I like and agree with Gary's forecast that without proper compensation/licensing, original content will dwindle and the online world will be all re-consumed vomit. That alone should be a reason to make sure that content creators are incentivized to continue creating original and truthful content. Otherwise, everyone is a loser in this closed system.
And it is rather unsolvable too. Users *want* the systems to be able to produce correct information. But the better the model is at producing correct results (based on correctness of what was in the training material) the more will it memorise and the bigger the plagiarism problem. https://ea.rna.nl/2023/12/26/memorisation-the-deep-problem-of-midjourney-chatgpt-and-friends/
I was betting money that 2024 would see the start of an AI winter. A bet I made because I thought the limits of LLMs would become insurmountable and development outside of refining existing use cases would plateau. Now I'm just glad nothing in the wager said anything about the cause: AI companies having to grapple with the owners and creators vast troves of data they have to consume will just make LLMs' flaws more relevant.
Your post could be entitled ‘the desperate race for massive undue profits from genAI’. The more I read about this plagiarism issue with today’s genAI tools, the more I find amazing that global scale robbery of IP is going on without a larger echo from media and political spheres. And that bad faith and ill will declarations from major concerned genAI companies are considered as reliable technical expert opinions and not directly denounced as fraudulent marketing rhetoric.
To get a little perspective on this period we are in with AI and LLMs, I read two books about the early days of airplanes: John Lancaster's "The Great Air Race," and "The Wright Brothers" by David McCullough. What I think can be learned from this period of innovation around a world-changing technology, was how much additional work, practice and innovation was needed before air flight had broad use by the government (military) and public/private business. Risks were taken in the early phases (like flying without air pressurization or seat belts or navigation systems) that no one in their right mind would do today. Even the Wright Brothers talked about what they were doing were demonstrations of the some core principles upon which further engineering and innovation could take place. That further innovation was part of its integration and use in broader applications. They knew early on there were serious risks. "The Great Race" book is almost comical in the descriptions of the accidents (running into mountain ranges). I think what is being suggested for AI/LLM and the like by Gary and others is this same careful development with additional creativity and innovation to address problems in actual real world use. But as he points out, the venture-backed financing doesn't want to experience the early losses, and want to tilt the law to protect them. Historically this kind of thing has been tried, but mostly failed as innovation and creativity were used to address the real world problems (seat belts, anti-lock braking, materials innovation, wind-tunnel testing, etc.) and better products won out in the market. The equivalent of safety, environmental, and other kinds of rules and regulations are needed for AI/LLMs and we need to get on with it so we have better products and services at fair prices with reasonable returns to investors.
If GenAI is generating so much value for society that these companies deserve to have rules rewritten to serve its needs, it should be generating enough value to compensate publishers. It's incredibly short sighted for them to risk putting those creating actual new content out of business. Absolutely maddening.
Separately, I appreciated the comment in the IEEE on whether increases in accuracy accompany increases in the likelihood of creating plagiaristic content, as this is what I've been wondering about looking at the Midjourney images.
> But it would be interesting to think about what copywrite law would be like if humans had the ability to memorize entire books and recite them with prompted to do so.
This is super-pedantic and doesn't really undermine the point, but: humans do have this ability. However, becoming a hafiz -- one who has memorized the Koran -- is a serious endeavor, a lifetime task, and that for only a single holy book. The mass data consumption of an AI is a very different matter.
Love this bang-on post from GM and for once I read all the comments before posting. Great comments below. They must have seen this coming - why does it feel they are in damage control mode? Why were they not out ahead of the issue? I know the answer... =)
I just randomly asked Copilot to generate some images of space marines doing stuff. All the space marines looked like the ones from "Warhammer 40K." I kind of expected that some would, but I also expected that some might look like the ones from "Aliens" or from the many other science fiction franchises. Nope, all Warhammer. I tried a few prompts, including specifically saying for them to not look like the ones from Warhammer. I also asked for them to not be wearing battle armor, that did not work either. Searching for "space troopers" gave two images of generic battle spacesuits and two of Imperial Storm Troopers from Star Wars.
I kind of assumed that a lot of these allegations of copyright infringement were fishing for it by asking specific prompts. But "space marine" is not a concept unique to the Warhammer franchise. I think certain image-word associations are so ubiquitous online that it's impossible to get them out of the databases of these image generators. I assume it's the same for some text combinations as well.
To be fair, as a software engineer, I mostly use ChatGPT to pull in public-domain knowledge that I lack to accomplish specific tasks. None of the links it comes up with for any of my prompts point to any paywall of any sort. It helps save me gobs of time and makes me more effective and efficient because it's able to take large amounts of information via web plugins, munge it together, and spit it out in the form I need. I feel I'm getting my money's worth.
With that said, I do empathize with copyright holders and feel sad that the same tool has crossed those lines. If all copyrighted content was pulled from ChatGPT I wouldn't shed a tear. Also, if I wanted ChatGPT to access copyrighted data, I wouldn't mind paying an optional fee for the privilege (think Spotify and how one monthly fee gets distributed to the copyright owners whose content you actually use).
