Pay per crawl of StackOverflow wouldn't encourage me to post more on StackOverflow. (Not that I was anyway.) Presumably you'd need to pay content creators, but that seems quite inefficient:
1. I pay OpenAI
2. OpenAI rev shares to StackOverflow
3. StackOverflow mostly keeps that money, but shares some with me for posting
4. I get some money back to help pay OpenAI?
This is nonsense. And if the frontier labs are right about simulated data, as Tesla seems to have been right with its FSD simulated visualization stack, does this really matter anyway? The value I get from an LLM far exceeds anything I have ever received from SO or an O'Reilly book (as much as I genuinely enjoy them collecting dust on a shelf).
If the argument is "fairness," I can sympathize but then shrug. If the argument is sustainability of training, I'm skeptical we need these payment models. And if the argument is about total value creation, I just don't buy it at all.
We’ve seen decades of growing wage gaps and erosion of labors strength. The current elites don’t really care to enrich the people. Why would they care to do anything about this problem? They likely don’t see it as a problem at all.
If they did actually stumble on AGI (assuming it didn’t eat them too) it would be used by a select few to enslave or remove the rest of us.
People should vote for more socialist governments pushing for UBI and automation tax on the companies..but which this comment get downvoted because of the capitalism religion.
Example 1 is bad, StackOverflow had clearly plateaued and was well into the downward freefall by the time ChatGPT was released.
Example 2 is apparently "open source" but it's actually just Tailwind which unfortunately had a very susceptible business model.
And I don't really think the framing here that it's eating its own tail makes sense.
It's also confusing to me why they're trying to solve the problem of it eating its own tail - there's a LOT of money being poured into the AI companies. They can try to solve that problem.
What I mean is - a snake eating its own tail is bad for the snake. It will kill it. But in this case the tail is something we humans valued and don't want eaten, regardless of the health of the snake. And the snake will probably find a way to become independent of the tail after it ate it, rather than die, which sucks for us if we valued the stuff the tail was made of, and of course makes the analogy totally nonsensical.
The actual solutions suggested here are not related to it eating its own tail anyway. They're related to the sentiment that the greed of AI companies needs to be reeled in, they need to give back, and we need solutions to the fact that we're getting spammed with slop.
I guess the last part is the part that ties into it "eating its own tail", but really, why frame it that way? Framing it that way means it's a problem for AI companies. Let's be honest and say it's a problem for us and we want it solved for our own reasons.
I feel like the only solution to the problem is democratized RLHF, where whenever we get a bad answer from an LLM, we can immediately tell it what was wrong and it can learn from that.
GenAI changes the dynamics of information systems so fundamentally that our entire notion of intellectual property is being upended.
Copyright was predicated on the notion that ideas and styles can not be protected, but that explicit expressive works can. For example, a recipe can't be protected, but the story you wrap around it that tells how your grandma used to make it would be.
LLMs are particularly challenging to wrangle with because they perform language alchemy. They can (and do) re-express the core ideas, styles, themes, etc. without violating copyright.
People deem this 'theft' and 'stealing' because they are trying to reconcile the myth of intellectual property with reality, and are also simultaneously sensing the economic ladder being pulled up by elites who are watching and gaming the geopolitical world disorder.
There will be a new system of value capture that content creators need to position for, which is to be seen as a more valuable source of high quality materials than an LLM, serving a specific market, and effectively acquiring attention to owned properties and products.
It will not be pay-per-crawl. Or pay-per-use. It will be an attention game, just like everything in the modern economy.
Attention is the only way you can monetize information.
This is an article that I agreed with more reading the headline than I did when I finished reading the article itself.
Stack Overflow peaked in 2014 before beginning it's downward decline. How is that at all related to GenAI? GPT4 is when we really started seeing these things get used to replace SO, etc., and that would be early 2023 - and indeed the drop gets worse there - but after the COVID era spike, SO was already crashing hard.
Tailwind's business model was providing a component library built on top of their framework. It's a business model that relies on the framework being good enough for people to want to use it to begin with, but being bad enough that they'd rather pay for the component library than build it themselves. The more comfortable it is to use, the more productive it is, the worse the value proposition is for the premium upsell. Even other "open core" business models don't have this inherent dichotomy, much less open source on the whole, so it's really weird to try and extrapolate this out.
