Anyone who has made copies of videotapes knows what happens to the quality of each successive copy. You’re not making a “treasure trove.” You’re making trash.
Uh, good.
As an engineer who cares a LOT about engineering ethics, it is absolutely fucking infuriating watching the absolute firehose of shit that comes out of LLMs and public-consumption audio, image, and video ML systems, juxtaposed with the outright refusal of companies and engineers who work there to accept ANY accountability or culpability for the systems THEY FUCKING MADE.
I understand the nuances of NNs. I understand that they’re much more stochastic than deterministic. So, you know, maybe it wasn’t a great idea to just tell the general public (which runs a WIDE gamut of intelligence and comprehension ability - not to mention, morality) “have at it”. The fact that ML usage and deployment in terms of information generating/kinda-sorta-but-not-really-aggregating “AI oracles” isn’t regulated on the same level as what you’d see in biotech or aerospace is insane to me. It’s a refusal to admit that these systems fundamentally change the entire premise of how “free speech” is generated, and that bad actors (either unrepentantly profit driven, or outright malicious) can and are taking disproportionate advantage of these systems.
I get it - I am a staunch opponent of censorship, and as a software engineer. But the flippant deployment of literally society-altering technology alongside the outright refusal to accept any responsibility, accountability, or culpability for what that technology does to our society is unconscionable and infuriating to me. I am aware of the potential that ML has - it’s absolutely enormous, and could absolutely change a HUGE number of fields for the better in incredible ways. But that’s not what it’s being used for, and it’s because the field is essentially unregulated right now.
Having now flooded the internet with bad AI content not surprisingly its now eating itself. Numerous projects that aren’t AI are suffering too as the quality of text reduces.
If we can work out which data conduits are patrolled more often by AI than by humans, we could intentionally flood those channels with AI content, and push Model Collapse along further. Get AI authors to not only vet for “true human content”, but also pay licensing fees for the use of that content. And then, hopefully, give the fuck up on their whole endeavor.
So AI:
- Scraped the entire internet without consent
- Trained on it
- Polluted it with AI generated rubbish
- Trained on that rubbish without consent
- Are now in need of lobotomy
Well duh. I think a lot of us here learned that lesson from watching the movie Multiplicity.
Would you recommend it?
Oh, shit. Ummm…it was a funny movie back when it came out, but I haven’t seen it in like 25 years so who knows how bad it seems now. Could still be good?
I couldn’t care less.
Lol
Good.
have we tried feeding them actual human beings yet ?
Billionaires are the smartest, give them the most knowledge first!
It’s like a human centipede where only the first person is a human and everyone else is an AI. It’s all shit, but it gets a bit worse every step.
Every single one of us, as kids, learned the concept of “garbage in, garbage out”; most likely in terms of diet and food intake.
And yet every AI cultist makes the shocked pikachu face when they figure out that trying to improve your LLM by feeding it on data generated by literally the inferior LLM you’re trying to improve, is an exercise in diminishing returns and generational degradation in quality.
Why has the world gotten both “more intelligent” and yet fundamentally more stupid at the same time? Serious question.
Why has the world gotten both “more intelligent” and yet fundamentally more stupid at the same time? Serious question.
Because it’s not actually always true that garbage in = garbage out. DeepMind’s Alpha Zero trained itself from a very bad chess player to significantly better than any human has ever been, by simply playing chess games against itself and updating its parameters for evaluating which chess positions were better than which. All the system needed was a rule set for chess, a way to define winners and losers and draws, and then a training procedure that optimized for winning rather than drawing, and drawing rather than losing if a win was no longer available.
Face swaps and deep fakes in general relied on adversarial training as well, where they learned how to trick themselves, then how to detect those tricks, then improve on both ends.
Some tech guys thought they could bring that adversarial dynamic for improving models to generative AI, where they could train on inputs and improve over those inputs. But the problem is that there isn’t a good definition of “good” or “bad” inputs, and so the feedback loop in this case poisons itself when it starts optimizing on criteria different from what humans would consider good or bad.
So it’s less like other AI type technologies that came before, and more like how Netflix poisoned its own recommendation engine by producing its own content informed by that recommendation engine. When you can passively observe trends and connections you might be able to model those trends. But once you start actually feeding back into the data by producing shows and movies that you predict will do well, the feedback loop gets unpredictable and doesn’t actually work that well when you’re over-fitting the training data with new stuff your model thinks might be “good.”
