22 ms·

Nobody is going to "generate training data" if they get precisely nothing out of it. The complete disregard for intellectual property rights being exhibited by

Nobody is going to "generate training data" if they get precisely nothing out of it. The complete disregard for intellectual property rights being exhibited by the "AI" community will inevitably come back to bite them in the arse one day.

ramesh31 4y ago
LLMs are not regurgitating copywrighted works. You're fundamentally missing what's happening here if you think that. They do the same thing you and I do; read the literature and come to a conclusion. To quote Wittgenstein, "My words are my world", i.e the sum of the textual and verbal output I am capable of producing are me to the rest of the world. In the same way, LLM AIs are approaching something indiscernible from human intelligence.
- partiallypro 4y ago
  This would be well and good if you couldn't occasionally ask ChatGPT for a response then Google it and find a near identical response elsewhere in what would have otherwise been copyright materials. Other AI tools are doing the same thing, and the most blatant is with art. If you honestly think ChatGPT etc aren't training on copyrighted data and just rewriting it for tons of responses (and images), you're out of your mind. It is just incredibly difficult to prove, and the laws currently are very blurred.
  - jacobn 4y ago
    Isn’t their point that humans also do that? Without it necessarily being called copyright infringement? Lots of subtleties: fair use, what’s not fair use, etc? It does seem like we’re in new territory, I can see merits on both sides and I guess we’ll figure it out via the courts / new laws over the next couple of years.
    - p1necone 4y ago
      It is called copyright infringement. If I write a book, and there's more than a few sentences in a row of content that's verbatim/almost verbatim the same as someone elses published writing (excluding intentional quotes/references) I do get in trouble. You can't just go "computer did it" and handwave away existing IP law.
      - stale2002 4y ago
        > there's more than a few sentences in a row of content that's verbatim/almost verbatim Ok, now image that it isn't giving out verbatim results. Which is a lot of content. Totally fine then right?
      - anonzzzies 4y ago
        > I do get in trouble. Unless the original was a bestseller and you are a bestseller, you almost certainly will not. This obviously happens all the time as most writers suck and do actually deliberately copy things and then self publish. AI is actually helping to now not do that.
    - precompute 4y ago
      Except, humans aren't machines or tools. Our laws don't answer to some logical / mathematical ideal, they're built to address social realities. LLMs aren't citizens.
  - philwelch 4y ago
    > This would be well and good if you couldn't occasionally ask ChatGPT for a response then Google it and find a near identical response elsewhere in what would have otherwise been copyright materials. Likewise, someone could go on Hacker News to start an argument about intellectual property and half of the responses would be nearly identical to comments on Slashdot twenty years ago.
- Dalewyn 4y ago
  >They do the same thing you and I do; read the literature and come to a conclusion. No.
  - flangola7 4y ago
    Very compelling argument you've made.
    - Dalewyn 4y ago
      [flagged]
      - Zetobal 4y ago
        Sounds illogical to force a conversation and than withdraw with a argument about logic while dismissing every counter argument. :D
      - ux-app 4y ago
        seems you're getting unnecessarily hung up on the "how" of it. planes fly but not like birds. does it matter from an "object moved from point A to B through the air" point of view? not really. LLMs are kind of the same. it's obvious from even the most cursory examination that they're doing something quite different to statistical word mashing. is it being achieved in the same way as your gray matter? no. in a black box sense, does it matter? also no.
        Dalewyn 4y ago
        Considering the argument is irrationally applying human rights and rights of living beings to these computers running mildly complicated software, yes the "How" does matter a lot.
        ace2358 4y ago
        The concepts of lift and flight actually are the same for birds and planes and insects and drones and hellicoptors…… or fish! And whales. And any other organism that uses voyance and lift to ‘fly’ through the ‘medium’. Every form of intelligence we know of is operating in a fundamentally different environment to what these programs are doing.
- srcreigh 4y ago
  The key difference is that we know how LLMs work but we still haven’t solved the problem of consciousness, or even neuroscience. You can’t equate something you understand with something you don’t understand. Something you can create with something you can’t.
- nostromo123 4y ago
  "They do the same thing you and I do; read the literature and come to a conclusion." I'm sorry for you if you can't do better than an LLM.
partiallypro 4y ago
There will be tons of lawsuits in the next decade over the data AI is using to generate its responses. Some of the most blatant and easy to prove have been in image generation with them literally including blurred Getty logos with their output.
- Zuiii 4y ago
  Are the blurry getty logos imposed on backdrops from getty as well? I think what's happening here is that the AI is treating this feature like any other feature (table, nose, paperclip) and included it because it assumed the context was appropriate for it. On the otherhand, if the AI can be coxed into producing nearly complete getty image with only minor modifications, then maybe there would be grounds for a copyright lawsuit. I think AI detractors should focus on passing new laws. Copyright is not the right tool for this.
  - lambertsimnel 4y ago
    That's an interesting point, but I wonder whether Getty could sue on other grounds, such as trademark infringement
- whywhywhywhy 4y ago
  Getty are going to be one of the biggest investors in the space if they’re smart. Why pay photographers when you already have the dataset to create them on demand.
donpark 4y ago