The thing is, people turn to LLMs to solve problems and answer questions. If they can't turn to the LLM to solve that problem or answer that question, they'll either turn elsewhere, in which case there is still a market for that book or blog post, or they'll drop the problem and question and move on. And if they were willing to drop the problem or question and move on without investigating post-LLM, were they ever invested enough to buy your book, or check more than the first couple of results on google?
I see what you mean, but the problem is that the LLM provider is trying to provide all the value from the book to the user without the user needing to look at the book at all. I agree if the LLM fails to do so then there is a market for the book. But the LLM provider is trying to minimize that as much as possible. And if the LLM succeeds at providing all the value of the book to the user, without providing any value to the book creator, then in the future there is no incentive to create the book at all, at which point the LLM has no value to provide, etc etc etc.
This is exactly the sentiment I have been trying to articulate myself.
The ONLY reason we are here today is because OpenAI, and Anthropic, by extension, took it upon themselves to launch chat bots trained on whatever datasources they could get in a short amount of time to quickly productize their investments. Their first versions didn't include any references to the source material, and just acted as if they knew everything.
When CoPilot was built as a better auto-complete engine, trained on opensource projects, it was an interesting idea, because it doing what people already did. They searched GitHub for examples of the solution or nudged them in that direction. However, the biggest difference, using other project code was stable, because it came with a LICENSE.md that you then agreed to, and paid it forward. (i.e. "I used code from this project").
CoPilot initially would just inject snippets for you, without you knowing the source. It was only later, they walked that back and if you did use CoPilot, it shows you the most-likely source of the code it used. This is exactly the direction all of the platforms seem headed.
It's not easy to walk back the free-for-all system (i.e. Napster), but I'm optimistic over time it'll become a more fair, pay to access system.
Companies will provide incentives for people to generate authentic content. Example X giving $1M reward for the top Articles. Because they can use it for training.
In fact, this might be overall good thing, because finally original content will be highly on demand since those companies now use to train their models. But we are probably just in a transition phase.
The other thing is that new sources of input will come, from LLM usage probably, so they cut the middle layer, users input in the LLM is also a form of input, and a hybrid co-creation between users/AI would generate content at much faster rater, which again would be used to train the model, and that would improve their quality.
> he took a PDF of my book Terraform: Up & Running, uploaded it into a GenAI tool, and asked the tool to follow the guidance in the book to generate Terraform code
This is ridiculous - AI doesn't need to be fed a PDF of a Terraform book to know how to Terraform. Blowing out context with hundreds of OCR'd pages of generic text on how to terraform isn't going to help anything.
The model that is broken is really ultimately going to be "content for hire". That's the industry that is going to be destroyed here because it's simply redundant now. Actual artwork, actual literature, actual music... these things are all safe as long as people actually want to experience the creations of others. Corporate artwork, simple documentation, elevator music.... these things are done; I'm sorry if you made a living making them but you were ultimately performing an artisinal task in a mostly soulless way.
I'm not talking about video game artists, mind you, I'm talking about the people who produced Corporate Memphis and Flat Design here. We'll all be better off if these people find a new calling in life.
Ironic that they created this post and I immediately felt like referencing ouroborous even when I hadn't looked at the website but then I opened it and the first thing I see is an AI generated image of Ouroborous
Like the irony is pretty deep with this one about this.
I am not sure if they could've gotten trademark from Inscryption/if they needed it but if they really wanted, I have found inscryption's ouroborous card to look the best and it was honestly how I discovered ouroborous in the first place! (became my favourite card, I love inscryption)
Even just searching Ouroborous on internet gave me some genuinely beautiful Ouroborous illustrations (Some stock photos, some not) but even using a stock photo might have made a better idea than using AI generated Ouroboros photo itself?
Everyone here throwing shade at Stack Overflow is clearly too young to remember the horror of Experts Exchange and every other technical help site prior to SO. For nearly a decade, it was absolutely transformative as a technical help resource. It certainly had its faults, but it was so far ahead of the other options as to be game-changing.