In fact, Github CoPilot has already been dinged for providing publicly accessible code in its responses and now it's possible to set it to generate original content only and not regurgitate public code verbatim. Anything is possible, if there's the political will to do so. We should continue to push back to our AI providers.
LLMs need a business model -- and they are only one step away from getting one.
No content creator is going to allow LLMs to train on their content and produce a trilion-dollar company without getting anything in return. Google is a trilion-dollar company because it brings as much value to content creators as well. Some even chose to pay money to Google to get their content scanned! (advertisers)
LLMs just need to accurately cite their sources. That's it. A trilion-dollar business in front of you.
I think the only thing left is OpenAI making the argument that Google and Facebook were allowed to crawl the web and create an ad-business out of it without paying license fees to content creators and publishers.
I’ve covered something similar in my recent post on Google and I think the backend consumption model needs to reflect attribution at the very least.
Imagine someone reading my posts and making a business out of it repeatedly by selling my insight for $20 a month when I take days of effort to research and document my analysis.
There is a long history of examples of people inventing/discovering something new, and then knowing what to do with it. One I read years ago, was that they thought people would use the telephone to listen to symphonies. The Wright Brothers did not imagine airplanes being used to bomb people to death in WWII (air mail maybe, but they were open after that). Well, here is a quote attributed to Mark Zuckerberg in the latest Bloomberg BusinessWeek article on Meta: "Hey Meta, look and tell me what pants to wear with this shirt," Zuckerberg said, holding one up. "Based on the image, it appears to be a striped shirt," the AI replied, showing off the real-world impact of decades of advancement in computer science. "A pair of dark washed jeans or sold color trousers would complement this shirt well." If there is a need for an indication of how far we have to go in using AI, I can find no better example. As a society we need to acknowledge that we don't know what we are really going to do with distributed data-based analytic capability using models that mirror human decision-making. We have no idea all the places to put it, and to do what, when and under what conditions, and with what oversight and containment. So-called experts, like I am assuming Zuckerberg and his crew must be considered have no crystal ball and their values and biases will handicap them just as it has others in the past in similar situations of new discovery. As we experience real life and addressing real needs and desires, we will move forward in small steps with learning and feedback loops.
I like and agree with Gary's forecast that without proper compensation/licensing, original content will dwindle and the online world will be all re-consumed vomit. That alone should be a reason to make sure that content creators are incentivized to continue creating original and truthful content. Otherwise, everyone is a loser in this closed system.
"We won't get fabulously rich if you don't let us steal, so please don't make stealing a crime."
Drop the Mic GM. That one line sums up everything! I love the complex simplicity of it.
"nobody is actually suggesting that OpenAI only use public domain works"
Maybe they should!
And it is rather unsolvable too. Users *want* the systems to be able to produce correct information. But the better the model is at producing correct results (based on correctness of what was in the training material) the more will it memorise and the bigger the plagiarism problem. https://ea.rna.nl/2023/12/26/memorisation-the-deep-problem-of-midjourney-chatgpt-and-friends/
I was betting money that 2024 would see the start of an AI winter. A bet I made because I thought the limits of LLMs would become insurmountable and development outside of refining existing use cases would plateau. Now I'm just glad nothing in the wager said anything about the cause: AI companies having to grapple with the owners and creators vast troves of data they have to consume will just make LLMs' flaws more relevant.
Your post could be entitled ‘the desperate race for massive undue profits from genAI’. The more I read about this plagiarism issue with today’s genAI tools, the more I find amazing that global scale robbery of IP is going on without a larger echo from media and political spheres. And that bad faith and ill will declarations from major concerned genAI companies are considered as reliable technical expert opinions and not directly denounced as fraudulent marketing rhetoric.
Thank you again Gary!! Great information.
To get a little perspective on this period we are in with AI and LLMs, I read two books about the early days of airplanes: John Lancaster's "The Great Air Race," and "The Wright Brothers" by David McCullough. What I think can be learned from this period of innovation around a world-changing technology, was how much additional work, practice and innovation was needed before air flight had broad use by the government (military) and public/private business. Risks were taken in the early phases (like flying without air pressurization or seat belts or navigation systems) that no one in their right mind would do today. Even the Wright Brothers talked about what they were doing were demonstrations of the some core principles upon which further engineering and innovation could take place. That further innovation was part of its integration and use in broader applications. They knew early on there were serious risks. "The Great Race" book is almost comical in the descriptions of the accidents (running into mountain ranges). I think what is being suggested for AI/LLM and the like by Gary and others is this same careful development with additional creativity and innovation to address problems in actual real world use. But as he points out, the venture-backed financing doesn't want to experience the early losses, and want to tilt the law to protect them. Historically this kind of thing has been tried, but mostly failed as innovation and creativity were used to address the real world problems (seat belts, anti-lock braking, materials innovation, wind-tunnel testing, etc.) and better products won out in the market. The equivalent of safety, environmental, and other kinds of rules and regulations are needed for AI/LLMs and we need to get on with it so we have better products and services at fair prices with reasonable returns to investors.