Derivative training data is already part of the training regime. Earliest example produced rotated images for more robust image recognition. IP laws are legal constructs and subject to changes like everything else depending on our needs. When and if enough of us see benefits from AI hugely outweighing benefits from maintaining IP framework, I'm pretty sure laws will be passed to make needed adjustments.
- chitowneats 4y ago
  Literal shape-rotating?? I thought it was only the wordcels who were in trouble.
- srcreigh 4y ago
  Any automated process for generating training data is part of the model, not part of the training data. Authentic human generated information has irreducible newness.
  - adrianN 4y ago
    >Authentic human generated information has irreducible newness I believe that point is open to debate.
    - xingyzt 4y ago
      AI couldn't have started any of the artistic and literary movements, which don't just blend between traditional styles but reject some aspect to forge something new. You can't get AI to generate cubism from training on Victorian art.
      - adrianN 4y ago
        So do you believe that human brains are somehow magic and don't follow the laws of physics and can't be simulated by Turing machines?
        4y ago
franciscop 4y ago
"Intellectual property rights" is a concept made by greedy proprietors to build monopolies and keep competition out. It's no chance it was strongly pushed by Disney, and in its current form goes against the society interests as a whole. It is sold maliciously as "it incentivizes creative work", but how does "life of the author + 50 years" incentivize the author to keep creating? It just incentivizes them to keep exploiting. Now I think we'd all agree to something more realistic to actually protect authors, like 10-30 years protection from publication. But every time someone brings "copyright! IP!" to ML I just laugh since it's one of the more corrupt laws that we have currently.
- wuqs 4y ago
  excuse me sir, are you blind by silicon? greedy is to train ML models on copyrighted data of hundreds of thousands of people with far beyond the amount of power Disney have! also how many people are responsible for stuff like ChatGPT... 300? getting the bulky of their trainning data from, again, who? it feels the monopoly history is re-inventing itself, now with fancy computers replicating human behaviour
  - stale2002 4y ago
    That sounds horrible. I guess the only solution is to get rid of copyright law so people don't exploit it so much. Problem solved. Let's all stop protecting this, and allow anyone to innovate by making cool and new awesome things.
    - thrown123098 4y ago
      Sounds great, and when someone leaks the model online I can make a product on top of it without paying anyone.
      - Dalewyn 4y ago
        What goes around comes around.
      - 4y ago
misssocrates 4y ago
Predictions on how it's going to come back to bite them?
- evdubs 4y ago
  5% chance the copyright industry will sue AI companies for verbatim reproduction of copyrighted works. 0.1% chance the Feds get involved and sue AI companies for criminal copyright infringement.
  - petesergeant 4y ago
    If you can induce ChatGPT to repeat a copyright work, I wonder if that’s definitive proof of copyright infringement. It has no problem reproducing song lyrics for me, which I’m pretty sure are absolutely copyright. Interestingly, it’ll start to reproduce a famous book for me, and then cut off in such a way as to suggest there’s a hard filter to stop it proceeding
- precompute 4y ago
  It will not bite them back, because they will argue that what was put in the dataset was "easy" and not worthy of copyright anyway. Then, they'll tout how smart their "AI" is. Two birds with one stone.
hn92726819 4y ago
> Nobody is going to "generate training data" if they get precisely nothing out of it. What? What do you call literally the entire training dataset of stable diffusion, dalle, gpt3, and gpt4? You think they compensated anyone that went into the dataset? Did you get paid for your comment, because guess what? Your comment will end up in the gpt4.1 or gpt5 model. Training data will likely continue to include everything public on the internet. Not everyone will stop posting things online for free, therefore humans will still generate training data for models.
- ipaddr 4y ago
  Soon we get content created for gpt and introduce a new level of spam.
- Dalewyn 4y ago
  You know what people treated like shit do? They go find greener pastures to spend their lives in, taking any of their benefits to humanity with them. Sure, random comments of insignificant value like these will probably continue being created. Valuable content, though? That's a different story. "This is why we can't have nice things." is a saying that accurately reflects what could happen if the "AI" community doesn't stop behaving like uncivilized barbarians. Personally, I'd prefer to have nice things.
  - ShamelessC 4y ago
    Why are you putting AI in quotes, out of curiosity? My priors tell me that great art will continue to be made, both with and without these new tools; and that people will simply re-adjust to a new normal of sorts. What are you suggesting exactly? That all the creators are so weak-willed that they simply stop creating? Or are you implying that they are going to all go live on some sort of "no-internet-allowed" island together?
    - Dalewyn 4y ago
      >Why are you putting AI in quotes, out of curiosity? Because it's not intelligent. It's definitely artificial, but it's not intelligent in any meaningful sense of the term. "Automatic Interpolation" or something similar would be a far more suitable use for the acronym. >What are you suggesting exactly? Noone is going to spend time making something if the next guy immediately goes "You made this? ... I made this." like clockwork. Even moreso if someone was making something professionally. If we do see creators and inventors go away, it would be because of their own volition and desire to find greener pastures worthy of their time rather than "AI" usurping their jobs.
matheusmoreira 4y ago
> The complete disregard for intellectual property rights Those "rights" should not even exist to begin with. Nobody is interested in bending over backwards just to appease a bunch of aging monopolists. We want to explore the full potential of technology, not see it crippled by artificial limitations due to "concerns" over imaginary property.
- phone8675309 4y ago
