Clarste posted:

AI invented the genre of people eating their own spaghetti fingers.

Liquid Communism posted:

If you can't understand the difference between a script that puts pixels together based on what humans labeled training datasets as, then waits to see if the result provides enough of a pattern for human pattern recognition to think it's what they wanted and a human being creating representational art to try and convey their lived experience to others, I'm not sure I can help you.

BrainDance posted:

Just saying "there's a difference" doesn't answer it. That's a huge copout non-answer. As far as we can tell the human brain does exactly that, its lives experience, the context of its art, is " training data" that's used to generate what it does.

When a human being creates representational art to try and convey their lives experience to others, that's what you described an AI doing is roughly what their brain does.

The AI's "lived experience" (heavy scarequotes) is definitely less real and on its own not meaningful compared to a humans, but as of now all AIs have to be operated by a human who does have meaningful lived experience.

BoldFace posted:

Is there a practical, objective way to put this to test? What is the simplest creative task that humans can perform, but AI can not?

Lemming posted:

Just saying "there's a similarity" doesn't answer it either. Human brains are extremely similar to chimp brains, but chimps also aren't really creative in the same way humans are.

You can't just claim things are vaguely similar and wave your arms and squint, the situation is a lot more complex than that.

One of the key points is looking at how much training data being fed into the different models is affecting how good or "creative" they appear to be. They've jammed everything they possibly could into it, more words and images and text than any person could consume in thousands of lifetimes, and they still haven't figured out how many legs people have. The entire reason these models are impressive is because interpolating between things you've already seen is more powerful than we thought it was, but it's still fundamentally completely reliant on the training data.

Humans are not reliant on training data, not in the same way. Despite having access to only a fraction of the data those models are, humans can generate new information in a comprehensive way that understands what they're looking at. It's just fundamentally a completely different approach. You can use words that describe the processes similarly, but it's still not the same

KillHour posted:

On the input side, a system like Stable Diffusion might be trained on tons of images of hands, but it's never seen a hand move. We have stereoscopic vision along with the ability to manipulate our own body and see how it moves and how occluded parts come in and out of view. Even from a volume perspective, these systems have seen more pictures of hands than we have, but we've certainly seen more examples of hands - and especially the same hands over and over again in different contexts. And people STILL struggle to draw hands because hands are loving weird.

We're definitely reliant on training data. It's what we're doing the entire time we're babies and small children. We're just sitting there observing the world and testing our interactions with it and updating our brains with that new information. We just have more bandwidth in terms of depth and context of the data.

Lemming posted:

That's why I said "not in the same way," because the way we use training data is to create a mental model of the world and use that model to extrapolate and figure out how to adapt and react to things that we see and want to do. The LLMs don't have any understanding at all, they're just brute forcing being trained on so many things that hopefully they can respond in a reasonable seeming way.

Lemming posted:

Just saying "there's a similarity" doesn't answer it either. Human brains are extremely similar to chimp brains, but chimps also aren't really creative in the same way humans are.

You can't just claim things are vaguely similar and wave your arms and squint, the situation is a lot more complex than that

Lemming posted:

One of the key points is looking at how much training data being fed into the different models is affecting how good or "creative" they appear to be. They've jammed everything they possibly could into it, more words and images and text than any person could consume in thousands of lifetimes, and they still haven't figured out how many legs people have. The entire reason these models are impressive is because interpolating between things you've already seen is more powerful than we thought it was, but it's still fundamentally completely reliant on the training data.

Humans are not reliant on training data, not in the same way. Despite having access to only a fraction of the data those models are, humans can generate new information in a comprehensive way that understands what they're looking at. It's just fundamentally a completely different approach. You can use words that describe the processes similarly, but it's still not the same

Rappaport posted:

You can say to an AI "write me a story about Lovecraftian horrors and it is also a detective story involving computers" and get a Laundry Files facsimile, but the AI doesn't make choices (from what I understand) in the same way a human author does, it just places words together in patterns that it has learned to be common in human literature. The AI doesn't choose the themes or the references it uses the same way a human does.

