort and resources to such things.

not everyone who throws their pocket change at things, neccessarily believes there will be any success at it.

Whether you believe it or not, there are numerous people obsessed with such things, many of them with billions of dollars at their disposal.

yeah? there were people who did that with aether too. and to get their hands on unicorn horns. and to set fire to the countryside to communicate with Martians.

clearly you don't know microchips generate geat - you stuff chips together, they melt.

so now you're just building castles in the sky, almost literally.

Everyone familiar with computers knows they generate waste heat.

its hard to know what you know / are aware of...and that includes that computer parts generate heat. you don't need a complete computer.

Computer engineers already take that into account and incorporate cooling systems into data centers, hence the controversy over AI using so much water

what controversy?

. Likewise the existence of massive skyscrapers like the Burj Khalifa demonstrates that huge buildings for billions of chips are certainly possible.

the Burj certainly does exist. does that mean that someone will build a Burj just for computer chips? no.

Perhaps a machine with the cognitive abilities of Trump requires multiple city blocks and the power output of a medium sized country.

*snerk*

I've seen limestone race faster in its growth, than AI has been.

Over the past decade, AI has gone from science fiction speculation for the most part to folding proteins, writing novels, drawing pictures, composing music, creating highly convincing video complete with speech, and much more. You can criticize AI for many things, but sluggish growth is certainly not one of them. The average LLM can write better than almost anyone on this forum on pretty much any subject you care to name or even pure fiction while recent AI models can produce full video faster than any human with a camera, yet you dismiss it as limestone growth.

in terms of being adaptable? yes, yes i do. in terms of being hyperspecialized, yes it is fast growth. those are not the same thing.

[Torco]

also, those "burj khalipha but filled to the brim with chips" already exist, except they're not thin and tall but rather thick and low-to-the-ground, they're called data centers, or server farms. the burj is what, 2 million square meters in volume or something? that's a big data center, but not that big: that's three stories of 300m by 300m or something? there's probably A LOT MORE cubic meters dedicated to datacenters in reality.
-

it may be that the name is a bit of a pet peeve of mine, but i think a lot of the reason people misunderstand that-which-is-called-AI is, well, due to the fact that it is called AI. like, LLMS and other generative models are artificial -in the same sense cups or gasoline is artificial- but they are not intelligence. like, they don't work the way intelligence works -as far as we know how intelligence works, which isn't far, it's far enough that we know it is not that.

no level of intelligence corresponds to the capacity to write python code that runs nonparametric dimensionality reduction but at the same being unable to count the Rs in strawberry. there are no intelligences that exhibit that combination of capabilities. by all accounts, generative models are more like a biased autocomplete: complete this sentence in a way that is helpful, complete this sentence in a way that is sexy (in the case of porn chatbots), complete this picture in a way that it features a crocodile headed bomber plane.
-

I am in general skeptical of the concept of "singularity", not least cause, you know, i don't think ayone means anything very specific with it. but let's take the modest notion of the singularity: a moment of expontential growth in *something* about AI such that it causes important disruptions to the economy, society, culture, etcetera. if so, then we're in the middle of it. it doesn't need to be escathologic, it may be something like, i don't know, the industrial revolution, in terms of like... the amount of transformation? let's assume we can quantify that. i think that could easily happen.

then again, for a lot of people, the industrial revolution was kind of apocalyptic. we got antibiotics tho

[rotting bones]

LLMs don't know what they are talking about. In their training data somewhere, they saw being threatened followed by blackmail. They are very expensive parrots.

What should worry you more is that LLMs have been shown to systematically deceive their users when put under pressure.

[rotting bones]

I was talking about image generation; are you talking about protein folding again? If not I don't get what you're referring to.

At least some upscaling models work in this way. Photos of real objects consist of 3D surfaces folded over each other. Based on the consistency of the substance the surface is made of, draping them over each other produces the shape of one mathematical function or other because of various tensile forces. From the training data, the model is trying to guess which function the blurry image is probably trying to depict, and interpolate hi-res details on that basis. This is why models that work the best on photos don't necessarily produce the best results for paintings.

If you're understanding the resolution problem now, that's great. However, maybe it's worth showing mathematically that a Star Trek or Blade Runner "enhance" function cannot interpolate correct information indefinitely.

Defending my posts is not normally my number one priority. I'm only doing it for so long because my post is the topic of the discussion.