  That's an incredibly reductive view. Copyright protects me when I do any creative work, I do not want my creative work to be part of any AI training dataset without my explicit permission, and I want laws that force people training generative AI models to copyright clear all of their training data and moreover be able to show which inputs corresponded to which outputs so the result is auditable. If Disney trains an AI good enough to just generate a good story then I think that everyone who fed training data into that model should get a piece of the revenue.
  - galaxyLogic 4y ago
    AI copies everything, except copyright notices. Curious why is that?
  - matheusmoreira 4y ago
    Reductive for you, maybe. I don't want this technology limited in literally any way. How many times must we go through this? Destroying perfectly good technology because it hurts the copyright industry's interests? I say it's time to end copyright instead. They're holding us back with these concerns.
    - meheleventyone 4y ago
      This is going to be awkward for the incentive to generate these models which are protected by the same rights.
      - PoignardAzur 4y ago
evdubs 4y ago
When you download a couple epubs from a torrent site, you're an evil pirate stealing books. When you're a FAANG-affiliated AI company and you download all the ebooks and charge for the derivative works your algorithm produces, you're a visionary changing the future.
- rhtgrg 4y ago
  Yeah, this is just capitalism in action. Almost every startup has stories like this, which appears to be why YC's startup school shies away from "goody two-shoes" type people and has questions like "please tell us about the time you most successfully hacked some (non-computer) system". For example, in the early days Amazon famously "leveraged" the purchase of out of print books from their wholesalers to obtain orders that would otherwise be too small to ship. That's just how capitalism works. You have to be either too small to bother with or too big to fail -- the problem area is in the middle.
  - shafyy 4y ago
    I hate that YC question. A "good" answer to that is almost always something that's borderline illegal or otherwise ethically bad.
    - DC-3 4y ago
      It's almost as if unethical behaviour is incentivized under capitalism...
    - mahathu 4y ago
      I automated applying for apartments in my city's public housing program as they regularly get hundreds of applications within a few minutes and automatically take down the ads again, often before their email notifications even arrive. That way I at least got invited to some viewings without having to check it every few minutes throughout the day, as many people I know have been doing for months. I don't see how that can be considered ethically bad. This is in Berlin.
XorNot 4y ago
No one's going to need to generate training data, because image generators will simply setup and run their own artist pages using LLMs to interact with their fans, and then weight their outputs by likes/clicks/web requests. The idea that complete original "art" is needed to run image generators is a very temporary state of affairs that bootstraps the proof of concept, but is in no way necessary to the process. The entire process can be adversarially trained solely against human aesthetic values - we know this, because image generators are already a derivative technology of building image classifiers - i.e. they functionally are adversarial against a given textual description in a classifier. The real limitation right now is speed: once we can diffuse at > 1 frame-per-second, data collection can be run solely from the consumer side.
scotty79 4y ago
People create stuff out of the sheer desire to create stuff. You only need to make sure they won't starve while they do it. And there's so many ways of ensuring that apart from paying them for copies of their output.
colordrops 4y ago
Intellectual property rights shouldn't exist as a concept. Digital scarcity is a scam. Property is a function of the natural scarcity of physical matter, which doesn't apply to data. Rent seekers and power brokers unnecessarily brought this concept into the digital realm.
_ZeD_ 4y ago
funny, people invented, discovered, and created stuff even before the "intellectual property rights". Those terms have been created by distributors, not creators, of contents.
imgabe 4y ago
> Nobody is going to "generate training data" if they get precisely nothing out of it. How much did you get paid to write this comment?
- nicbou 4y ago
  People are happy to share their opinion, but don't expect people to write long, well-researched content about anything. In my case, the content I write involves hours of research, and correspondence with a network of experts I have built up over 5 years. I constantly update and revise that content. The website on which it's hosted was also developed with great attention over the same period. I don't think that I'd keep doing if I was deprived of revenue, and from the kind feedback of readers. I certainly couldn't do it full time.
  - imgabe 4y ago
    If GPT could do the same research - reading a bunch of primary sources and condensing and summarizing the information - why wouldn't we want to have GPT do that rather than having a human do it? We don't employ roomfuls of people to solve math equations by hand to calculate trajectories anymore, because the computer can do that faster. If GPT can't do research and produce content with the same level of accuracy as you (which, so far, it can't) than surely people would be happy to continue compensating you to do so.
    - nicbou 4y ago
      Sometimes I am the primary source. I got the information from calling various government offices and interviewing people. I write with a certain amount of specific context that a LLM does not have. To illustrate my point, an LLM might summarise legal texts, but only a lawyer can tell you the odds. I tried ChatGPT. I asked it questions about the work I do, and it was about as correct as the cheapest copywriter. The writing was very good (if verbose), but the information was shallow but credible bullshit. There is no efficient market that allocates money to the best products. In reality, the winner takes it all. Therein lies the problem. If people find credible bullshit before valid information, they might stop there. It deprives them of valid information, and me of revenue. There is plenty of credible bullshit already. GPT just makes it cheaper to produce.
bontaq 4y ago
