Read the article, that's not what this is about. Title is clickbait, they fired contractors for using AI to label data when the whole point of labeling data is to distill human knowledge into weights, not distill weights into weights.
You're replying to a joke, and yes, that's all it's about. For all the tales of rapid self-improvement, all labs are real careful not to taint their precious training sets with anything that comes out of their systems. Website reputation modeling, fingerprints in generated text, firing contractors that rely on AI.
I've been told on HN about a year ago that the era of scraping is over and that it's all AI-training-AI now. My web server logs and stories like that disagree.
There's a difference between shoving LLM output back into the input and RSI more broadly. Using synthetic training data didn't work super well, that would have been one path to RSI, but it doesn't work. The main path towards RSI people are talking about today is using the models to do research and experimentation for training new models. Machine learning is a very empirical field, you need to run tons of experiments, tune hyperparameters, and try different architectures. It's something agents are extremely good at doing, just because one path to RSI doesn't work doesn't mean it won't work at all.
I've been told on HN that 1+1=3, so, yes, skepticism is required on broad claims.
There are more subtle truths on this. Things that can be proven algorithmically are much more apt to be in recursive AI loops now. Hence things like programming and hacking keep improving steadily over time with much less human training data being added.
Think of labeling as establishing a ground truth. It's more important than training, but it's far less complex than training at an individual level. A kid can tell you what an ice cream cone looks like, but they cannot tell you the best algorithm to use to get the best model with the least power usage.
Labeling is more like working a checkout at Walmart. Just about anyone can do it with the smallest amount of training, but you have to ensure your labelers are not just scanning one item multiple times and bagging up the rest as your dataset can skew from reality since AI cannot just capture this data fully reliably at this point (well in many fields it can or can do even better than humans, but it's still lumpy as to where and why).
Bubble bursting behavior in action - AI company that pushes everyone to use AI, replace everything and everyone with AI, fires people for using AI to make the AI.
Earlier this year it was so dissonant to see so many companies explaining how they are going to replace their engineers with AI from both OpenAI and Anthropic, while both companies were instead (and are still) hiring all the human talent they could
Typically, for AI training, you need human feedback. If AI were enough, then OpenAI would do it themselves. Though we all know how bad AI slop is, running it in a loop can lead to unreadable sentences.
They hired contractors on the condition that they provide human feedback without AI; those people broke the rules, so their contracts ended prematurely.
Not sure why this is news uncovered by an investigative journalist.
"Percentage of humans that read headlines" > "Percentage of humans that read articles"
Time and time again I see headlines that are designed to catch eyeballs and are totally refuted by the article. The problem is meat statocastic parrots read the headline and hallucinate their version of the story that is most likely.
“Suppose the government puts a certain drug in the water supply … A couple of conspiracy nuts say it makes your fingers fall off one by one, but the government says that’s ridiculous … However, government employees are all observed drinking bottled water exclusively, and if anyone suggests that government employees might also want to take the completely innocuous drug, they freak out … If by chance you manage to slip a little bit of tap water into a government employee’s drink, and he finds out about it, he runs around shrieking like a banshee and occasionally yelling “AAAAAAH! MY FINGERS! MY PRECIOUS FINGERS!”. At some point you might start to wonder whether the government was being entirely honest with you.”
What is it, exactly, that makes AI-generated text so poisonous to AIs but totally harmless to humans? What is mode/model collapse, and why can it only happen to AIs with too much AI text in their training data and not to, say, human students with too much AI text in their textbooks? The people who know the most about this phenomenon seem much more careful about contamination than they are encouraging us to be.
I don't think anyone is saying it's poisonous here? The people were hired to do a job and they didn't do it. Feels like you're being disingenuous, if there's one thing i know about AI employees, it's that they use LLMs, like, all the time.
Are you having trouble following the metaphor? In it, I would be the conspiracy theorist saying it’s poisonous, and OpenAI would be the government saying it’s harmless but it’s also a fireable offense to feed it to us and we’ve hired a second set of contractors to watch for any violation of this rule.
(I’m not trying to be disingenuous, though I am trying to express a niche viewpoint. Hopefully, my willingness to take the metaphorical role of conspiracy theorist is understood as epistemic humility.)
Nope, not having trouble following it, but the metaphor leaves out an important link between AI output and the AI itself. The question is, did the AI gain anything from this training? And the answer is, it did not, since the training used it's own answer with no other input. There's nothing in the water metaphor that draws this line.