I have always said it's not always correct. That's why the inference is "statistical". That it's possible for these models to work at all is fascinating. If you look at the literature, some architectures are better than others at recovering details for specific image types. So the details are not fully random just because they are statistical. The relative image sizes is what makes this task a hard problem. But since the details being inferred are not random, the full space of details that could be inferred is not as devastating a counterpoint as you might think. (Re: the number of pixels.)

Did Star Trek ever claim that Enhance maintains perfect fidelity up to infinite magnification? They were probably just thinking of a navy telescope, right? Algorithmic upscaling, whether it uses sophisticated inference techniques or an old fashioned Lanczos filter, is much more amazing because of the breathtaking results it can obtain despite the depicted object being absent.

[malloc]

no level of intelligence corresponds to the capacity to write python code that runs nonparametric dimensionality reduction but at the same being unable to count the Rs in strawberry. there are no intelligences that exhibit that combination of capabilities. by all accounts, generative models are more like a biased autocomplete: complete this sentence in a way that is helpful, complete this sentence in a way that is sexy (in the case of porn chatbots), complete this picture in a way that it features a crocodile headed bomber plane.

Suppose someone traveled through the jungles of Brazil and found an animal with the same capabilities as current AI models. Imagine if you will a parrot who could compose entire novels or symphonies or a monkey who could draw photorealistic images or a sloth who could fold proteins. We would undoubtedly consider that animal remarkably intelligent. Yet for some reason when computers achieve the same thing, we insist otherwise and focus on all the mistakes they make, as if humans never struggle to count or suffer from hallucinations.

yeah? there were people who did that with aether too. and to get their hands on unicorn horns. and to set fire to the countryside to communicate with Martians.

Except that intelligence demonstrably exists whereas unicorns and Martians are clearly fantasy. We already know that intelligence is possible and that it doesn't require anything particularly magical, just a kilogram of electrolytes flowing through tubes of fat. Unlike faster than light travel or time machines, intelligence contradicts no laws of physics and nothing in theory prevents us from creating machines that replicate it. Given enough time and research, we will eventually figure out AGI.

[zompist]

If you're understanding the resolution problem now, that's great. However, maybe it's worth showing mathematically that a Star Trek or Blade Runner "enhance" function cannot interpolate correct information indefinitely.

Defending my posts is not normally my number one priority. I'm only doing it for so long because my post is the topic of the discussion.

I have always said it's not always correct. That's why the inference is "statistical". That it's possible for these models to work at all is fascinating. If you look at the literature, some architectures are better than others at recovering details for specific image types. So the details are not fully random just because they are statistical. The relative image sizes is what makes this task a hard problem. But since the details being inferred are not random, the full space of details that could be inferred is not as devastating a counterpoint as you might think. (Re: the number of pixels.)

OK, that you are speaking this way says, unfortunately, that you completely missed the point. My argument is not "I don't think upscaling is cool." It's that upscaling is mathematically incapable of interpolating correct data to disambiguate merged pictures.

It doesn't matter what the algorithm is. It doesn't have to be AI; it can be God's Own Algorithm. You cannot map from 16.7 billion possible pictures into anything but 16.7 billion pictures.

And if you and Uncle Anand share the same 32x32 picture, even God's Own Algorithm cannot turn that into two pictures again; that information was lost because of the limits on resolution.

Now, for this merging to happen, you and Uncle Anand have to, in fact, look pretty similar. But since the number of people in the world far exceeds the part of the image space that looks like headshots, this will happen at that image size. It doesn't matter if God's Own Algorithm turns it into a nice picture of you; then it won't be a nice picture of Uncle Anand.

Obviously the problem recedes if you have a hi-res image. But you're fooling yourself if you think that more realistic picture sizes never have this problem. In a picture of a street scene, or in low light, or if it's moving, a face might be effectively as low-res as the examples I'm detailing. (And in fact that's precisely the sort of thing the movies show the computer doing. They never show the image quadrupling in size, as you did; they show it expanding a millionfold.)

Algorithmic upscaling, whether it uses sophisticated inference techniques or an old fashioned Lanczos filter, is much more amazing because of the breathtaking results it can obtain despite the depicted object being absent.

"It's amazing" isn't an argument— we all find some things amazing and some not. All I'm concerned about is not deluding ourselves that we can see things that aren't there.