This is entirely my thesis on why they'll win. Of course google wishes they could ignore all attribution. They are slighting the law just like airbnb and uber did. They offer a better form of consumption. This is disruption at it's finest, and it will be in the courts forever.
patrickk 4y ago
> Nobody is going to "generate training data" if they get precisely nothing out of it. It's using common crawl and wikipedia among other sources. The original purpose is something else. Another anecdote that I find amusing is that Google Translate was trained on EU legal texts (among other sources), which are (generally) translated into 26/27 languages of the EU members. A huge, multinational bureaucracy, generated the texts by hand, but Google gets all the credit. I'm guessing that's part of the reason that Chinese isn't so accurate compared with other translations.
DuckFeathers 4y ago
"Intellectual property rights" is a scam implemented by corrupt western cronies and is completely unfair to the people. No person owns the right to their "invention" because it is leeched off of other people's work. No corporation owns the right to whatever the people they employed came up with. The IP thing needs to be eradicated. We need more countries to be strong and brave and say no to westerners forcing them to be unable to copy technologies... like China has done to some extent... but that might take too long. The path where AI lawyers can sort IP disputes instantly and hence competitors can easily find out how to copy technology without violating laws would be another nail in the coffin of this corrupt practice. > Nobody is going to "generate training data" if they get precisely nothing out of it. The people who we don't want generating training data will not generate it... and that is a good thing. You have been generating too much crap in the name of innovation anyway. Real tangible useful innovation gets copied naturally in the marketplace in absence of corruption.
- precompute 4y ago
  IP is the promise that you get the dues for your work, guaranteed by the government. It is what enables the entire "information economy". Without it, we'd soon need to weed out people that don't contribute to IP.
ChuckNorris89 4y ago
>Nobody is going to "generate training data" if they get precisely nothing out of it. "Before accessing your account, please solve this puzzle to verify you're not a robot."
- precompute 4y ago
  "Your OpenAI free tier requires you to install the OpenAI app and accrue 30 OpenStat points for every 10000 tokens. Please note, you can only earn 1 OpenStat point every two hours. Please do not deny Location Access, Microphone Access, Camera Access or Storage Access permissions, or this program will stop working."
dragonwriter 4y ago
> Nobody is going to "generate training data" if they get precisely nothing out of it. Everyone that has a job and gets paid for it generates (with the appropriate monitoring) some kind of data that could be training data in the course of doing their job. (And doesn’t raise the issue of “if they get precisely nothing out of it”.) Once no humans have jobs because AIs are better at them (should that occur), then the AIs themselves will be the best source of training data for the next generation. There are many problems to address with AI; “Where is training data going to come from,” however, is nowhere on the list. It is a distraction from real issues.
zirgs 4y ago
Copyright activists should not have pushed Mickey Mouse protection acts , sued fangame creators and the like. I have zero sympathy for copyright activists now. Copyright reform is long overdue.
- whywhywhywhy 4y ago
  People just side with it as a way to get what they want. Crazy how many artists who’s portfolios are full of their own drawings of Disney IP shouting that Disney should use IP law to stop stable diffusion. The hypocrisy completely lost on them.
version_five 4y ago
People write open source software and get "nothing" out of it. And people post stuff on the public internet and get "nothing" out of it. It's not obvious to me that everyone who has shared something permissively is unhappy if someone else uses it for training data. In many cases, it's just loud people who want a new revenue stream, which is the problem with intellectual property generally, the concept breeds entitlement. Why should I retroactively profit from what I willingly shared because somebody else found a use for it.
- nicbou 4y ago
  Let's say that I repair donated bicycles and give them away for free. It's a fun activity that does some good for the community. Now, let's say that on my morning walk, I see someone selling a bike I gave them a week ago. Is it fair to be offended by this?
  - version_five 4y ago
    There are two issues I see here. One, it's not analogous because you're talking about physical property which is not intellectual property. One might be offended because it means the person they gave the bike to doesn't value it and really just wants to get money. They could have given it directly to someone who really wanted a bike. That's irrelevant when something can be copied for free. It would be more like if you were really good at fixing bikes and showed someone how to be good at it to, and they used what they learned to open a paid bike repair shop. Re the actual bike thing (which has nothing to do with the discussion) I wouldn't be offended if the person sold it. In the hypothetical example, I got the enjoyment of fixing the bike, apparently gave it to someone no strings attached, and he's found a way that maximizes the utility he gets out of it. What's the problem?
- account42 4y ago
  I have released works under copyleft licenses and am not OK with Microsoft using those for profit without respecting the license (e.g. without opening up the models derived from copyleft works) when Microsoft themselves have been going after others copying their works when that benefited them. They shouldn't get to have it both ways - either we reform copyright or everyone should be required to follow it. If anyhting, large corporations should be subject to stricter enforcement since they already can benefit more due to their size.
PoignardAzur 4y ago
> Nobody is going to "generate training data" if they get precisely nothing out of it. This is trivially false. Eg, fanfiction exists. Maybe monetizing art will become hard as cheap art becomes ubiquitous, but that doesn't mean people will stop creating stuff.