Maybe a better metaphor is something involving drinking your own piss, but I don't feel the need to flesh that one out
So you’re saying LLMs are feeding us their piss instead of water
Well dang. That’s a problem ain’t it.
GPT-5 came and went, and we’re still here adding decimals to the model numbers hoping this fundamental fuckup will go away
Feels like everyone, including you and the guy you’re arguing with, and the researchers at openAI, knows this is a problem. Is it laziness? Lack of creativity? Are we seriously all out of ideas other than scaling compute?
When are we planning on figuring this one out boys. Who’s actually working on this today
Religions are an example of a system feeding back into itself and going off the rails. This is not an AI problem, this is the problem of any evolving system.
In humans this problem is solved by dying. If you get too stoooopid, well you get culled by your own stupidity. Luckily we run in a massively parallel fashion and those that aren't fatally dumb carry on.
This is also why human civilization fundamentally changed (not humans, our civilization) after wide adoption of the scientific method. Humanities growth in relation to our potential intelligence level was horrifically slow. It's really easy to say "hur hur, machine dumb" when we could have had rockets 10,000 years ago if we weren't dumb ourselves.
Because this isn't the case, training AI on its own output is more akin to what you get from pointing a camera at a TV showing its own output. Its a feedback loop that causes noise to blow up out of proportion and leaves nothing meaningful behind but a self-recursive loop. This is "what makes ai generated text poisonous to ais".
It's not a clickbait title? Click bait has a very different meaning. The click bait title would be something like, "Open AI mysteriously fired these workers, find out why"
This title says why. The article then goes on to explain why that's a problem. Pretty straightforward.
Is your objection to the word “train” in headline? Otherwise, you’re just restating the headline while calling it clickbait (which, btw, it’s not, perhaps you meant to say it is misleading, which is a different thing.)
labeling is more specific than training, and changes the implied situation.
Its the same difference between "I was arrested for having liquid in my car while driving" and "I was arrested for holding an open bottle of whiskey while driving"
Yeah, that's a pretty big difference. The people "training" OpenAI models are getting paid like pro football players, have PhDs in the field, and are the superstars of the company. The people "labeling" data are low level contractors that get paid basically nothing to read chat responses and grade them, anyone who can type can do it.
I don't know if it's deliberately misleading or if the journalist doesn't understand the difference, but the output is functionally the same.
Why exactly it is clickbaity? It makes perfect sense to me, I understand title as: people told not use AI fired for using AI - without any need for additional context.
Mechanical Turk was shut down for exactly same reason, it was always a race to the bottom and using LLMs was cheaper than even third world gig workers. Verifying if task were done by humans is probably harder than doing them in the first place.
As I see it, the distinction is between the work of designing and managing the training process, and the work of providing human training data. It's generally considered a good idea to automate the former, but there's a conceptual (information theoretical) issue with automating the latter. I don't know if it makes things more fair, but it's not quite a contractor vs FTE issue.
Really? It’s clickbait because it paints OpenAI as hypocritical: “look they’re marketing AI as valuable, and they won’t even let their own people use it”. When in fact the job description is to not use AI.
Please reread. I am not making the claim that OpenAI is hypocritical.
I am making the claim that the article headline is clickbait because it paints OpenAI as hypocritical for firing workers for using AI to do their job, which is a mischaracterization because the whole job is to give human input.
So OpenAI is not a believer in recursive self-improvement after all?
I think people want the service they are paying for.
Read the article, that's not what this is about. Title is clickbait, they fired contractors for using AI to label data when the whole point of labeling data is to distill human knowledge into weights, not distill weights into weights.
You're replying to a joke, and yes, that's all it's about. For all the tales of rapid self-improvement, all labs are real careful not to taint their precious training sets with anything that comes out of their systems. Website reputation modeling, fingerprints in generated text, firing contractors that rely on AI.
I've been told on HN about a year ago that the era of scraping is over and that it's all AI-training-AI now. My web server logs and stories like that disagree.
There's a difference between shoving LLM output back into the input and RSI more broadly. Using synthetic training data didn't work super well, that would have been one path to RSI, but it doesn't work. The main path towards RSI people are talking about today is using the models to do research and experimentation for training new models. Machine learning is a very empirical field, you need to run tons of experiments, tune hyperparameters, and try different architectures. It's something agents are extremely good at doing, just because one path to RSI doesn't work doesn't mean it won't work at all.
