The brain does not work the way you think… (I work in the field, bio-informatics). What you call “neural networks” come from an early misunderstanding of how the brain stores information. It’s a LOT more complicated and frankly, barely understood.
It’s a LOT more complicated and frankly, barely understood.
Yet you confidently state that the brain doesn’t work the way LLMs do?
Obviously it doesn’t work exactly the same way that LLMs do, if only because of the completely different substrates. But when you get to more nebulous concepts like “creativity” and “inspiration” it’s not so clear.
The part where brain and neural net differ is in the learning via backpropagation, that seem to be done different in the brain, as there is no mechanism to go backwards through the network and jiggle the weights.
That aside, they seem to work very similar once they are trained, as the knowledge they are able to extract from data ends up being basically the same that a human would be able to extract. There is surprisingly little weirdness in AI and a surprising amount of human-like capabilities.
Wikipedia: In copyright law, a derivative work is an expressive creation that includes major copyrightable elements of a first, previously created original work.
I think you may be off a bit on what a derivative work is. I don’t see LLMs spouting out major copyrightable elements of books. They can give a summary sure, but Cliff Notes would like to have a word if you think that’s copyright infringement.
I would be, and I don’t understand why you think this would be a problem. I wouldn’t want the government to be preventing activities that there weren’t any actual laws prohibiting.
That’s an interesting take, I didn’t know software could be inspired by other people’s works. And here I thought software just did exactly as it’s instructed to do. These are language models. They were given data to train those models. Did they pay for the data that they used to train for it, or did they scrub the internet and steal all these books along with everything everyone else has said?
Saying software can be inspired is like saying a rock can feel pain.
The rock doesn’t do anything similar to pain. The LLM on the other side does a lot of things similar to inspiration. I can give the LLM a very trivial question and it will answer with a mountain of text. Did my question or the books it was trained on “inspire” the LLM to write that? Maybe, depends of course how far reaching you want to define the word. But either way, the LLM produced something by itself, that was neither a copy of my prompt nor the training data.
Software can do a lot of things that rocks can’t do, that’s not a good analogy.
Whether software can feel “pain” depends a lot on your definitions, but I think there are circumstances in which software can be said to feel pain. Simple worms can sense painful stimuli and react to it, a program can do the same thing.
We’ve reached the point where the simplistic prejudices about artificial intelligence common in science fiction are no longer useful guidelines for talking about real artificial intelligence. Sci-fi writers have long assumed that AIs couldn’t create art and now it turns out it’s one of the things they’re actually rather good at.
AIs in their training stages are simply just running extreme statistical analysis on the input material. They’re not “learning” they’re not “inspired” they’re not “understanding”
The anthropomorphism of these models is a major problem. They are not human, they don’t learn like humans.
The anthropomorphism of these models is a major problem.
People attributing any kind of person hood or sentience is certainly a problem, the models are fundamentally not capable of that (no loops, no hidden thought). At least for now. However what you are doing isn’t really much better, just utterly wrong in the opposite direction.
Those models are very definitely do “learn” and “understand” by every definition of the word. Simply playing around with that will quickly show that and it’s baffling that anybody would try to claim otherwise. Yes, there are limits to what they can understand and there are plenty things that they can’t do, but the amount of questions they can answer goes far beyond what is directly in the training data. Heck, even the fact that they hallucinate is proof that they understand, since it would be impossible to make completely plausible, but incorrect, stuff up without having a deep understanding of the topics. Also humans make mistakes too and they’ll also make stuff up, so this isn’t even anything AI specific.
Hallucinations happen when there’s gaps in the training data and it’s just statistically picking what’s most likely to be next. It becomes incomprehensible when the model breaks down and doesn’t know where to go. However, the model doesn’t see a difference between hallucinating nonsense and a coherent sentence. They’re exactly the same to the model.
The model does not learn or understand anything. It statistically knows what the next word is. It doesn’t need to have seen something before to know that. It doesn’t understand what it’s outputting, it’s just outputting a long string that is gibberish to it.
I have formal training in AI and 90%+ of what I see people claiming AI can do is a complete misunderstanding of the tech.
It doesn’t understand what it’s outputting, it’s just outputting a long string that is gibberish to it.
Which is obviously nonsense, as I can ask it questions about its output. It can find mistakes in its own output and all that. It obviously understands what it is doing.
They weren’t given data. They were shown data then the company spent tens of millions of dollars on cpu time to do statistical analysis of the data shown.
