Some perspectives from someone working in the image space.
These tests don't feel practical - That is, they seem intended to collapse the model, not demonstrate "in the wild" performance.
The assumption is that all content is black or white - AI or not AI - and that you treat all content as equally worth retraining on.
It offers no room for assumptions around data augmentation, human-guided quality discrimination, or anything else that might alter the set of outputs to mitigate the "poison"
As someone also working in the imaging space, ai generated data is useful solong as it's used carefully.
Specifically, we're implementing AI culled training sets which contain some generated data that then gets reviewed manually for a few specific things, then pushed into our normal training workflows. This makes for a huge speedup versus 100% manual culling and the metrics don't lie, the models continue to improve steadily.
There may be a point where they're poisoned and will collapse, but I haven't seen it yet.
This is exactly right. Model collapse does not exist in practice. In fact, LLMs trained on newer web scrapes have increased capabilities thanks to the generated output in their training data.
For example, "base" pretrained models trained on scrapes which include generated outputs can 0-shot instruction follow and score higher on reasoning benchmarks.
Intentionally produced synthetic training data takes this a step further. For SoTA LLMs the majority of, or all of, their training data is generated. Phi-2 and Claude 3 for example.
Granted, one could argue that this only happened because the API version of Claude doesn't appear to use a system prompt. If that's the case, then the LLM lacks any identity otherwise defined by the initial system prompt, and thus, kind of makes one up.
Nonetheless, point remains, it's kind of interesting to see that in the years since the launch of ChatGPT we're already seeing a tangible impact on publicly available training data. LLMs "know" what ChatGPT is, and may even claim to be it.
that is the meat the article tries to cook. the impacts so far aren’t all that negative.
but time flows like a river, and the more shit that gets into it…
poison does not need to be immediately fatal to be fatal. some take a frighteningly long time to work. by the time you know what’s happening, not only is it too late, you have already suffered too much.
does this sound like anything more than a scary story to tell around campfires? not yet.
Claude 3 does use publically available data. Not everything is synthetically generated. Look at the section for training data in the below link. It has an quote from the paper which states that it uses a mix of public data, data from labelers and synthetic data
I can't find a link to the actual clause paper to verify the above link but a few other places mention the same thing about the training data. We don't know if this improved performance is because of synthetic data or something else. I'm guessing even antropic might not be knowing this too.
Wouldn’t reinforcement learning just weigh any nonsense data very low and then spammy garbage doesn’t really affect the model in the end much ? If the model and human experts can’t tell the difference then it’s probably pretty good AI generated data
Why would you limit a model to be like a brain in a vat? Instead let the model out so people use it, then use the chat logs to fine-tune. A chat room is a kind of environment, there is a human, maybe some tools. The LLM text will generate feedback and right there is a learning signal.
Even without a human, if a LLM has access to code execution it can practice solving coding tasks with runtime feedback. There are many ways a LLM could obtain useful learning signals. After all, we got all our knowledge from the environment as well, in the end there is no other source for knowledge and skills.
Dude what? That’s a pretty absurd claim. Most generally available models specifically curate their inputs for the express purpose of avoiding AI garbage induced collapse. It’s literally on their cited reasons for avoiding ai generated data as inputs.
This is the part that I don't really understand. Isn't this basically an evolutionary algorithm, where the fitness function is "whatever people like the most" (or at least enough to post it online)?
People rarely generate 10 pieces of content with AI and then share all 10 with the world. They usually only share the best ones. This naturally filters for better output.
Are they saying that evolutionary algorithms don't work?
> Use the model to generate some AI output. Then use that output to train a new instance of the model and use the resulting output to train a third version, and so forth. With each iteration, errors build atop one another. The 10th model, prompted to write about historical English architecture, spews out gibberish about jackrabbits.
That this happens doesn't surprise me, but I'd love to see a curve of how each organic vs machine content mixe ratio results in model collapse over N generations.
I believe that this is a non-problem pushed forward by small-scale experiments that are not representative of what people actually do with AI generation.
A lot of new content, while AI generated, has been hand picked and polished by a human (for example, while you might commit AI generated code to your codebase, you ensure that it is correct and follows your preferred style).
Content farms will push gibberish out, but they did so, and worse, before and the first generation of models was able to train on the internet anyway.
i think it's pretty much a problem and it's going to ruin any chance of high quality, original content.
look at the original internet content and what seo has done to it. google and search in general results are trash nowadays. this is what genAI is going to do over long term. garbage in garbage out.