I wonder if that's why AI sounds so wooden - so that they can clearly distinguish AI generated text from manmade.
I've been told on HN that 1+1=3, so, yes, skepticism is required on broad claims.
There are more subtle truths on this. Things that can be proven algorithmically are much more apt to be in recursive AI loops now. Hence things like programming and hacking keep improving steadily over time with much less human training data being added.
That is exactly what "training AI" needs: labeled data.
What is the other way to interpret the title in your mind?
3rd party contractor hired to generate human labels uses AI generate labels, gets fired. "Training" is a lot different than "data labeling".
Why is it a lot different? Is the labeling less important or something?
Think of labeling as establishing a ground truth. It's more important than training, but it's far less complex than training at an individual level. A kid can tell you what an ice cream cone looks like, but they cannot tell you the best algorithm to use to get the best model with the least power usage.
Labeling is more like working a checkout at Walmart. Just about anyone can do it with the smallest amount of training, but you have to ensure your labelers are not just scanning one item multiple times and bagging up the rest as your dataset can skew from reality since AI cannot just capture this data fully reliably at this point (well in many fields it can or can do even better than humans, but it's still lumpy as to where and why).
They’re a believer, but they won’t be paying third parties for that once it becomes feasible
But isn't the point of recurive self-improvement that you can fire people training your AI?
Correct, hence we don't have RSI yet. RSI'ing before you have RSI doesn't make a workable RSI.
Bubble bursting behavior in action - AI company that pushes everyone to use AI, replace everything and everyone with AI, fires people for using AI to make the AI.
The more you think about it, the worse it gets.
Earlier this year it was so dissonant to see so many companies explaining how they are going to replace their engineers with AI from both OpenAI and Anthropic, while both companies were instead (and are still) hiring all the human talent they could
Fires people for using AI to label data. That is something that has always been clear from the start is something that needs to be done by humans.
Makes sense, they are buying distillation of human brains not their own weights.
Typically, for AI training, you need human feedback. If AI were enough, then OpenAI would do it themselves. Though we all know how bad AI slop is, running it in a loop can lead to unreadable sentences.
They hired contractors on the condition that they provide human feedback without AI; those people broke the rules, so their contracts ended prematurely.
Not sure why this is news uncovered by an investigative journalist.
How is it a failure of journalistic understanding? They seem to say pretty much the same as you
"Percentage of humans that read headlines" > "Percentage of humans that read articles"
Time and time again I see headlines that are designed to catch eyeballs and are totally refuted by the article. The problem is meat statocastic parrots read the headline and hallucinate their version of the story that is most likely.
“Suppose the government puts a certain drug in the water supply … A couple of conspiracy nuts say it makes your fingers fall off one by one, but the government says that’s ridiculous … However, government employees are all observed drinking bottled water exclusively, and if anyone suggests that government employees might also want to take the completely innocuous drug, they freak out … If by chance you manage to slip a little bit of tap water into a government employee’s drink, and he finds out about it, he runs around shrieking like a banshee and occasionally yelling “AAAAAAH! MY FINGERS! MY PRECIOUS FINGERS!”. At some point you might start to wonder whether the government was being entirely honest with you.”
What is it, exactly, that makes AI-generated text so poisonous to AIs but totally harmless to humans? What is mode/model collapse, and why can it only happen to AIs with too much AI text in their training data and not to, say, human students with too much AI text in their textbooks? The people who know the most about this phenomenon seem much more careful about contamination than they are encouraging us to be.
I don't think anyone is saying it's poisonous here? The people were hired to do a job and they didn't do it. Feels like you're being disingenuous, if there's one thing i know about AI employees, it's that they use LLMs, like, all the time.
Are you having trouble following the metaphor? In it, I would be the conspiracy theorist saying it’s poisonous, and OpenAI would be the government saying it’s harmless but it’s also a fireable offense to feed it to us and we’ve hired a second set of contractors to watch for any violation of this rule.
(I’m not trying to be disingenuous, though I am trying to express a niche viewpoint. Hopefully, my willingness to take the metaphorical role of conspiracy theorist is understood as epistemic humility.)
Nope, not having trouble following it, but the metaphor leaves out an important link between AI output and the AI itself. The question is, did the AI gain anything from this training? And the answer is, it did not, since the training used it's own answer with no other input. There's nothing in the water metaphor that draws this line.
Maybe a better metaphor is something involving drinking your own piss, but I don't feel the need to flesh that one out
So you’re saying LLMs are feeding us their piss instead of water
Well dang. That’s a problem ain’t it.