If GenAI is generating so much value for society that these companies deserve to have rules rewritten to serve its needs, it should be generating enough value to compensate publishers. It's incredibly short sighted for them to risk putting those creating actual new content out of business. Absolutely maddening.
Separately, I appreciated the comment in the IEEE on whether increases in accuracy accompany increases in the likelihood of creating plagiaristic content, as this is what I've been wondering about looking at the Midjourney images.
> But it would be interesting to think about what copywrite law would be like if humans had the ability to memorize entire books and recite them with prompted to do so.
This is super-pedantic and doesn't really undermine the point, but: humans do have this ability. However, becoming a hafiz -- one who has memorized the Koran -- is a serious endeavor, a lifetime task, and that for only a single holy book. The mass data consumption of an AI is a very different matter.
What about patent infringement? Where lies my liability when using AI and it provides a patented process, but no reference to the patent. User beware!
Love this bang-on post from GM and for once I read all the comments before posting. Great comments below. They must have seen this coming - why does it feel they are in damage control mode? Why were they not out ahead of the issue? I know the answer... =)
I just randomly asked Copilot to generate some images of space marines doing stuff. All the space marines looked like the ones from "Warhammer 40K." I kind of expected that some would, but I also expected that some might look like the ones from "Aliens" or from the many other science fiction franchises. Nope, all Warhammer. I tried a few prompts, including specifically saying for them to not look like the ones from Warhammer. I also asked for them to not be wearing battle armor, that did not work either. Searching for "space troopers" gave two images of generic battle spacesuits and two of Imperial Storm Troopers from Star Wars.
I kind of assumed that a lot of these allegations of copyright infringement were fishing for it by asking specific prompts. But "space marine" is not a concept unique to the Warhammer franchise. I think certain image-word associations are so ubiquitous online that it's impossible to get them out of the databases of these image generators. I assume it's the same for some text combinations as well.
To be fair, as a software engineer, I mostly use ChatGPT to pull in public-domain knowledge that I lack to accomplish specific tasks. None of the links it comes up with for any of my prompts point to any paywall of any sort. It helps save me gobs of time and makes me more effective and efficient because it's able to take large amounts of information via web plugins, munge it together, and spit it out in the form I need. I feel I'm getting my money's worth.
With that said, I do empathize with copyright holders and feel sad that the same tool has crossed those lines. If all copyrighted content was pulled from ChatGPT I wouldn't shed a tear. Also, if I wanted ChatGPT to access copyrighted data, I wouldn't mind paying an optional fee for the privilege (think Spotify and how one monthly fee gets distributed to the copyright owners whose content you actually use).
In fact, Github CoPilot has already been dinged for providing publicly accessible code in its responses and now it's possible to set it to generate original content only and not regurgitate public code verbatim. Anything is possible, if there's the political will to do so. We should continue to push back to our AI providers.
LLMs need a business model -- and they are only one step away from getting one.
No content creator is going to allow LLMs to train on their content and produce a trilion-dollar company without getting anything in return. Google is a trilion-dollar company because it brings as much value to content creators as well. Some even chose to pay money to Google to get their content scanned! (advertisers)
LLMs just need to accurately cite their sources. That's it. A trilion-dollar business in front of you.
I think the only thing left is OpenAI making the argument that Google and Facebook were allowed to crawl the web and create an ad-business out of it without paying license fees to content creators and publishers.
I’ve covered something similar in my recent post on Google and I think the backend consumption model needs to reflect attribution at the very least.
Imagine someone reading my posts and making a business out of it repeatedly by selling my insight for $20 a month when I take days of effort to research and document my analysis.
There is a long history of examples of people inventing/discovering something new, and then knowing what to do with it. One I read years ago, was that they thought people would use the telephone to listen to symphonies. The Wright Brothers did not imagine airplanes being used to bomb people to death in WWII (air mail maybe, but they were open after that). Well, here is a quote attributed to Mark Zuckerberg in the latest Bloomberg BusinessWeek article on Meta: "Hey Meta, look and tell me what pants to wear with this shirt," Zuckerberg said, holding one up. "Based on the image, it appears to be a striped shirt," the AI replied, showing off the real-world impact of decades of advancement in computer science. "A pair of dark washed jeans or sold color trousers would complement this shirt well." If there is a need for an indication of how far we have to go in using AI, I can find no better example. As a society we need to acknowledge that we don't know what we are really going to do with distributed data-based analytic capability using models that mirror human decision-making. We have no idea all the places to put it, and to do what, when and under what conditions, and with what oversight and containment. So-called experts, like I am assuming Zuckerberg and his crew must be considered have no crystal ball and their values and biases will handicap them just as it has others in the past in similar situations of new discovery. As we experience real life and addressing real needs and desires, we will move forward in small steps with learning and feedback loops.