I believe that the main reason for SO's decline starting around 2018 was that most of the core technical questions had been answered. There was an enormous existing corpus of accepted answers around fundamental topics, and technology just doesn't change fast enough to sustain the site. Then the LLMs digested the site's (beautifully machine-readable) corpus along with the rest of the internet and now the AIs can give users that info directly, resulting in a downward spiral of traffic to SO, fewer new questions, etc.
Vale, Stack Overflow. You helped me solve many tricky problems.
Isn't this a self solving problem? If LLMs won't be able to solve a certain problem, arbitrage opportunities open to solve it as a human (by writing a blog post for example), just like in the old days, people typically did this for fame. I think LLMs can be understood as a really good way to search existing solutions, including combining multiple solutions in the fly. Only if something cannot be "found", we go back to the dynamics of before LLMs.
There may like 0.001% of the general population who would ever write a blog post, most of the LLM data in future would be fake blog posts used to trick LLMs or just spam.
The new world is one where someone can have an LLM assisted insight, post it on their blog for free, have it indexed by every agentic search engine, and it becomes part of the zeitgeist. That’s the new data that’ll feed the new models: a better information diet over time. And
guess what else: models are getting better at identifying - at scale - the high quality info that’s worth using as training data.
Strange nobody provided counterpoint. I would argue that questions meant for stackoverflow have shifted to genai tools and they have even richer data, questions, background of user and accepted answer so they dont need so data to improve. Side-effect of this is whoever gets most users wins and continues to win leading to monopoly as data is not public
I have a lot of opinions about LLMs, and most are not positive due to the error rate, the security nightmare, the IP issues, the ecological impact and the hype, hype, hype that stands in direct opposition from my years of experiments (starting in 2019).
But, I have to admit that a few years ago LLMs became part of my daily workflow for low-risk repetitive tasks, and sometimes even "sound boarding".
Foundation model sponsors already pay humans to generate authentic content, especially in technical areas that are underrepresented in general internet scrapes. I would imagine that this trend will continue.
Further, the "model collapse" hypothesis of 2020/2021 seems to have failed to materialize. Maybe we're still too early, and we're not yet seeing negative effects of OpenAI training on OpenAI output. But maybe "slop" is not being rewarded as much as human content, and having humans in the loop (even as readers) is preventing a slide into incoherence.
Will LLMs eventually disincentivize people from producing and publishing new original content? If that content is easily replicated by an LLM query, maybe. And maybe it's not the worst thing in the world. 5 years ago I would have bought an "FFmpeg Cookbook" from O'Reilly, but now I would just tell Claude exactly what I'm trying to achieve. As a consumer, I'm better off, and arguably we've saved the author of a hypothetical FFmpeg Cookbook weeks out of their precious life. Weeks they could spend doing something—anything—more valuable than rewording FFmpeg documentation.
The “snake eating its own tail” frame is real, but it’s not mystical — it’s incentives + sampling.
If the web gets flooded with LLM output and you train on it naively, you’re effectively training on your own prior. That pushes models toward the mean: less surprise, less specificity, more template-y phrasing. It’s like photocopying a photocopy: the sharp edges disappear.
The fix isn’t “never use synthetic data.” It’s to treat it like a controlled ingredient: tag provenance, keep a high-quality human/grounded core, filter aggressively, and anchor training to things that don’t self-contaminate (code that compiles/tests, math with verifiable proofs, retrieval with citations, real user feedback). Otherwise the easiest path is content volume, and volume is exactly what kills signal.
32 comments
[ 2.7 ms ] story [ 63.5 ms ] thread1. I pay OpenAI 2. OpenAI rev shares to StackOverflow 3. StackOverflow mostly keeps that money, but shares some with me for posting 4. I get some money back to help pay OpenAI?
This is nonsense. And if the frontier labs are right about simulated data, as Tesla seems to have been right with its FSD simulated visualization stack, does this really matter anyway? The value I get from an LLM far exceeds anything I have ever received from SO or an O'Reilly book (as much as I genuinely enjoy them collecting dust on a shelf).
If the argument is "fairness," I can sympathize but then shrug. If the argument is sustainability of training, I'm skeptical we need these payment models. And if the argument is about total value creation, I just don't buy it at all.