And here I thought software just did exactly as it’s instructed to do.
AI isn’t software. Everything the AI knows is from the books. There is no human instructing the AI what to do. All the human does is build the scaffolding to let the AI learn, everything else is in the data.
Are you saying the writers of these programs have read all these books, and were inspired by them so much they wrote millions of books? And all this software is doing is outputting the result of someone being inspired by other books?
Clearly not. He’s saying that other authors have done the same as the software does. The software creators implemented the same principle into their llm. You are being daft on purpose.
It’s not the same principle. Large language models aren’t ‘inspired’ to write new works. Software can’t be inspired. It follows instructions. Even though large language models might feel like somebody is talking back to you and giving you new information, it’s just code following instructions designed to predict output based on the input provided and the data supplied. There’s no inspiration to be had, and to attribute inspiration to language models is a huge mischaracterization of what’s happening under the hood. Can a language model, without being told what to do, actually use any of the data it was fed to create something? No. Every single large language model requires some sort of input from a user to act as a seed before any sort of response can begin.
This is why it’s so stupid to call this shit AI, because people start thinking it’s actual intelligence. Really, It’s just a fancy illusion.
It is using the term as defined. Maybe stop being a stupid parrot just repeating crap you heard else where and use your brain for a moment. I am losing hope that humans are capable of thought reading all this junk.
They purchased their books to get inspiration from, the original author gets paid, and the author consented to selling it. That’s the difference.
Also the LLM can post entire snippets or chapters of books, which of course you’ll take at face value even if it hallucinates and makes the author look like a worse author then they are.
Generally they probably bought the books they read though.
If George RR Martin torrented Tolkien, wouldn’t he be infringing on the copyright no matter how he subsequently incorporated it into future output?
I completely agree that the training as infringement argument is ludicrous.
But OpenAI exposed themselves to IP infringement by sailing the high seas in how they obtained the works in the first place.
I hate that the world we live in is one where so much data is gated behind paywalls, but the law is what it is, and if the government was going to come down hard on Aaron Swartz for trying to bypass paywalls for massive amounts of written text, it’s not exactly fair if there’s a double standard for OpenAI doing the same thing in an even more closed fashion.
But yes, the degree of entitled focus on the premise of training an AI as equivalent of infringing is weird as heck to see from authors drawing quite clearly from earlier works in their own output.
I have to assume that openAI also paid for the books. if yes then i consider it the same as me reciting passages from memory or coming up with derivative text.
if no, then by all means, go after them and any model trainer for the cost of one book.
Asking an LLM to recite an entire novel isn’t even vaguely a thing yet.
Well, here’s straight from one of the suits against them:
“The OpenAI Books2 dataset can be estimated to contain about 294,000 titles. The only ‘internet-based books corpora’ that have ever offered that much material are notorious ‘shadow library’ websites like Library Genesis (aka LibGen), Z-Library (aka B-ok), Sci-Hub, and Bibliotik. The books aggregated by these websites have also been available in bulk via torrent systems.”
I’m not even sure how they would have logistically gone about purchasing 294,000 books in bulk in digital form to be fed into training. Using the existing collections seems much more likely, but I suppose we’ll see what turns up in litigation.
Also, the penalty for downloading copyrighted material if willful infringement is up to $250,000 per work. So it’s quite a bit more than the cost of one book on the line…
God that Aaron/jstor thing makes me see red every time. Swartz was scraping jstor to publish it for the benefit of everyone, openai is doing it to make billions of dollars. Don’t forget who the bad guys are (and donate to sci-hub)
I certainly hope that none of these authors have ever read a book before or have been inspired by something written by another author.
That would be a much better comparison if it was artificial intelligence, but these are just reinforcement learning models. They do not get inspired.
More to the point: they replicate patterns of words.
So do humans.
That’s a Bingo!
…like the naturally occuring neural networks are.
The brain does not work the way you think… (I work in the field, bio-informatics). What you call “neural networks” come from an early misunderstanding of how the brain stores information. It’s a LOT more complicated and frankly, barely understood.
Yeah, accurately simulating a single pyramidal neuron requires an eight-layer deep neural network:
https://www.cell.com/neuron/pdf/S0896-6273(21)00501-8.pdf
that was an interesting read, thank you
Yet you confidently state that the brain doesn’t work the way LLMs do?