ramesh31 4y ago

LLMs are not regurgitating copywrighted works. You're fundamentally missing what's happening here if you think that. They do the same thing you and I do; read the literature and come to a conclusion. To quote Wittgenstein, "My words are my world", i.e the sum of the textual and verbal output I am capable of producing are me to the rest of the world. In the same way, LLM AIs are approaching something indiscernible from human intelligence.

partiallypro 4y ago
This would be well and good if you couldn't occasionally ask ChatGPT for a response then Google it and find a near identical response elsewhere in what would have otherwise been copyright materials. Other AI tools are doing the same thing, and the most blatant is with art. If you honestly think ChatGPT etc aren't training on copyrighted data and just rewriting it for tons of responses (and images), you're out of your mind. It is just incredibly difficult to prove, and the laws currently are very blurred.
- jacobn 4y ago
  Isn’t their point that humans also do that? Without it necessarily being called copyright infringement? Lots of subtleties: fair use, what’s not fair use, etc? It does seem like we’re in new territory, I can see merits on both sides and I guess we’ll figure it out via the courts / new laws over the next couple of years.
  - p1necone 4y ago
    It is called copyright infringement. If I write a book, and there's more than a few sentences in a row of content that's verbatim/almost verbatim the same as someone elses published writing (excluding intentional quotes/references) I do get in trouble. You can't just go "computer did it" and handwave away existing IP law.
    - stale2002 4y ago
      > there's more than a few sentences in a row of content that's verbatim/almost verbatim Ok, now image that it isn't giving out verbatim results. Which is a lot of content. Totally fine then right?
    - anonzzzies 4y ago
      > I do get in trouble. Unless the original was a bestseller and you are a bestseller, you almost certainly will not. This obviously happens all the time as most writers suck and do actually deliberately copy things and then self publish. AI is actually helping to now not do that.
  - precompute 4y ago
    Except, humans aren't machines or tools. Our laws don't answer to some logical / mathematical ideal, they're built to address social realities. LLMs aren't citizens.
- philwelch 4y ago
  > This would be well and good if you couldn't occasionally ask ChatGPT for a response then Google it and find a near identical response elsewhere in what would have otherwise been copyright materials. Likewise, someone could go on Hacker News to start an argument about intellectual property and half of the responses would be nearly identical to comments on Slashdot twenty years ago.
Dalewyn 4y ago
>They do the same thing you and I do; read the literature and come to a conclusion. No.
- flangola7 4y ago
  Very compelling argument you've made.
  - Dalewyn 4y ago
    [flagged]
    - Zetobal 4y ago
      Sounds illogical to force a conversation and than withdraw with a argument about logic while dismissing every counter argument. :D
    - ux-app 4y ago
      seems you're getting unnecessarily hung up on the "how" of it. planes fly but not like birds. does it matter from an "object moved from point A to B through the air" point of view? not really. LLMs are kind of the same. it's obvious from even the most cursory examination that they're doing something quite different to statistical word mashing. is it being achieved in the same way as your gray matter? no. in a black box sense, does it matter? also no.
      - Dalewyn 4y ago
        Considering the argument is irrationally applying human rights and rights of living beings to these computers running mildly complicated software, yes the "How" does matter a lot.
      - ace2358 4y ago
        The concepts of lift and flight actually are the same for birds and planes and insects and drones and hellicoptors…… or fish! And whales. And any other organism that uses voyance and lift to ‘fly’ through the ‘medium’. Every form of intelligence we know of is operating in a fundamentally different environment to what these programs are doing.
srcreigh 4y ago
The key difference is that we know how LLMs work but we still haven’t solved the problem of consciousness, or even neuroscience. You can’t equate something you understand with something you don’t understand. Something you can create with something you can’t.
nostromo123 4y ago
"They do the same thing you and I do; read the literature and come to a conclusion." I'm sorry for you if you can't do better than an LLM.

partiallypro 4y ago

There will be tons of lawsuits in the next decade over the data AI is using to generate its responses. Some of the most blatant and easy to prove have been in image generation with them literally including blurred Getty logos with their output.

Zuiii 4y ago
Are the blurry getty logos imposed on backdrops from getty as well? I think what's happening here is that the AI is treating this feature like any other feature (table, nose, paperclip) and included it because it assumed the context was appropriate for it. On the otherhand, if the AI can be coxed into producing nearly complete getty image with only minor modifications, then maybe there would be grounds for a copyright lawsuit. I think AI detractors should focus on passing new laws. Copyright is not the right tool for this.
- lambertsimnel 4y ago
  That's an interesting point, but I wonder whether Getty could sue on other grounds, such as trademark infringement
whywhywhywhy 4y ago
Getty are going to be one of the biggest investors in the space if they’re smart. Why pay photographers when you already have the dataset to create them on demand.