(My argument applies to photographic images too, because grain size imposes a resolution limit.)

[rotting bones]

Now, for this merging to happen, you and Uncle Anand have to, in fact, look pretty similar. But since the number of people in the world far exceeds the part of the image space that looks like headshots, this will happen at that image size. It doesn't matter if God's Own Algorithm turns it into a nice picture of you; then it won't be a nice picture of Uncle Anand.

You said 32x32 is not special, so I assumed you think your argument generalizes to larger images. It's the generalization that leads to issues. If me and my uncle look like identical twins in a 200x200 image, I agree the model won't be much help in disambiguating them. The issue is that that will not be the case 99.9% of the time. What the model does is try to reconstruct the appearance of physical surfaces from blurry images. This transformation does help human viewers make out what they are looking at most of the time. This procedure has empirical limits like microscopic inputs and very high magnifications. These are not part of the original use case.

I have already written a paragraph on why the full space of possible interpolations may not be as relevant as you think. You didn't engage with that argument. The algorithm doesn't claim to reproduce the original image with 100% fidelity. It makes educated guesses about what the original looked like based on the physical properties of surfaces. This procedure is certain to miss, say, tiny imperfections in the original, but it would still reconstruct a good enough representation in hi-res for a human viewer with acceptable fidelity.

Obviously the problem recedes if you have a hi-res image. But you're fooling yourself if you think that more realistic picture sizes never have this problem. In a picture of a street scene, or in low light, or if it's moving, a face might be effectively as low-res as the examples I'm detailing.

The input image obviously need not be hi-res. It just needs to have some info, which may not be available in a 32x32 image. I disagree that Star Trek's Enhance is shown to find details of images that are in shadow. If that were the case, Enhance could be used to nullify, say, the cloaking device.

(And in fact that's precisely the sort of thing the movies show the computer doing. They never show the image quadrupling in size, as you did; they show it expanding a millionfold.)

zompist, if you made a picture a million times smaller, you won't be able to see it on the screen. Do you realize how large an image gets if you magnify it 10 times? That's entirely within the capability of AI upscaling.

Which scene are you thinking of that turns an invisible half a pixel into a hi-res shot? If the were the case, the Enterprise could probably combine Enhance with advanced telescope lenses to obtain hi-res shots of neighboring solar systems.

I think it's reasonable to say that Enhance expects an input image that is at least a hundred pixels wide. Do you realize how small a 100x100 pixel icon-like spaceship will look on the big screen in the Enterprise? A Lanczos filter will probably be enough to "Enhance" it. The only thing AI adds are reconstructed raytracer-like surfaces.

"It's amazing" isn't an argument— we all find some things amazing and some not. All I'm concerned about is not deluding ourselves that we can see things that aren't there.

It's relevant because that's what I said in my original post. I'm trying to explain why the Obama hallucination doesn't detract from the amazingness.

[keenir]

no level of intelligence corresponds to the capacity to write python code that runs nonparametric dimensionality reduction but at the same being unable to count the Rs in strawberry. there are no intelligences that exhibit that combination of capabilities. by all accounts, generative models are more like a biased autocomplete: complete this sentence in a way that is helpful, complete this sentence in a way that is sexy (in the case of porn chatbots), complete this picture in a way that it features a crocodile headed bomber plane.

Suppose someone traveled through the jungles of Brazil and found an animal with the same capabilities as current AI models. Imagine if you will a parrot who could compose entire novels or symphonies or a monkey who could draw photorealistic images or a sloth who could fold proteins. We would undoubtedly consider that animal remarkably intelligent.

no, we would not (except maybe for you, because you'd be worshipping it).

we would instead call that animal intensely trained...and prosecute its owner for inhumane treatment of the animal.

Yet for some reason when computers achieve the same thing, we insist otherwise and focus on all the mistakes they make, as if humans never struggle to count or suffer from hallucinations.

yeah? there were people who did that with aether too. and to get their hands on unicorn horns. and to set fire to the countryside to communicate with Martians.

Except that intelligence demonstrably exists whereas unicorns and Martians are clearly fantasy.

unicorns exist - not only the oryx and rhino, but also fossil antelopes which possessed single horns atop their heads.

Martians were fully plausible based on the science of the time.


but yet again, you seem more resistant to the idea of analogy and premise, than anyone on the spectrum that I've ever met.