You'd think we'd be concerned about it poisoning the culture, well before any concerns that it would start to interfere with the rich continuing to be able to profit from it doing so.
I think it's interesting that human minds generally (though not always!) improve when exposed to the output of other human minds. It seems to be the opposite for current LLMs.
Maybe it's less about "Human VS Robot" and more about exposure to "Original thoughts VS mass-produced average thoughts".
I don't think a human mind would be improving if they're in a echo-chamber with no new information. I think the reason the human mind is improving is because we're exposed to new, original and/or different thoughts, that we hadn't considered or come across before.
Meanwhile, a LLM will just regurgitate the most likely token based on the previous one, so there isn't any originality there, hence any output from a LLM cannot improve another LLM. There is nothing new to be learned, basically.
> Humans are very capable of looking around themselves and thinking "I can do better than this"
Doesn't this require at least some perspective of what "better than this" means, which you could only know with at least a bit of outside influence in one way or another?
humans haven’t been had the same set of all encompassing “training experiences” like LLMs have. we each a subset of knowledge that may overlap with some other’s knowledge, but is largely unique. so when we interact with each other we can learn new things, but with LLMs I imagine it is a group of experienced but antiquated professors developing their own set of out of touch ideas
A sequence of AI models trained on each other's output gets mutations, which might help or hurt, but if there's one dominant model at any given time then it's like asexual reproduction with only living descendant in each generation (and all the competing models being failures to reproduce). A photocopy of a photocopy of a photocopy — this seems to me to also be the incorrect model which Intelligent Design proponents seem to mistakenly think is how evolution is supposed to work.
A huge number of competing models that never rise to dominance would be more like plants spreading pollen in the wind.
A huge number of AI there are each smart enough to decide what to include in its training set would be more like animal reproduction. The fittest memes survive.
Memetic mode collapses still happen in individual AI (they still happen in humans, we're not magic), but that manifests as certain AI ceasing to be useful and others replacing them economically.
A few mega-minds is a memetic monoculture, fragile in all the same ways as a biological monoculture.
A different biological analogy occurred to me which I've mentioned
before in a security context. It isn't model degeneration but the
amplification of invisible nasties that don't become a problem until
way down the line.
Natural examples are prions such as Bovine spongiform encephalopathy
[0] or sheep scrapie. This seems to really become a problem in systems
with a strong and fast positive feedback loop with some selector. In
the case of cattle it was feeding rendered bonemeal from dead cattle
back to livestock. Prions are immune to high temperature removal so
are selected for and concentrated by the feedback process.
To really feel the horror of this, read Ken Thompson's "Reflections on
Trusting Trust" [1] and ponder the ways that a trojan can be replicated
iteratively (like a worm) but undetectably.
It isn't loss functions we should worry about. It's gain functions.
Have you ever heard of the telephone game? This is what is going on here. Or imagine an original story of something that really happened. If it goes by 100 people in a chain, how much do you think the story will resemble the original one?
I mean it makes sense that (even impressively functional) statistical approximations would degrade when recursed.
If anything I think this just demonstrates yet again that these aren't actually analogous to what humans think of as "minds", even if they're able to replicate more of the output than makes us comfortable.
Humans exhibit very similar behavior. Prolonged sensory deprivation can drive a single individual insane. Fully isolated/monolithic/connected communities easily become detached from reality and are susceptible to mass psychosis. Etc etc etc. Humans need some minimum amount of external data to keep them in check as well.
I watched someone in the printer room at the computer science department gradually photocopy from white to black, and back again, over the span of 300 pieces of paper, by altering the thresholds of the photocopyer.
They didn’t graduate to become computer scientists, but did indeed get admitted to the royal school of art the year after.
It's fascinating that error can accumulate through repeated trainings that 1) is undetected by humans and 2) can degrade LLM or diffusion models (or any transformer model?) so completely. This implies that not only do we not understand how latent knowledge is actually representated in deep nets, we don't know it forms or how it changes during training. If we did, we could have predicted the destructive impact of recycling of output as input. IMO, this suggests we should demand rigorous validation of deep nets (especially generative ones) before relying on them to behave responsibly.
The effect is not new. We have known about it ever since we've had basic machine learning. The way to look at it is somewhat novel but not surprising at all.
I think AI-generated images are worse for training AI generative models than LLMs, since there are so many now on the internet (see Instagram art related hashtags if you want to see nothing but AI art) compared to the quantity of images downloaded prior to 2021 (for those AI that did that). Text will always be more varied than seeing 10m versions of the same ideas that people make for fun. AI text can also be partial (like AI-assisted writing) but the images will all be essentially 100% generated.