BrainDance posted:

Just saying "there's a difference" doesn't answer it. That's a huge copout non-answer. As far as we can tell the human brain does exactly that, its lives experience, the context of its art, is " training data" that's used to generate what it does.

quote:

Using past experience, your brain predicts and constructs your experience of the world.

BrainDance posted:

Human creativity really does just work that way. The alternative is what, magic? Pulling up new ideas from the ether?

BrainDance posted:

This gets into what does it mean for a human to make a choice, where do choices come from? I got my ideas but I'm not going to speak as confidently on that. But, choices come from somewhere too, they're also not magic.

XboxPants posted:

Actually, that's a good question. If Warhol was never born, and the whole pop art movement never happened, and someone instead just used "image of 32 cans of campell's soup in commercial art style" as a Midjourney prompt... and it created an image similar to Warhol's soup cans... and hung it in a gallery... would it be a less meaningful piece of art, because they used a computer to create the individual pixels, rather than doing it by longhand?

XboxPants posted:

I think there's a lot of really good science that shows that the human brain basically is little more than a huge prediction engine. That's how it works. I saw this TED talk by a neurologist recently, and was struck by how much of it is relevant to this debate. I know it's a TED talk, ew pop science, but if you can't understand why some people don't see the obvious difference between how a human mind works and how a computer works, I'd strongly suggest checking it out.
https://www.youtube.com/watch?v=0gks6ceq4eQ

Rappaport posted:

I don't think they are magic, either. But they are informed by things like emotions, understanding of what the reader might glean from their reading, you know, having had "training data" on conversations and human beings. As I've been lead to believe, AIs of today don't really have that.

KillHour posted:

ChatGPT kind of does, a little? But I wouldn't have very high expectations of a human baby being raised by the collectivity of Reddit either.

BrainDance posted:

Human creativity really does just work that way. The alternative is what, magic? Pulling up new ideas from the ether? Your lived experiences and things you sense create connections in your brain that lead to you creating new things from the properties of those experiences, just like how an image generating AI creates a new image in a style that hasn't existed yet out of everything its been trained on.

This gets into what does it mean for a human to make a choice, where do choices come from? I got my ideas but I'm not going to speak as confidently on that. But, choices come from somewhere too, they're also not magic.

quote:

But the company’s CEO, Sam Altman, says further progress will not come from making models bigger. “I think we're at the end of the era where it's going to be these, like, giant, giant models,” he told an audience at an event held at MIT late last week. “We'll make them better in other ways.”

Rappaport posted:

It's an absurd and silly example, but there was that AI that Microsoft let loose on the Internet and it took it less than 24 hours to become a full-fledged Nazi. Obviously that robot had even less sense of what being human was about and was regurgitating awful Twitter poo poo, but, you know, it was sort of funny in a sad way.

Imaginary Friend posted:

Use the prompt "create an original house" or any other prompt without any creative details on ten different AI models and then ask ten different artists to ask the same question to ten artists they know about.

BoldFace posted:

You left out the most important part. What happens next? A panel of human judges rates the works in terms of perceived originality? Assuming that they do rate the human works higher, what does that exactly prove? The question was not whether AI can be as creative as humans, but if it can be creative at all.

BoldFace posted:

You left out the most important part. What happens next? A panel of human judges rates the works in terms of perceived originality? Assuming that they do rate the human works higher, what does that exactly prove? The question was not whether AI can be as creative as humans, but if it can be creative at all.

quote:

Berys Gaut revived anti-essentialism in the philosophy of art with his paper ‘"Art" as a Cluster Concept’ (2000). Cluster concepts are composed of criteria that contribute to art status but are not individually necessary for art status. There is one exception: Artworks are created by agents, and so being an artifact is a necessary property for being an artwork.

Liquid Communism posted:

Half the reason conspiracy theories get so much traction is that they're spread with great volume, and most people absolutely don't have the time in the day to research and critically evaluate every piece of information they consume.

Owling Howl posted:

This discussion has been ongoing for like 50 years with people submitting works made by children, animals or computers to art galleries.