Another great example (from DeepMind) is AlphaFold. Because there’s relatively little amounts of data on protein structures (only 175k in the PDB), you can’t really build a model that requires millions or billions of structures. Coupled with the fact that getting the structure of a new protein in the lab is really hard, and that most proteins are highly synonymous (you share about 60% of your genes with a banana).
So the researchers generated a bunch of “plausible yet never seen in nature” protein structures (that their model thought were high quality) and used them for training.
Granted, even though AlphaFold has made incredible progress, it still hasn’t been able to show any biological breakthroughs (e.g. 80% accuracy is much better than the 60% accuracy we were at 10 years ago, but still not nearly where we really need to be).
Image models, on the other hand, are quite sophisticated, and many of them can “beat” humans or look “more natural” than an actual photograph. Trying to eek the final 0.01% out of a 99.9% accurate model is when the model collapse happens–the model starts to learn from the “nearly accurate to the human eye but containing unseen flaws” images.
good commentary, covered a lot of ground - appreciate the effort to write it up :)
Because the people with power funding this shit have pretty much zero overlap with the people making this tech. The investors saw a talking robot that aced school exams, could make images and videos and just assumed it meant we have artificial humans in the near future and like always, ruined another field by flooding it with money and corruption. These people only know the word “opportunity”, but don’t have the resources or willpower to research that “opportunity”.
Remember Trump every time he’s weighed in on something, like suggesting injecting people with bleach, or putting powerful UV lights inside people, or fighting Covid with a “solid flu vaccine” or preventing wildfires by sweeping the forests, or suggesting using nuclear weapons to disrupt hurricane formation, or asking about sharks and electric boat batteries? Remember these? These are the types of people who are in charge of businesses, they only care about money, they are not particularly smart, they have massive gaps in knowledge and experience but believe that they are profoundly brilliant and insightful because they’ve gotten lucky and either are good at a few things or just had an insane amount of help from generational wealth. They have never had anyone, or very few people genuinely able to tell them no and if people don’t take what they say seriously they get fired and replaced with people who will.
Because the dumdums have access to the whole world at the tip of the fingertip without having to put any efforts in.
In a time without that, they would be ridiculed for their stupid ideas and told to pipe down.
Now they can find like minded people and amplify their stupidity, and be loud about it.
So every dumdum becomes an AI prompt engineer (whatever the fuck that means) and know how to game the LLM, but do not understand how it works. So they are basically just snake oil salesmen that want to get on the gravy train.
oh no are we gonna have to appreciate the art of human beings? ew. what if they want compensation‽
Oh no. Anyways…
Let’s go, already!
How you can help: If you run a website and can filter traffic by user agent, get a list of the known AI scrapers agent strings and selectively redirect their requests to pre-generated AI slop. Regular visitors will see the content and the LLM scraper bots will scrape their own slop and, hopefully, train on it.
It’s kinda interesting in how it actually roughly parallels the dawn of the nuclear age in some specific ways. Namely, that there’s a clear “purity” line established by the advent of the technology - and I mean that literally, not figuratively. Content on the internet is going to have a very similar dividing line. But it’s also going to be way harder to definitively source data from before that line, unless someone clairvoyant happened to offline and archive a huge storage array with a complete internet snapshot right before ML made its public debut. And I know exactly what the scale of that storage commitment would be, and how much it would cost. So I’m certain nobody has done that.
Are there any good lists of known AI user agents? Ideally in a dependency repo so my server can get the latest values when the list is updated.
This would ideally become standardized among web servers with an option to easily block various automated aggregators.
Regardless, all of us combined are a grain of rice compared to the real meat and potatoes AI trains on - social media, public image storage, copyrighted media, etc. All those sites with extensive privacy policies who are signing contracts to permit their content for training.
Without laws (and I’m not sure I support anything in this regard yet), I do not see AI progress slowing. Clearly inbreeding AI models has a similar effect as in nature. Fortunately there is enough original digital content out there that this does not need to happen.
Regardless, all of us combined are a grain of rice compared to the real meat and potatoes AI trains on
Absolutely. It’s more a matter of principle for me. Kind of like the digital equivalent of leaving fake Amazon packages full of dog poo out front to make porch pirates have a bad day.