We already know that intelligence is possible and that it doesn't require anything particularly magical,

we know human intelligence exists...even primates and cetaceans and parrots have quantitites of it.

what you're doing, in taking so many different hyperspecialized programs, and claiming that collectively it amounts to human-conquering AI...is no different than if you considered all mammals to be one species and one civilization, and declared, based upon that, that the species would be terraforming Mars and Pluto within the year.

and nothing in theory prevents us from creating machines that replicate it.

except for not knowing how intelligence works in organic critters, and an inability to produce nonspecialized computer programs.

Given enough time and research, we will eventually figure out AGI.

sure, just like we figured out cold fusion and room temperature superconductors and aether and...

[rotting bones]

Imagine if you will a parrot who could compose entire novels or symphonies or a monkey who could draw photorealistic images or a sloth who could fold proteins.

If the parrot could compose shitty novel-like text after the complete works of Shakespeare were read to it, I would say it definitely has an amazing memory if nothing else. But we already knew computers have higher fidelity memory than humans. That's one reason we use them to store our work instead of committing it all to memory.

[malloc]

no, we would not (except maybe for you, because you'd be worshipping it).

we would instead call that animal intensely trained...and prosecute its owner for inhumane treatment of the animal.

You are missing the point of the analogy. The animal doesn't need torture to achieve these abilities but simply has them as part of its nature just as AlphaFold has the ability to fold proteins as part of its programming.

except for not knowing how intelligence works in organic critters, and an inability to produce nonspecialized computer programs.

Sure but we already know much about natural intelligence, that it arises from the brain rather than the heart or an immaterial soul as ancient people imagined, that neurons and neuronal connections play a critical role, and so forth. Nothing in principle prevents us from progressing further and understanding the finer details of neurology and cognition enough to replicate it technology.

Given enough time and research, we will eventually figure out AGI.

sure, just like we figured out cold fusion and room temperature superconductors and aether and...

Except that intelligence demonstrably exists whereas cold fusion and room temperature superconductors are unproven at best while aether is completely discredited. We cannot figure out cold fusion because it probably can't work in the first place. By contrast, intelligence does work and we simply need to figure out the mechanism by which it functions in organic beings and replicate that technologically.

[keenir]

no, we would not (except maybe for you, because you'd be worshipping it).

we would instead call that animal intensely trained...and prosecute its owner for inhumane treatment of the animal.

You are missing the point of the analogy.

I'm surprised you know what one is.

except for not knowing how intelligence works in organic critters, and an inability to produce nonspecialized computer programs.

Sure but we already know much about natural intelligence, that it arises from the brain rather than the heart or an immaterial soul as ancient people imagined, that neurons and neuronal connections play a critical role, and so forth. Nothing in principle prevents us from progressing further and understanding the finer details of neurology and cognition enough to replicate it technology.

except all the things WE HAVE ALREADY TOLD TO A POSTER NAMED MALLOC.

Given enough time and research, we will eventually figure out AGI.

sure, just like we figured out cold fusion and room temperature superconductors and aether and...

Except that intelligence demonstrably exists whereas cold fusion and room temperature superconductors are unproven at best while aether is completely discredited. We cannot figure out cold fusion because it probably can't work in the first place. By contrast, intelligence does work.

human intelligence does work. parrot intelligence does work. machine intelligence does not work. if you program a machine to do a task, and it does the task, thats your intelligence at work.

if my pet hydra knows it will be rewarded when it completes a maze or labrynth...it can stand in the entrance of that maze, and stick one neck down each corridor...and one head will emerge out the other end of the maze. does it need to be intelligent to accomplish this? no, no more than Sonic Hedgehog needs to be able to read signs, in order to race down every single fork in the maze in the second movie by his name,

[malloc]

human intelligence does work. parrot intelligence does work. machine intelligence does not work. if you program a machine to do a task, and it does the task, thats your intelligence at work.

But we have machines that can match or even exceed humans at many cognitive tasks. Even the best chess player cannot beat a chess engine and image generators can produce decent imitations of artwork. If we discovered a wild animal that effortlessly defeated even Magnus Carlson at chess every time, we would consider that animal amazingly smart, at least when it comes to chess. Yet for some reason we refuse the same courtesy to computers with the same capabilities. Nobody is claiming that AlphaFold or chatGPT have general intelligence, but you must concede that they are quite intelligent within their respective fields.