donpark 4y ago

Derivative training data is already part of the training regime. Earliest example produced rotated images for more robust image recognition. IP laws are legal constructs and subject to changes like everything else depending on our needs. When and if enough of us see benefits from AI hugely outweighing benefits from maintaining IP framework, I'm pretty sure laws will be passed to make needed adjustments.

chitowneats 4y ago
Literal shape-rotating?? I thought it was only the wordcels who were in trouble.
srcreigh 4y ago
Any automated process for generating training data is part of the model, not part of the training data. Authentic human generated information has irreducible newness.
- adrianN 4y ago
  >Authentic human generated information has irreducible newness I believe that point is open to debate.
  - xingyzt 4y ago
    AI couldn't have started any of the artistic and literary movements, which don't just blend between traditional styles but reject some aspect to forge something new. You can't get AI to generate cubism from training on Victorian art.
    - adrianN 4y ago
      So do you believe that human brains are somehow magic and don't follow the laws of physics and can't be simulated by Turing machines?
      - 4y ago

franciscop 4y ago

"Intellectual property rights" is a concept made by greedy proprietors to build monopolies and keep competition out. It's no chance it was strongly pushed by Disney, and in its current form goes against the society interests as a whole. It is sold maliciously as "it incentivizes creative work", but how does "life of the author + 50 years" incentivize the author to keep creating? It just incentivizes them to keep exploiting. Now I think we'd all agree to something more realistic to actually protect authors, like 10-30 years protection from publication. But every time someone brings "copyright! IP!" to ML I just laugh since it's one of the more corrupt laws that we have currently.

wuqs 4y ago
excuse me sir, are you blind by silicon? greedy is to train ML models on copyrighted data of hundreds of thousands of people with far beyond the amount of power Disney have! also how many people are responsible for stuff like ChatGPT... 300? getting the bulky of their trainning data from, again, who? it feels the monopoly history is re-inventing itself, now with fancy computers replicating human behaviour
- stale2002 4y ago
  That sounds horrible. I guess the only solution is to get rid of copyright law so people don't exploit it so much. Problem solved. Let's all stop protecting this, and allow anyone to innovate by making cool and new awesome things.
  - thrown123098 4y ago
    Sounds great, and when someone leaks the model online I can make a product on top of it without paying anyone.
    - Dalewyn 4y ago
      What goes around comes around.
    - 4y ago

misssocrates 4y ago

Predictions on how it's going to come back to bite them?

evdubs 4y ago
5% chance the copyright industry will sue AI companies for verbatim reproduction of copyrighted works. 0.1% chance the Feds get involved and sue AI companies for criminal copyright infringement.
- petesergeant 4y ago
  If you can induce ChatGPT to repeat a copyright work, I wonder if that’s definitive proof of copyright infringement. It has no problem reproducing song lyrics for me, which I’m pretty sure are absolutely copyright. Interestingly, it’ll start to reproduce a famous book for me, and then cut off in such a way as to suggest there’s a hard filter to stop it proceeding
precompute 4y ago
It will not bite them back, because they will argue that what was put in the dataset was "easy" and not worthy of copyright anyway. Then, they'll tout how smart their "AI" is. Two birds with one stone.

hn92726819 4y ago

> Nobody is going to "generate training data" if they get precisely nothing out of it. What? What do you call literally the entire training dataset of stable diffusion, dalle, gpt3, and gpt4? You think they compensated anyone that went into the dataset? Did you get paid for your comment, because guess what? Your comment will end up in the gpt4.1 or gpt5 model. Training data will likely continue to include everything public on the internet. Not everyone will stop posting things online for free, therefore humans will still generate training data for models.

ipaddr 4y ago
Soon we get content created for gpt and introduce a new level of spam.
Dalewyn 4y ago
You know what people treated like shit do? They go find greener pastures to spend their lives in, taking any of their benefits to humanity with them. Sure, random comments of insignificant value like these will probably continue being created. Valuable content, though? That's a different story. "This is why we can't have nice things." is a saying that accurately reflects what could happen if the "AI" community doesn't stop behaving like uncivilized barbarians. Personally, I'd prefer to have nice things.
- ShamelessC 4y ago
  Why are you putting AI in quotes, out of curiosity? My priors tell me that great art will continue to be made, both with and without these new tools; and that people will simply re-adjust to a new normal of sorts. What are you suggesting exactly? That all the creators are so weak-willed that they simply stop creating? Or are you implying that they are going to all go live on some sort of "no-internet-allowed" island together?
  - Dalewyn 4y ago
    >Why are you putting AI in quotes, out of curiosity? Because it's not intelligent. It's definitely artificial, but it's not intelligent in any meaningful sense of the term. "Automatic Interpolation" or something similar would be a far more suitable use for the acronym. >What are you suggesting exactly? Noone is going to spend time making something if the next guy immediately goes "You made this? ... I made this." like clockwork. Even moreso if someone was making something professionally. If we do see creators and inventors go away, it would be because of their own volition and desire to find greener pastures worthy of their time rather than "AI" usurping their jobs.

matheusmoreira 4y ago

> The complete disregard for intellectual property rights Those "rights" should not even exist to begin with. Nobody is interested in bending over backwards just to appease a bunch of aging monopolists. We want to explore the full potential of technology, not see it crippled by artificial limitations due to "concerns" over imaginary property.