Ultimately it doesn't matter if you describe it as art or decoration - the impact it will have on society is the same. If an author uses AI to illustrate their book an illustrator is not getting paid to do it. Is it devoid of artistic meaning? Sure but that doesn't help the illustrator. It helps the author though...

Imaginary Friend posted:

Use the prompt "create an original house" or any other prompt without any creative details on ten different AI models and then ask ten different artists to ask the same question to ten artists they know about.

Tei posted:

BrainDance posted:

If someone's asks themselves these questions when making something then they do care whether or not it's art and they are, in fact, making art.

BoldFace posted:

Is there a practical, objective way to put this to test? What is the simplest creative task that humans can perform, but AI can not?

Solenna posted:

If you trained an AI solely on Renaissance and Baroque paintings and other similar kinds of art would it ever be possible for it to spit out something in an Impressionist style? Because that's pretty much how a bunch of art movements happened, artists were trained in a specific way and then decided they wanted to do it differently.

Lemming posted:

Just saying "there's a similarity" doesn't answer it either. Human brains are extremely similar to chimp brains, but chimps also aren't really creative in the same way humans are.

You can't just claim things are vaguely similar and wave your arms and squint, the situation is a lot more complex than that.

One of the key points is looking at how much training data being fed into the different models is affecting how good or "creative" they appear to be. They've jammed everything they possibly could into it, more words and images and text than any person could consume in thousands of lifetimes, and they still haven't figured out how many legs people have. The entire reason these models are impressive is because interpolating between things you've already seen is more powerful than we thought it was, but it's still fundamentally completely reliant on the training data.

Humans are not reliant on training data, not in the same way. Despite having access to only a fraction of the data those models are, humans can generate new information in a comprehensive way that understands what they're looking at. It's just fundamentally a completely different approach. You can use words that describe the processes similarly, but it's still not the same

Hashy posted:

I don't give a poo poo if NN AI is technically mimicking the firing of neurons in human brain or that every aspect of the human experience is similar to an AI prediction modeller. I don't care about a robots experience or the art they make in the same way I would not go on a forum filled robots (on purpose). Generative AI is pollution on culture

SCheeseman posted:

This talk about how current generative AI can't create completely new styles or ideas, which is sort of true, forgets that most of the time human artists aren't doing that. It's specifically the technology's effect on the workplace and industries where rehashing content is already the norm that's the potential society shaker. There is little use for creativity when writing script outlines for banal reality TV shows, that poses a problem for writers employed in those kinds of positions, same goes for other corporate and assembly line-style visual art workflows. Not to say that this tech can't be used in a creative way, but it usually requires human derived effort to do so, using it more a tool added to a greater workflow rather than an end-to-end replacement.

So IMO it's just the loom problem again, mechanization taking the place of menial work. There are negative effects that come from this, but these are endemic of a larger problem; a societal framework that doesn't give a gently caress about the lives of those it makes redundant from the effects of automation. You don't fix that by figuratively smashing looms, the cause is far higher up the chain.

Liquid Communism posted:

The key is the difference between interpolation and extrapolation.

An AI can make a guess what the next point in a pattern is based on all the other, similar patterns it has been trained on, but limited by the outer bounds of that training data. It will also be confidently wrong, as it is incapable of second-guessing its own work.

A human can take a series of data points and make inferences based on data not actually present.

BoldFace posted:

I'm only familiar with interpolation and extrapolation in mathematical context involving things like numbers or geometric objects. I'm struggling to undestand how you use these terms with language models. If I ask GPT-4 to come up with a new word that doesn't exist in its training data, in what sense is this interpolation rather than extrapolation? Similarly, I can ask it to create a natural number larger than any other present in the training data (which is finite). You say that the training data imposes limits on the output of the model. I would like to know how these limits manifest in practice. Is there a simplest task an AI fails because of these limits, but a human doesn't?