If they did actually stumble on AGI (assuming it didn’t eat them too) it would be used by a select few to enslave or remove the rest of us.
Example 1 is bad, StackOverflow had clearly plateaued and was well into the downward freefall by the time ChatGPT was released.
Example 2 is apparently "open source" but it's actually just Tailwind which unfortunately had a very susceptible business model.
And I don't really think the framing here that it's eating its own tail makes sense.
It's also confusing to me why they're trying to solve the problem of it eating its own tail - there's a LOT of money being poured into the AI companies. They can try to solve that problem.
What I mean is - a snake eating its own tail is bad for the snake. It will kill it. But in this case the tail is something we humans valued and don't want eaten, regardless of the health of the snake. And the snake will probably find a way to become independent of the tail after it ate it, rather than die, which sucks for us if we valued the stuff the tail was made of, and of course makes the analogy totally nonsensical.
The actual solutions suggested here are not related to it eating its own tail anyway. They're related to the sentiment that the greed of AI companies needs to be reeled in, they need to give back, and we need solutions to the fact that we're getting spammed with slop.
I guess the last part is the part that ties into it "eating its own tail", but really, why frame it that way? Framing it that way means it's a problem for AI companies. Let's be honest and say it's a problem for us and we want it solved for our own reasons.
Actually we can. And we will.
Copyright was predicated on the notion that ideas and styles can not be protected, but that explicit expressive works can. For example, a recipe can't be protected, but the story you wrap around it that tells how your grandma used to make it would be.
LLMs are particularly challenging to wrangle with because they perform language alchemy. They can (and do) re-express the core ideas, styles, themes, etc. without violating copyright.
People deem this 'theft' and 'stealing' because they are trying to reconcile the myth of intellectual property with reality, and are also simultaneously sensing the economic ladder being pulled up by elites who are watching and gaming the geopolitical world disorder.
There will be a new system of value capture that content creators need to position for, which is to be seen as a more valuable source of high quality materials than an LLM, serving a specific market, and effectively acquiring attention to owned properties and products.
It will not be pay-per-crawl. Or pay-per-use. It will be an attention game, just like everything in the modern economy.
Attention is the only way you can monetize information.
Stack Overflow peaked in 2014 before beginning it's downward decline. How is that at all related to GenAI? GPT4 is when we really started seeing these things get used to replace SO, etc., and that would be early 2023 - and indeed the drop gets worse there - but after the COVID era spike, SO was already crashing hard.
Tailwind's business model was providing a component library built on top of their framework. It's a business model that relies on the framework being good enough for people to want to use it to begin with, but being bad enough that they'd rather pay for the component library than build it themselves. The more comfortable it is to use, the more productive it is, the worse the value proposition is for the premium upsell. Even other "open core" business models don't have this inherent dichotomy, much less open source on the whole, so it's really weird to try and extrapolate this out.
The thing is, people turn to LLMs to solve problems and answer questions. If they can't turn to the LLM to solve that problem or answer that question, they'll either turn elsewhere, in which case there is still a market for that book or blog post, or they'll drop the problem and question and move on. And if they were willing to drop the problem or question and move on without investigating post-LLM, were they ever invested enough to buy your book, or check more than the first couple of results on google?
Before the LLM time there was actually the problem that google often showed SEO spam sites that harvested content from stackoverflow.
The ONLY reason we are here today is because OpenAI, and Anthropic, by extension, took it upon themselves to launch chat bots trained on whatever datasources they could get in a short amount of time to quickly productize their investments. Their first versions didn't include any references to the source material, and just acted as if they knew everything.
When CoPilot was built as a better auto-complete engine, trained on opensource projects, it was an interesting idea, because it doing what people already did. They searched GitHub for examples of the solution or nudged them in that direction. However, the biggest difference, using other project code was stable, because it came with a LICENSE.md that you then agreed to, and paid it forward. (i.e. "I used code from this project").
CoPilot initially would just inject snippets for you, without you knowing the source. It was only later, they walked that back and if you did use CoPilot, it shows you the most-likely source of the code it used. This is exactly the direction all of the platforms seem headed.