phone8675309 4y ago
That's an incredibly reductive view. Copyright protects me when I do any creative work, I do not want my creative work to be part of any AI training dataset without my explicit permission, and I want laws that force people training generative AI models to copyright clear all of their training data and moreover be able to show which inputs corresponded to which outputs so the result is auditable. If Disney trains an AI good enough to just generate a good story then I think that everyone who fed training data into that model should get a piece of the revenue.
- galaxyLogic 4y ago
  AI copies everything, except copyright notices. Curious why is that?
- matheusmoreira 4y ago
  Reductive for you, maybe. I don't want this technology limited in literally any way. How many times must we go through this? Destroying perfectly good technology because it hurts the copyright industry's interests? I say it's time to end copyright instead. They're holding us back with these concerns.
  - meheleventyone 4y ago
    This is going to be awkward for the incentive to generate these models which are protected by the same rights.
    - PoignardAzur 4y ago

evdubs 4y ago

When you download a couple epubs from a torrent site, you're an evil pirate stealing books. When you're a FAANG-affiliated AI company and you download all the ebooks and charge for the derivative works your algorithm produces, you're a visionary changing the future.

rhtgrg 4y ago
Yeah, this is just capitalism in action. Almost every startup has stories like this, which appears to be why YC's startup school shies away from "goody two-shoes" type people and has questions like "please tell us about the time you most successfully hacked some (non-computer) system". For example, in the early days Amazon famously "leveraged" the purchase of out of print books from their wholesalers to obtain orders that would otherwise be too small to ship. That's just how capitalism works. You have to be either too small to bother with or too big to fail -- the problem area is in the middle.
- shafyy 4y ago
  I hate that YC question. A "good" answer to that is almost always something that's borderline illegal or otherwise ethically bad.
  - DC-3 4y ago
    It's almost as if unethical behaviour is incentivized under capitalism...
  - mahathu 4y ago
    I automated applying for apartments in my city's public housing program as they regularly get hundreds of applications within a few minutes and automatically take down the ads again, often before their email notifications even arrive. That way I at least got invited to some viewings without having to check it every few minutes throughout the day, as many people I know have been doing for months. I don't see how that can be considered ethically bad. This is in Berlin.

XorNot 4y ago

No one's going to need to generate training data, because image generators will simply setup and run their own artist pages using LLMs to interact with their fans, and then weight their outputs by likes/clicks/web requests. The idea that complete original "art" is needed to run image generators is a very temporary state of affairs that bootstraps the proof of concept, but is in no way necessary to the process. The entire process can be adversarially trained solely against human aesthetic values - we know this, because image generators are already a derivative technology of building image classifiers - i.e. they functionally are adversarial against a given textual description in a classifier. The real limitation right now is speed: once we can diffuse at > 1 frame-per-second, data collection can be run solely from the consumer side.

scotty79 4y ago

People create stuff out of the sheer desire to create stuff. You only need to make sure they won't starve while they do it. And there's so many ways of ensuring that apart from paying them for copies of their output.

colordrops 4y ago

Intellectual property rights shouldn't exist as a concept. Digital scarcity is a scam. Property is a function of the natural scarcity of physical matter, which doesn't apply to data. Rent seekers and power brokers unnecessarily brought this concept into the digital realm.

_ZeD_ 4y ago

funny, people invented, discovered, and created stuff even before the "intellectual property rights". Those terms have been created by distributors, not creators, of contents.

imgabe 4y ago

> Nobody is going to "generate training data" if they get precisely nothing out of it. How much did you get paid to write this comment?

nicbou 4y ago
People are happy to share their opinion, but don't expect people to write long, well-researched content about anything. In my case, the content I write involves hours of research, and correspondence with a network of experts I have built up over 5 years. I constantly update and revise that content. The website on which it's hosted was also developed with great attention over the same period. I don't think that I'd keep doing if I was deprived of revenue, and from the kind feedback of readers. I certainly couldn't do it full time.
- imgabe 4y ago
  If GPT could do the same research - reading a bunch of primary sources and condensing and summarizing the information - why wouldn't we want to have GPT do that rather than having a human do it? We don't employ roomfuls of people to solve math equations by hand to calculate trajectories anymore, because the computer can do that faster. If GPT can't do research and produce content with the same level of accuracy as you (which, so far, it can't) than surely people would be happy to continue compensating you to do so.
  - nicbou 4y ago
    Sometimes I am the primary source. I got the information from calling various government offices and interviewing people. I write with a certain amount of specific context that a LLM does not have. To illustrate my point, an LLM might summarise legal texts, but only a lawyer can tell you the odds. I tried ChatGPT. I asked it questions about the work I do, and it was about as correct as the cheapest copywriter. The writing was very good (if verbose), but the information was shallow but credible bullshit. There is no efficient market that allocates money to the best products. In reality, the winner takes it all. Therein lies the problem. If people find credible bullshit before valid information, they might stop there. It deprives them of valid information, and me of revenue. There is plenty of credible bullshit already. GPT just makes it cheaper to produce.

bontaq 4y ago

This is entirely my thesis on why they'll win. Of course google wishes they could ignore all attribution. They are slighting the law just like airbnb and uber did. They offer a better form of consumption. This is disruption at it's finest, and it will be in the courts forever.

patrickk 4y ago

> Nobody is going to "generate training data" if they get precisely nothing out of it. It's using common crawl and wikipedia among other sources. The original purpose is something else. Another anecdote that I find amusing is that Google Translate was trained on EU legal texts (among other sources), which are (generally) translated into 26/27 languages of the EU members. A huge, multinational bureaucracy, generated the texts by hand, but Google gets all the credit. I'm guessing that's part of the reason that Chinese isn't so accurate compared with other translations.