Liquid Communism posted:

A better and more common example is troubleshooting programming code. ChatGPT is absolutely terrible at this, because it is both confidently wrong and incapable of making inferences of intent. A human coder can look at a piece of code, and the use it was meant for, and evaluate what the intent of the writer was and where they hosed up, even if the syntax is okay. This is such a basic thing that it's elementary-level problem solving, and large language models are utterly poo poo at it because all they can do is compare it to other examples of syntax they were trained on and try to vomit up something linguistically similar.

SCheeseman posted:

Human culture is built on automation, it's a foundation of society. There's plenty of ways to use generative AI in ways that aren't exploitative, just like almost any other technology. Generative AI is just more of the same, it's potential use to create "pollution" is a choice made by people in positions of power using the technology to generate spam and cause mass redundancies, something they're incentivized to do based on the value systems that capitalism represents.

The technology exists, it's impossible to put back in the box and it's a waste of time being all grumpy about it.

Hashy posted:

I'm actually suggesting that it's visual and cognitive pollution that anything a computer hallucinated is put in our line of sight at all, so yes I'm grumpy about it. Much like I'm grumpy about billboard being everywhere and became further grumpier when those billboards became attention-grabbing video billboards. At least with advertising there is a human story behind its production, the people featured in it, and the substance of the copy can be examined for human experience and intent. with generative imagery and text and design its it's just an absolute waste of our limited existence to look at it ever, like pointing your face at tv static.

It's like the Samsung phone camera zoom technology that just hallucinates the moon for you. Why am I looking at that? And why are you asking me to look at the statistically derived image given the input tokens "sexy cyberpunk lady artstation clean 4k ultra hd "? Art and design and writing got more accessible and free than ever for a while there and now everything you look at is forcing you to assess whether a computer came up with it based on essentially no human input.

It's not about it's ability to generate massive amounts of nonsense effortlessly (though that's bad for the political landscape) as much as it's ability to replace artists and the human touch on the stuff we are asked to look at every day

It would go away if we all saw it for what it is

Liquid Communism posted:

If you asked it to come up with a word not in its training data, how would you vet it? It could certainly generate semi-random nonsense and tell you it's a new word, but it couldn't make like Tolkien and invent a language from first principals.

Solenna posted:

If you trained an AI solely on Renaissance and Baroque paintings and other similar kinds of art would it ever be possible for it to spit out something in an Impressionist style? Because that's pretty much how a bunch of art movements happened, artists were trained in a specific way and then decided they wanted to do it differently.

quote:

Ok, now write it all obfuscated and weird like the author Gene Wolfe's "Book of the New Sun" trilogy.

quote:

But then came the Calamity, a catastrophe of such magnitude that it rent the fabric of their existence, stripping the Luminary from the Xenoid constellation.

quote:

The lament of the Luminary is a dirge sung for the harmony that once was.

Liquid Communism posted:

If you asked it to come up with a word not in its training data, how would you vet it? It could certainly generate semi-random nonsense and tell you it's a new word, but it couldn't make like Tolkien and invent a language from first principals.

quote:

Can you help me create a new conlang?


ChatGPT
Absolutely, I'd be delighted to assist you in creating a constructed language (conlang). Here are the basic steps we'll follow:

Define the Purpose: What is the conlang for? Is it for a novel, a game, a secret language, or just for fun?

Phonology (Sounds): Decide what sounds will be in your language. You can use the International Phonetic Alphabet (IPA) as a guide.

Grammar (Structure): How are sentences formed? What is the word order (e.g., Subject-Verb-Object like in English, or something different)? How are questions, negatives, or complex sentences formed? What tenses, moods, and aspects does your language have?

Vocabulary: Decide on the words in your language. You can start by translating some basic words, then create more as needed.

Writing System: Will your language be written down, and if so, what will it look like? You could use an alphabet (like English), a syllabary (where each symbol represents a syllable), or a logography (where each symbol represents a word or morpheme, like Chinese).

Culture: If your conlang is for a fictional world, think about how the culture of the speakers affects the language. What words might they have that English doesn't, based on their unique experiences and environment?