Obviously it doesn’t work exactly the same way that LLMs do, if only because of the completely different substrates. But when you get to more nebulous concepts like “creativity” and “inspiration” it’s not so clear.
The part where brain and neural net differ is in the learning via backpropagation, that seem to be done different in the brain, as there is no mechanism to go backwards through the network and jiggle the weights.
That aside, they seem to work very similar once they are trained, as the knowledge they are able to extract from data ends up being basically the same that a human would be able to extract. There is surprisingly little weirdness in AI and a surprising amount of human-like capabilities.
people have a definite fear of being defined as machines… not sure why we think were so special…
so its barely understood, but this definitely is not it. got it.
But you, random stranger on the internet, knows better than the guy that literally works in the field. Got it.
i do? where did i claim that?
Tell you what, you get a landmark legal decision classifying LLM as people and then we’ll talk.
Until then it’s software being fed content in a way not permitted by its license i.e. the makers of that software committing copyright infringement.
What exactly was not permitted by the license? Reading?
Using it to (create a tool to) create derivatives of the work on a massive scale.
Wikipedia: In copyright law, a derivative work is an expressive creation that includes major copyrightable elements of a first, previously created original work.
I think you may be off a bit on what a derivative work is. I don’t see LLMs spouting out major copyrightable elements of books. They can give a summary sure, but Cliff Notes would like to have a word if you think that’s copyright infringement.
An AI model is not a derivative work. It does not contain the copyrighted expression, just information about the copyrighted expression.
Better tell that Google and their search index, book scanning project and knowledge graph.
I didn’t know those were LLMs, TIL.
Well when that happens we have laws. So no problems
Would you be okay with applying that argument for any crime?
I would be, and I don’t understand why you think this would be a problem. I wouldn’t want the government to be preventing activities that there weren’t any actual laws prohibiting.
Ever heard of the early 21st century classic Minority Report
That’s an interesting take, I didn’t know software could be inspired by other people’s works. And here I thought software just did exactly as it’s instructed to do. These are language models. They were given data to train those models. Did they pay for the data that they used to train for it, or did they scrub the internet and steal all these books along with everything everyone else has said?
Well, now you know; software can be inspired by other people’s works. That’s what AIs are instructed to do during their training phase.
Does that mean software can also be afraid, or angry? What about happy software? Saying software can be inspired is like saying a rock can feel pain.
If it is programmed/trained that way, sure. I recommend having a listen to Geoffrey Hinton on the topic (41:50).
The rock doesn’t do anything similar to pain. The LLM on the other side does a lot of things similar to inspiration. I can give the LLM a very trivial question and it will answer with a mountain of text. Did my question or the books it was trained on “inspire” the LLM to write that? Maybe, depends of course how far reaching you want to define the word. But either way, the LLM produced something by itself, that was neither a copy of my prompt nor the training data.
Here is an alternative Piped link(s):
Geoffrey Hinton on the topic
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source; check me out at GitHub.
Software can do a lot of things that rocks can’t do, that’s not a good analogy.
Whether software can feel “pain” depends a lot on your definitions, but I think there are circumstances in which software can be said to feel pain. Simple worms can sense painful stimuli and react to it, a program can do the same thing.
We’ve reached the point where the simplistic prejudices about artificial intelligence common in science fiction are no longer useful guidelines for talking about real artificial intelligence. Sci-fi writers have long assumed that AIs couldn’t create art and now it turns out it’s one of the things they’re actually rather good at.
Software cannot be “inspired”
AIs in their training stages are simply just running extreme statistical analysis on the input material. They’re not “learning” they’re not “inspired” they’re not “understanding”
The anthropomorphism of these models is a major problem. They are not human, they don’t learn like humans.
People attributing any kind of person hood or sentience is certainly a problem, the models are fundamentally not capable of that (no loops, no hidden thought). At least for now. However what you are doing isn’t really much better, just utterly wrong in the opposite direction.
Those models are very definitely do “learn” and “understand” by every definition of the word. Simply playing around with that will quickly show that and it’s baffling that anybody would try to claim otherwise. Yes, there are limits to what they can understand and there are plenty things that they can’t do, but the amount of questions they can answer goes far beyond what is directly in the training data. Heck, even the fact that they hallucinate is proof that they understand, since it would be impossible to make completely plausible, but incorrect, stuff up without having a deep understanding of the topics. Also humans make mistakes too and they’ll also make stuff up, so this isn’t even anything AI specific.