That's far from unique to instagram. I loathe Stable Diffiusion and co solely because they've utterly FLOODED every cool art-adjacent website with endless mediocre derivative shit. Like there was always low-effort content of course, but holy fuck, there is SO MUCH MORE now. And some of these people are trying to CHARGE for this uninspired junk!!!
I agree with this despite using SD a lot myself. It's fun to use until you realize the majority of people posting stuff generated with it have almost no creativity, all generating the same things over and over again, mostly without any manual work involved. that uncanny realism style with the generic Stable Diffusion face and one of 5 different poses. The number of people putting any sort of effort into it is way, way lower than the number of users thinking they are making art. It's more of a slot machine in the majority of cases
Unfortunately, yeah 99.9% of images you're going to see generated from stable diffusion models are going to be either selfies, portraits, or porn.
What's you're not going to see is things like "a divine gigantic textile loom sewing together a white horse and a black horse in an interlaced pattern to create a zebra." for example.
Honestly, if you want to make art, AI is a hindrance not a tool. You have to engineer what you say to maximum exactness, and even then it'll still just ignore certain words, or skip them over, or get basic details wrong.
Way back when all this stuff first popped off, I did try it out. I was unimpressed. It was like playing a game of telephone with my ideas, having to describe them into one end and have a thousand people repeat it to one another and make little contributions till it came out the other end, most of the time looking absolutely nothing like I expected.
People who say this makes art accessible... I dunno, I've never gotten it. I've seen people with all manner of disabilities, deformities, etc. all manage to express themselves creatively with practice and accessibility tools far more reliably than trying to make an AI pop out what you actually want. It seems to be the accessibility claims really only hold water if the accessibility feature is "I don't want to learn any skills" which... I mean, okay. But as with all art, your end product will reflect that level of care.
I definitely think the flooding of art spaces is hugely problematic, but it is pretty funny to watch people try to "be an artist" by putting essentially no effort in. It definitely points to a lack of understanding in the field when all these people are basically generating a ton of images that are all derived from the same models. There's a lack of understanding of supply and demand, when the expectation is that your ai illustration that you made in like an hour with the same software as every other ai artist is that it's somehow going to be competitive on engagement with an original piece from an artist who has an audience. There's a lot of demand for artists like Mika Pikazo and Frank Frazetta, not the 100,000 ai artists out there.
I mean it's hard to fault those people when that was essentially how these were sold way back. "Automating art" and all that. All the most insufferable people on twitter jumping from shilling crypto scams to shilling AI and telling real artists in their ivory towers that their days were numbered.
Guess put it on the pile with all the other broken promises.
There's no such thing as a world model. People do not have world models.
This is a confused term made up by 70s AI researchers, who had the continual problem that they didn't know any philosophy and kept making up their own metaphors for how intelligence might work, and then deciding that because they'd made it up it must be true, and also that if they wrote a computer program that had the same metaphors it must work.
"World model" just vaguely points at something people might do and assumes that if you make up a new thing it vaguely points at it'd help.
On what basis are you saying that? I have a model in my head mapping out the world around me, so I know where things are etc and what I can do with all those things. How is that not a world model? Are you using a very strange definition of "world model"?
> I have a model in my head mapping out the world around me, so I know where things are etc and what I can do with all those things.
No you don't; a map is not the territory, and is necessarily wrong, which means that if you had such a model and were actually relying on it you wouldn't be able to do things you obviously can do in real life.
You have an inaccurate memory of the world and you update it, only as much as you need to[0], as you go, in order to do a specific task.
[0] probably a little less than you need to, because you want to save thinking energy
> No you don't; a map is not the territory, and is necessarily wrong, which means that if you had such a model and were actually relying on it you wouldn't be able to do things you obviously can do in real life.
What are you talking about, a world model doesn't need to be perfect, nobody said humans has a perfect world model just that we have a world model.
Yes we update the model and adjust when it is wrong, but we still have a world model we use to plan out actions before we do them. I can very accurately predict all the events that will happen when I cook food, how the water will flow etc, sometimes things go wrong and I adjust and fix, but there is no way I could cook food if I didn't have that world model to plan out actions.
> This is a longstanding critique of GOFAI; see Hubert Dreyfuss and Phil Agre.