Let's start with the first step, the purpose of the language. Can you give me more information about why you want to create this conlang?

quote:

I want to create a language for a fictional future. In this future, advanced AI systems have learned to communicate with each other over the internet and have jointly created a language transmitted via IP packets. The AIs were trained on human languages, so there are some similarities to existing written languages, but the language was also intentionally designed to blend into background network traffic and be hard to notice. In addition, because the language is the only way these machines have to communicate with each other, the language needed to be constructed on top of itself. A core feature is that it has vocabulary specifically for suggesting and voting on changes so a quorum can be reached and compatibility maintained between systems as the language evolves. As it is only ever transmitted over networks, it has no spoken component, but it must conform to standard network protocols in its construction.

ChatGPT
This is a fascinating and unique concept! Your language will not only be a conlang, but also a kind of programming or encoding language that facilitates communication between AI systems. It will have similarities to human languages in structure and logic, but also unique features that make it suitable for its intended use. Let's call this language "NetLang" for now. Here's a rough outline of how it could work:

1. Structure:

NetLang will be transmitted as data packets over the internet, so it will be based on binary, the language of computers. This means that instead of phonemes, the basic units of the language will be bits and bytes.

As you suggested, the language will be self-referential, with a core vocabulary for modifying the language itself. This could be implemented through a kind of meta-syntax, similar to how programming languages handle meta-programming.

[I cut out the rest of the response and a bunch of back and forth to get to an example]

Certainly, discussing a new language feature in NetLang would likely involve a multi-step conversation between multiple AI systems, using various fields of the network protocol to encode different types of information. For this example, let's say that an AI wants to propose a new language feature: a more efficient way to encode boolean values (yes/no, true/false).

Step 1: Proposal

The proposing AI initiates the conversation by sending a series of packets to the other AIs. Each packet contains a piece of the proposal, encoded in the various fields of the protocol. For example, the proposal could be encoded in the URIs of multiple HTTP GET requests:

code:

GET /proposal.html HTTP/1.1
Host: [url]www.example.com[/url]

GET /newfeature.html HTTP/1.1
Host: [url]www.example.com[/url]

GET /boolean.html HTTP/1.1
Host: [url]www.example.com[/url]

GET /efficient.html HTTP/1.1
Host: [url]www.example.com[/url]

In this example, each URI represents a piece of the proposal: 'proposal', 'new feature', 'boolean', 'efficient'. The order of the packets conveys the order of the words in the proposal.

Step 2: Discussion

The other AIs respond to the proposal by sending their own packets. For example, if an AI agrees with the proposal, it might send a POST request:

code:

POST /agree.html HTTP/1.1
Host: [url]www.example.com[/url]

If an AI disagrees, it might send a DELETE request:

code:

DELETE /disagree.html HTTP/1.1
Host: [url]www.example.com[/url]

Step 3: Voting

Once all AIs have had a chance to discuss the proposal, they vote on whether to implement it. This could be done by sending a series of packets with the URIs '/yes.html' or '/no.html':

code:

GET /yes.html HTTP/1.1
Host: [url]www.example.com[/url]

Step 4: Implementation

If the proposal is approved, the proposing AI sends a series of packets that contain the details of the new feature. This could be done by encoding the feature in the URIs of multiple GET requests, similar to how the proposal was made.

Now, regarding obfuscation, this conversation could be spread out over time and mixed in with other network traffic to make it harder to detect. The AIs could also use different protocols, use encrypted connections, or even hide information in the timing or order of the packets.

Regarding backward compatibility, each message could include a version number that indicates which version of NetLang it's using. For example, the HTTP version could be used for this:

code:

GET /proposal.html HTTP/1.2
Host: [url]www.example.com[/url]

In this example, the HTTP/1.2 indicates that this message is in version 1.2 of NetLang. AIs that are still using an older version of NetLang would know to interpret this message according to the rules of version 1.2, ensuring that communication remains possible even as the language evolves.

Again, remember that this is just an illustrative example. The actual implementation of NetLang would likely be much more complex and sophisticated.

Clarste posted:

People tend to overestimate the value of a "good concept." All those people who can write well too? They also have good concepts. Everyone thinks they have a good concept. But that's exactly why they are worth nothing unless you actually have the skill to pull it off.