Well it means they need some ability to reject some content, which means they need a level of transparency they would never want otherwise.
Okay but I like using perchance cus they dont profit off anything 👉👈
a large chunk of that site is some dudes lil hobby project and its kinda neat interacting with the community and seein how the code works. Its the only bot I’ll ever use cus they arent profiting off of other people shit. the only money they get is from ads and thats it.
Dont kill me with downvotes, I like making up cool OC concepts or poses n stuff and then drawing em.
They’ll just start using a chrome user agent
Only if enough people do it. Then again, loads scrapers outside of AI already pretend to be normal browsers.
You can validate that against user telemetry data expected from a browser.
AI already long ago stopped being trained on any old random stuff that came along off the web. Training data is carefully curated and processed these days. Much of it is synthetic, in fact.
These breathless articles about model collapse dooming AI are like discovering that the sun sets at night and declaring solar power to be doomed. The people working on this stuff know about it already and long ago worked around it.
I mean, we’ve seen already that AI companies are forced to be reactive when people exploit loopholes in their models or some unexpected behavior occurs. Not that they aren’t smart people, but these things are very hard to predict, and hard to fix once they go wrong.
Also, what do you mean by synthetic data? If it’s made by AI, that’s how collapse happens.
The problem with curated data is that you have to, well, curate it, and that’s hard to do at scale. No longer do we have a few decades’ worth of unpoisoned data to work with; the only way to guarantee training data isn’t from its own model is to make it yourself
Also, what do you mean by synthetic data? If it’s made by AI, that’s how collapse happens.
But that’s exactly my point. Synthetic data is made by AI, but it doesn’t cause collapse. The people who keep repeating this “AI fed on AI inevitably dies!” Headline are ignorant of the way this is actually working, of the details that actually matter when it comes to what causes model collapse.
If people want to oppose AI and wish for its downfall, fine, that’s their opinion. But they should do so based on actual real data, not an imaginary story they pass around among themselves. Model collapse isn’t a real threat to the continuing development of AI. At worst, it’s just another checkbox that AI trainers need to check off on their “am I ready to start this training run?” Checklist, alongside “have I paid my electricity bill?”
The problem with curated data is that you have to, well, curate it, and that’s hard to do at scale.
It was, before we had AI. Turns out that that’s another aspect of synthetic data creation that can be greatly assisted by automation.
For example, the Nemotron-4 AI family that NVIDIA released a few months back is specifically intended for creating synthetic data for LLM training. It consists of two LLMs, Nemotron-4 Instruct (which generates the training data) and Nemotron-4 Reward (which curates it). It’s not a fully automated process yet but the requirement for human labor is drastically reduced.
the only way to guarantee training data isn’t from its own model is to make it yourself
But that guarantee isn’t needed. AI-generated data isn’t a magical poison pill that kills anything that tries to train on it. Bad data is bad, of course, but that’s true whether it’s AI-generated or not. The same process of filtering good training data from bad training data can work on either.
Both can be true.
Preserved and curated datasets to train AI on, gathered before AI was mainstream. This has the disadvantage of being stuck in time, so-to-speak.
New datasets that will inevitably contain AI generated content, even with careful curation. So to take the other commenter’s analogy, it’s a shit sandwich that has some real ingredients, and doodoo smeared throughout.
They’re not both true, though. It’s actually perfectly fine for a new dataset to contain AI generated content. Especially when it’s mixed in with non-AI-generated content. It can even be better in some circumstances, that’s what “synthetic data” is all about.
The various experiments demonstrating model collapse have to go out of their way to make it happen, by deliberately recycling model outputs over and over without using any of the methods that real-world AI trainers use to ensure that it doesn’t happen. As I said, real-world AI trainers are actually quite knowledgeable about this stuff, model collapse isn’t some surprising new development that they’re helpless in the face of. It’s just another factor to include in the criteria for curating training data sets. It’s already a “solved” problem.
The reason these articles keep coming around is that there are a lot of people that don’t want it to be a solved problem, and love clicking on headlines that say it isn’t. I guess if it makes them feel better they can go ahead and keep doing that, but supposedly this is a technology community and I would expect there to be some interest in the underlying truth of the matter.