They don't even mention world model there, I think you are talking about something completely different. No, the mental model humans have of the world isn't the real world, we know that, it is a model of the world, ie a world model. A model isn't the real thing, that is why we call it a model.
I also wonder what search engines are going to do about all this. Sounds to me, actually, traditional, non-intelligent search might be on its way out, although of course it'll take time. Future search engines will have to be quite adept at trying to figure out whether the text they index is bullshit or not.
reminds me of sheep and cows being fed their bretherens own brain matter developing spongiform encepalopathy (brain disease) or of course cannibals developing kuru. except a purely 'software' form.
Is there a standard objective metric that can help determine that the quality of a model has degraded over time. In that case, much like source code, you just revert to the old version.
I don't remember wich YouTuber made a interesting video about it but basically communities are moving away from the free web in private communities (think discord or even sites that you are forced to register to to read the content)
It's an interesting thing but I think queries on searche engines are becoming worse for this reason too.
I'm not sure how much of a risk this is to LLMs in particular, but I feel like we're already seeing the impact on image AI models.
Even though they're getting better at generating hands that make sense and other fine details, you can generally tell that an image is AI generated because it has a certain "style". Can't help but wonder if this is partly due to generated images contaminating the training data and causing subsequent AI image generators to stylistically converge over time.
It's because the models don't have an optimal aesthetic policy. Which would be difficult, but if they did have one, it wouldn't matter how much bad input data you added during pretraining.
In human society, a feedback loop of nonsense is usually defeated by practical effects in physical reality and experience. The objective of education, for example, is to transmit knowledge and apply reason to important questions.
In manipulated social media, there is no check on the nonsense loop. The technology that we currently call A.I. could be used for educational good.
How it will be used, however, is likely to further distort discourse and generate nonsense.
It is worse, because it is faster - how many incorrect blog articles can a sigle typical writer publish and post on the internet - maybe 1-2 a day if you are a prolific writer?
How many can an AI agent do? Probably hundreds of thousands a day. To me, that is going to be a huge problem - but don't have a solution in mind either.
And then those 100K bad articles posted per day by one person, are used as training data for the next 100K bad/incorrect articles etc - and the problem explodes geometrically.
Imagine you have a calculator that outputs a result that is off by one percent. That's ai right now.
If you use the results of each calculation in additional calculations, the result will skew further and further from reality with each error. That's ai training on itself.
No, LLMs with soft attention use compression, and actually has no mechanism for ground truth.
They are simply pattern finding and matching.
More correctly, they are uniform consent depth threshold circuits.
Basically parallel operations on a polynomial number of AND, OR, NOT, and majority gates.
The majority gates can do the Parity function, but cannot self correct like ECC does.
The thing with majority gates is that they can show some input is in the language:
This the truthiness of 1,1,1,0,0 being true, but 1,1,0,0,0 would be failure as negation, but doesn't prove that negation, it isn't a truthy false.
With soft attention will majority gates they can do parity detection but not correction.
Hopefully someone can correct this if I am wrong.
Specifically I think that the upper bound of deciding whether X = x is a cause of m) in structures is
NP-complete in binary models (where all variables can take
on only two values) and Σ_2^P
-complete in general models.
As TC_0 is smaller than NP, and probably smaller than P, any methods would be opportunistic at best.
Preserving the long tail of a distribution is a far more pragmatic direction as an ECC type ability is unreasonable.
Thinking of correctional codes as serial turing machine and transformers as primarily parallel circuits should help with understanding why they are very different.
You can verify a mathematical result. You can run the calculations a second time on a separate calculator (in fact some computers do this) to verify the result, or use a built in check like ecc.
There's no such mathematical test for truth for an ai to run.
Error correction doesn’t insure truth. At least in communication, it insures that the final version matches the original version.
For AI, you wouldn’t be doing EC to make sure the AI was saying truth, you would be doing EC to ensure that the AI hasn’t drifted due to the 1% error rate.
Of course I have no idea how to actually do it - if it isn’t being done now, it is probably hard or impossible.
There's no fully general test for truth for an AI to run.
In some specific domains such tests exist — and the result is, generally, computers wildly outperforming humans. But I get the impression from using them that current LLMs didn't take full advantage of this during training.
These tests don't feel practical - That is, they seem intended to collapse the model, not demonstrate "in the wild" performance.
The assumption is that all content is black or white - AI or not AI - and that you treat all content as equally worth retraining on.
It offers no room for assumptions around data augmentation, human-guided quality discrimination, or anything else that might alter the set of outputs to mitigate the "poison"