Yeah, that’s just flat out wrong
Hallucinations happen when there’s gaps in the training data and it’s just statistically picking what’s most likely to be next. It becomes incomprehensible when the model breaks down and doesn’t know where to go. However, the model doesn’t see a difference between hallucinating nonsense and a coherent sentence. They’re exactly the same to the model.
The model does not learn or understand anything. It statistically knows what the next word is. It doesn’t need to have seen something before to know that. It doesn’t understand what it’s outputting, it’s just outputting a long string that is gibberish to it.
I have formal training in AI and 90%+ of what I see people claiming AI can do is a complete misunderstanding of the tech.
Than why do you keep talking such bullshit? You sound like you never even tried ChatGPT.
Yes, that’s understanding. What do you think your brain does differently? Please define whatever weird definition you have of “understand”.
You are aware of Emergent World Representations? Or have a listen to what Ilya Sutskever has to say on the topic, one of the people behind GPT-4 and AlexNet.
Which is obviously nonsense, as I can ask it questions about its output. It can find mistakes in its own output and all that. It obviously understands what it is doing.
They weren’t given data. They were shown data then the company spent tens of millions of dollars on cpu time to do statistical analysis of the data shown.
A computer being shown data is a computer being given data. I don’t understand your argument.
The data is gone by the time a user interacts with the AI. ChatGPT has no access to any books.
AI isn’t software. Everything the AI knows is from the books. There is no human instructing the AI what to do. All the human does is build the scaffolding to let the AI learn, everything else is in the data.
deleted by creator
Are you saying the writers of these programs have read all these books, and were inspired by them so much they wrote millions of books? And all this software is doing is outputting the result of someone being inspired by other books?
Clearly not. He’s saying that other authors have done the same as the software does. The software creators implemented the same principle into their llm. You are being daft on purpose.
It’s not the same principle. Large language models aren’t ‘inspired’ to write new works. Software can’t be inspired. It follows instructions. Even though large language models might feel like somebody is talking back to you and giving you new information, it’s just code following instructions designed to predict output based on the input provided and the data supplied. There’s no inspiration to be had, and to attribute inspiration to language models is a huge mischaracterization of what’s happening under the hood. Can a language model, without being told what to do, actually use any of the data it was fed to create something? No. Every single large language model requires some sort of input from a user to act as a seed before any sort of response can begin.
This is why it’s so stupid to call this shit AI, because people start thinking it’s actual intelligence. Really, It’s just a fancy illusion.
It is using the term as defined. Maybe stop being a stupid parrot just repeating crap you heard else where and use your brain for a moment. I am losing hope that humans are capable of thought reading all this junk.
They purchased their books to get inspiration from, the original author gets paid, and the author consented to selling it. That’s the difference.
Also the LLM can post entire snippets or chapters of books, which of course you’ll take at face value even if it hallucinates and makes the author look like a worse author then they are.
Generally they probably bought the books they read though.
If George RR Martin torrented Tolkien, wouldn’t he be infringing on the copyright no matter how he subsequently incorporated it into future output?
I completely agree that the training as infringement argument is ludicrous.
But OpenAI exposed themselves to IP infringement by sailing the high seas in how they obtained the works in the first place.
I hate that the world we live in is one where so much data is gated behind paywalls, but the law is what it is, and if the government was going to come down hard on Aaron Swartz for trying to bypass paywalls for massive amounts of written text, it’s not exactly fair if there’s a double standard for OpenAI doing the same thing in an even more closed fashion.
But yes, the degree of entitled focus on the premise of training an AI as equivalent of infringing is weird as heck to see from authors drawing quite clearly from earlier works in their own output.
I have to assume that openAI also paid for the books. if yes then i consider it the same as me reciting passages from memory or coming up with derivative text.
if no, then by all means, go after them and any model trainer for the cost of one book.
Asking an LLM to recite an entire novel isn’t even vaguely a thing yet.
Well, here’s straight from one of the suits against them:
I’m not even sure how they would have logistically gone about purchasing 294,000 books in bulk in digital form to be fed into training. Using the existing collections seems much more likely, but I suppose we’ll see what turns up in litigation.
Also, the penalty for downloading copyrighted material if willful infringement is up to $250,000 per work. So it’s quite a bit more than the cost of one book on the line…
God that Aaron/jstor thing makes me see red every time. Swartz was scraping jstor to publish it for the benefit of everyone, openai is doing it to make billions of dollars. Don’t forget who the bad guys are (and donate to sci-hub)