GPT-5 came and went, and we’re still here adding decimals to the model numbers hoping this fundamental fuckup will go away
Feels like everyone, including you and the guy you’re arguing with, and the researchers at openAI, knows this is a problem. Is it laziness? Lack of creativity? Are we seriously all out of ideas other than scaling compute?
When are we planning on figuring this one out boys. Who’s actually working on this today
Religions are an example of a system feeding back into itself and going off the rails. This is not an AI problem, this is the problem of any evolving system.
In humans this problem is solved by dying. If you get too stoooopid, well you get culled by your own stupidity. Luckily we run in a massively parallel fashion and those that aren't fatally dumb carry on.
This is also why human civilization fundamentally changed (not humans, our civilization) after wide adoption of the scientific method. Humanities growth in relation to our potential intelligence level was horrifically slow. It's really easy to say "hur hur, machine dumb" when we could have had rockets 10,000 years ago if we weren't dumb ourselves.
It’s a good metaphor, your point is valid
He’s being too literal
Because this isn't the case, training AI on its own output is more akin to what you get from pointing a camera at a TV showing its own output. Its a feedback loop that causes noise to blow up out of proportion and leaves nothing meaningful behind but a self-recursive loop. This is "what makes ai generated text poisonous to ais".
Clickbait title, they fired contractors for using AI to label data, which defeats the purpose of labeling data in this case.
It's not a clickbait title? Click bait has a very different meaning. The click bait title would be something like, "Open AI mysteriously fired these workers, find out why"
This title says why. The article then goes on to explain why that's a problem. Pretty straightforward.
Is your objection to the word “train” in headline? Otherwise, you’re just restating the headline while calling it clickbait (which, btw, it’s not, perhaps you meant to say it is misleading, which is a different thing.)
labeling is more specific than training, and changes the implied situation.
Its the same difference between "I was arrested for having liquid in my car while driving" and "I was arrested for holding an open bottle of whiskey while driving"
Yeah, that's a pretty big difference. The people "training" OpenAI models are getting paid like pro football players, have PhDs in the field, and are the superstars of the company. The people "labeling" data are low level contractors that get paid basically nothing to read chat responses and grade them, anyone who can type can do it.
I don't know if it's deliberately misleading or if the journalist doesn't understand the difference, but the output is functionally the same.
Why exactly it is clickbaity? It makes perfect sense to me, I understand title as: people told not use AI fired for using AI - without any need for additional context. Mechanical Turk was shut down for exactly same reason, it was always a race to the bottom and using LLMs was cheaper than even third world gig workers. Verifying if task were done by humans is probably harder than doing them in the first place.
Training is far more broad than labeling
"People training OAI AI" has a colloquial meaning of "People working at OAI"
So the headline is built to imply:
"People who work for OAI got fired for using AI to do their job"
When a non-clickbait headline would be
"3rd party workers contracted to distill their human knowledge fired for using AI instead"
Not nearly as "clicky"
Just make it "contractors training..." And then it's fine?
As I see it, the distinction is between the work of designing and managing the training process, and the work of providing human training data. It's generally considered a good idea to automate the former, but there's a conceptual (information theoretical) issue with automating the latter. I don't know if it makes things more fair, but it's not quite a contractor vs FTE issue.
All data labeling is training, but not all training is data labeling. In this sense, the headline is correct, I don’t get the nitpicking.
Some people would die of dehydration if they read the headline "H2O is a chemical that often kills people".
Really? It’s clickbait because it paints OpenAI as hypocritical: “look they’re marketing AI as valuable, and they won’t even let their own people use it”. When in fact the job description is to not use AI.
I did not get the hypocritical implication from the title and I don't think it's necessarily implied.
This is an odd take. There are many kinds of problems that must be solved using a different system than the first system that got the answer to ensure
1. That system isn't broken in a difficult to tell way.
2. The systems implementation is incorrect.
3. That the theory that particular implementation uses isn't incorrect.
There is nothing hypocritical at all about that.
Please reread. I am not making the claim that OpenAI is hypocritical.
I am making the claim that the article headline is clickbait because it paints OpenAI as hypocritical for firing workers for using AI to do their job, which is a mischaracterization because the whole job is to give human input.
Oh, sorry, I read it wrong.
So by using AI to label datasets, the data used to train the AI was generated by AI… and we arrive at the title of the article
yo dawg
pimp my agent
I heard you like AI