DuckFeathers 4y ago

"Intellectual property rights" is a scam implemented by corrupt western cronies and is completely unfair to the people. No person owns the right to their "invention" because it is leeched off of other people's work. No corporation owns the right to whatever the people they employed came up with. The IP thing needs to be eradicated. We need more countries to be strong and brave and say no to westerners forcing them to be unable to copy technologies... like China has done to some extent... but that might take too long. The path where AI lawyers can sort IP disputes instantly and hence competitors can easily find out how to copy technology without violating laws would be another nail in the coffin of this corrupt practice. > Nobody is going to "generate training data" if they get precisely nothing out of it. The people who we don't want generating training data will not generate it... and that is a good thing. You have been generating too much crap in the name of innovation anyway. Real tangible useful innovation gets copied naturally in the marketplace in absence of corruption.

precompute 4y ago
IP is the promise that you get the dues for your work, guaranteed by the government. It is what enables the entire "information economy". Without it, we'd soon need to weed out people that don't contribute to IP.

ChuckNorris89 4y ago

>Nobody is going to "generate training data" if they get precisely nothing out of it. "Before accessing your account, please solve this puzzle to verify you're not a robot."

precompute 4y ago
"Your OpenAI free tier requires you to install the OpenAI app and accrue 30 OpenStat points for every 10000 tokens. Please note, you can only earn 1 OpenStat point every two hours. Please do not deny Location Access, Microphone Access, Camera Access or Storage Access permissions, or this program will stop working."

dragonwriter 4y ago

> Nobody is going to "generate training data" if they get precisely nothing out of it. Everyone that has a job and gets paid for it generates (with the appropriate monitoring) some kind of data that could be training data in the course of doing their job. (And doesn’t raise the issue of “if they get precisely nothing out of it”.) Once no humans have jobs because AIs are better at them (should that occur), then the AIs themselves will be the best source of training data for the next generation. There are many problems to address with AI; “Where is training data going to come from,” however, is nowhere on the list. It is a distraction from real issues.

zirgs 4y ago

Copyright activists should not have pushed Mickey Mouse protection acts , sued fangame creators and the like. I have zero sympathy for copyright activists now. Copyright reform is long overdue.

whywhywhywhy 4y ago
People just side with it as a way to get what they want. Crazy how many artists who’s portfolios are full of their own drawings of Disney IP shouting that Disney should use IP law to stop stable diffusion. The hypocrisy completely lost on them.

version_five 4y ago

People write open source software and get "nothing" out of it. And people post stuff on the public internet and get "nothing" out of it. It's not obvious to me that everyone who has shared something permissively is unhappy if someone else uses it for training data. In many cases, it's just loud people who want a new revenue stream, which is the problem with intellectual property generally, the concept breeds entitlement. Why should I retroactively profit from what I willingly shared because somebody else found a use for it.

nicbou 4y ago
Let's say that I repair donated bicycles and give them away for free. It's a fun activity that does some good for the community. Now, let's say that on my morning walk, I see someone selling a bike I gave them a week ago. Is it fair to be offended by this?
- version_five 4y ago
  There are two issues I see here. One, it's not analogous because you're talking about physical property which is not intellectual property. One might be offended because it means the person they gave the bike to doesn't value it and really just wants to get money. They could have given it directly to someone who really wanted a bike. That's irrelevant when something can be copied for free. It would be more like if you were really good at fixing bikes and showed someone how to be good at it to, and they used what they learned to open a paid bike repair shop. Re the actual bike thing (which has nothing to do with the discussion) I wouldn't be offended if the person sold it. In the hypothetical example, I got the enjoyment of fixing the bike, apparently gave it to someone no strings attached, and he's found a way that maximizes the utility he gets out of it. What's the problem?
account42 4y ago
I have released works under copyleft licenses and am not OK with Microsoft using those for profit without respecting the license (e.g. without opening up the models derived from copyleft works) when Microsoft themselves have been going after others copying their works when that benefited them. They shouldn't get to have it both ways - either we reform copyright or everyone should be required to follow it. If anyhting, large corporations should be subject to stricter enforcement since they already can benefit more due to their size.

PoignardAzur 4y ago

> Nobody is going to "generate training data" if they get precisely nothing out of it. This is trivially false. Eg, fanfiction exists. Maybe monetizing art will become hard as cheap art becomes ubiquitous, but that doesn't mean people will stop creating stuff.