It's not easy to walk back the free-for-all system (i.e. Napster), but I'm optimistic over time it'll become a more fair, pay to access system.
I do not know what will replace it, but I will not miss websites trying to monetise my attention
In fact, this might be overall good thing, because finally original content will be highly on demand since those companies now use to train their models. But we are probably just in a transition phase.
The other thing is that new sources of input will come, from LLM usage probably, so they cut the middle layer, users input in the LLM is also a form of input, and a hybrid co-creation between users/AI would generate content at much faster rater, which again would be used to train the model, and that would improve their quality.
Oh boy I can already see what kind of articles those would be.
This is ridiculous - AI doesn't need to be fed a PDF of a Terraform book to know how to Terraform. Blowing out context with hundreds of OCR'd pages of generic text on how to terraform isn't going to help anything.
The model that is broken is really ultimately going to be "content for hire". That's the industry that is going to be destroyed here because it's simply redundant now. Actual artwork, actual literature, actual music... these things are all safe as long as people actually want to experience the creations of others. Corporate artwork, simple documentation, elevator music.... these things are done; I'm sorry if you made a living making them but you were ultimately performing an artisinal task in a mostly soulless way.
I'm not talking about video game artists, mind you, I'm talking about the people who produced Corporate Memphis and Flat Design here. We'll all be better off if these people find a new calling in life.
Like the irony is pretty deep with this one about this.
I am not sure if they could've gotten trademark from Inscryption/if they needed it but if they really wanted, I have found inscryption's ouroborous card to look the best and it was honestly how I discovered ouroborous in the first place! (became my favourite card, I love inscryption)
https://static1.thegamerimages.com/wordpress/wp-content/uplo...
Even just searching Ouroborous on internet gave me some genuinely beautiful Ouroborous illustrations (Some stock photos, some not) but even using a stock photo might have made a better idea than using AI generated Ouroboros photo itself?
I believe that the main reason for SO's decline starting around 2018 was that most of the core technical questions had been answered. There was an enormous existing corpus of accepted answers around fundamental topics, and technology just doesn't change fast enough to sustain the site. Then the LLMs digested the site's (beautifully machine-readable) corpus along with the rest of the internet and now the AIs can give users that info directly, resulting in a downward spiral of traffic to SO, fewer new questions, etc.
Vale, Stack Overflow. You helped me solve many tricky problems.
Rather we became the product.
The new world is one where someone can have an LLM assisted insight, post it on their blog for free, have it indexed by every agentic search engine, and it becomes part of the zeitgeist. That’s the new data that’ll feed the new models: a better information diet over time. And guess what else: models are getting better at identifying - at scale - the high quality info that’s worth using as training data.
But, I have to admit that a few years ago LLMs became part of my daily workflow for low-risk repetitive tasks, and sometimes even "sound boarding".
Further, the "model collapse" hypothesis of 2020/2021 seems to have failed to materialize. Maybe we're still too early, and we're not yet seeing negative effects of OpenAI training on OpenAI output. But maybe "slop" is not being rewarded as much as human content, and having humans in the loop (even as readers) is preventing a slide into incoherence.
Will LLMs eventually disincentivize people from producing and publishing new original content? If that content is easily replicated by an LLM query, maybe. And maybe it's not the worst thing in the world. 5 years ago I would have bought an "FFmpeg Cookbook" from O'Reilly, but now I would just tell Claude exactly what I'm trying to achieve. As a consumer, I'm better off, and arguably we've saved the author of a hypothetical FFmpeg Cookbook weeks out of their precious life. Weeks they could spend doing something—anything—more valuable than rewording FFmpeg documentation.
If the web gets flooded with LLM output and you train on it naively, you’re effectively training on your own prior. That pushes models toward the mean: less surprise, less specificity, more template-y phrasing. It’s like photocopying a photocopy: the sharp edges disappear.
The fix isn’t “never use synthetic data.” It’s to treat it like a controlled ingredient: tag provenance, keep a high-quality human/grounded core, filter aggressively, and anchor training to things that don’t self-contaminate (code that compiles/tests, math with verifiable proofs, retrieval with citations, real user feedback). Otherwise the easiest path is content volume, and volume is exactly what kills signal.