It reads to me that locking yourself / AI up in a room with only yourself to converse with causes it to go MAD / out of touch with reality.
If you interact with people / fresh data regularly you'll avoid that.
I'm not sure how much is down to evolving requirements and data formats vs degradation of the training data.
The idea sounds common-sense but it's also triggering a bullshit warning for me, because I'd expect more consistency per AI model, and differences to occur on new model interations - regardless of the dataset.
milk is unfit for some humans though and I believe excessive AI interaction will come with it's own set of issues for parts of our society. I am of the AI companion startups, that shit can't and won't end pretty.
well it not really a good analogy because pigs are very robust and have existed a long time and 'AI'(LLMs) are very brittle fragile embryos that can die from a sneeze.
a more relevant analogy for llms is a bullshiter that don't know shit about anything but like talk about everything and they're good at that (talking like they know), I'm sure we all at least know or knew in the past someone like that.
at a low enough percentage these bullshiters can thrive and get praise from other people but when there a lot of them and not enough of the real deal, well everyone is in trouble including the bullshiters since they no longer have who to mimic.
As a relative statement, I'm not sure if AI or humans are better or worse (or even "it depends on what you're measuring"), but in an absolute sense humans are also pretty suggestible and easy to influence, and it's only countervailing influences pushing in other directions that turn it into noise.
This space is where the Venn diagram of marketing and genocides overlap.
It's more like feeding humans to humans or cows to cows. They go mad too. Feeding cows to humans instead typically works, of course there are ALWAYS exceptions and it's never perfect or without harms. Both with AI and cows.
Humans eating human meat making people go crazy is actually a myth I think. The reason people thought it was a thing was because of a small group of Indonesian (I think) indigenous people who ate dead relatives as part of funeral rites. Only children and women were getting sick with a "laughing disease" because they ate the dead. The underlying cause was a prion disease that could be carried without symptoms and that just ripped through their population over a few generations. That account is why people think cannibalism is bad for you, but if I recall correctly, prion diseases are not actually widespread enough for this to be a major concern. In terms of health, most human meat should be safe for consumption.
Obviously ethically we probably should not normalize eating people though.
But we teach humans from outside sources as well as using local cyclical information passing. Things they aren't already familiar with (especially during the early years when practically everything is new), things from other cultures, new science results, etc.
Training AI on AI output is less like organised education systems and the less formal written & oral traditions that predate them, and more like dark reddit/chan/other communities eating each other's rhetoric, or some political¹ and religious groups in the wider world. These relatively insular (but sometimes large) groups often descend into a pit they have difficulty reasoning themselves out of.
Maybe the answer for AI is to work out what sort of external mixing helps to keep humans on track, and if it can be emulated in the training models used, so their reasoning continues to grow instead of falling into this pothole.
This might be a crap idea (too little sleep, caffeine has just made me tired and jittery) but perhaps the scale of the information we pile into the process and the fact the retraining at least partly on AI output seems inevitable, means we need to look at training an AI more like training a small population of humans rather than a single unit human.
----
[1] mostly the far right, but that might in part be because they tend to be loud so noticed – far anything is a problem
I usually get told not to be the other way around: dehumanising us! We are are essentially an advanced pattern matching system that happens to have developed higher order reasoning, and not as special as many like to credit us with being. Our evolution and programming methods differ from those of these "AI" models of course, but we and those are essentially information processing systems and there may be some similarity (so similar failure modes) in the respective "learn, do, relearn" feedback loops despite the obvious differences.
The same thing sort of happens with humans teaching humans though: you get things like linguistic drift where the structure of language and the meaning of words changes over time, and artistic styles and musicial fashions gradually shift as well. What's different here is that there's an ouside reality of what (for example) faces look like which human perception is grounded in but AI training doesn't have access to unless it's added to the training data. This means we also probably shouldn't worry about humans consuming AI generated stuff having the same effect, since there's no reason to think it should be any worse than humans consuming human-produced stuff.
You should better compare it with humans consuming stuff they created themselves. Or in a perhaps more sickening way, an author who only reads what they have written themself.
We know that there is more than one AI software but there is not as much variation as for humans.
And as someone else already wrote here, humans have something like a ratcheting mechanism. I would paraphrase it as "Humans have learnt to stand on shoulders of giants". AI does not have this.
Perhaps it means cultures that look inward and constantly recycle their own tropes are more likely to be unstable than cultures that look outwards and have a positive attitude to novelty.
Whether it's an AI eating its own training data, a human stuck in an empty cell and forced to live with just his/her own thoughts, or a culture believing its own myths and aggressively punishing insiders and outsiders who challenge them - it's all solitary confinement. And we know how that goes.
> If feeding AI generated stuff to train an AI is bad then what does that tell us about humans consuming AI generated stuff?
It's not unqualified bad. It is bad if you do it for many times in a loop, and without any external inputs. But you can put fresh material right in the prompt and it won't suffer from retraining on synthetic data.
If you train a small model from scratch on purely synthetic data, like the Microsoft Phi models, it comes out competent and 5x more efficient than models trained on web text. So there's the flip side. You can do it in bad way, and you can do it in a good way.
the problem isn't AI generated stuff, the problem is the models own AI generated stuff. If different models cover domains the current model doesn't do well in a particularly effectively, then there is plenty of merit for information to exchange between these models.
There is a difference to talking only to yourself and talking with other people.
It's articles like this that makes me think it's plausible that we are a small N iterations away from stumbling into a transformer(/newfangled thing) architecture that rivals the reasoning capacity of humans.
Perhaps some madness in humans (and other animals with sufficiently complex brains) is part of how we grow our reasoning capabilities. Like a fighter jet that is designed to be a little unstable because that makes them more manoeuvrable with computer aided controls (a stable design will tend to resist changes in flight path more), going off-track is part of how we have ideas: random crap fires in our brains and something triages it, keeping just the useful stuff, or if the less useful stuff gets more attention it is recorded as “let's not do that sort of thing” for future reference. Sometimes those random things are intrusive thoughts (“I wonder what would happen if I snapped and pushed him down those stairs”) sometimes they are bright ideas (“hmm, what if the ground was attracted to the Apple as well?”): the distinction isn't how they are made it is whether the end result would be good or bad overall. Madness in people is sometimes when the control structures stop working to differentiate properly between the two sorts of thoughts, and a dullness in reasoning can be due to these controls being too stringent. Perhaps a future step forward in training AI models for better reasoning and/or more general reasoning is find a way to better emulate this sort of triaged randomness, and some AIs having problems is something we end up accepting as much as we accept natural intelligences sometimes having problems that look similar?
This is, among other things, a very natural consequence of some of the equations surrounding and involved in Shannon's original noisy channel capacity theorem, where the noise is (in many ways) conditioned upon the structure of the model itself.
It is not at all necessarily surprising, I think, from a purely high-level perspective, but I do personally think that I find that it is good to have the analysis. From a purely professional standpoint, I do not believe it is unique or distinctive enough as an individual method to need its own separate name for day-to-day use. From a personal perspective, however, I thought the mad cow disease reference was hilarious and applaud whoever came up with the acronym.
I find the benefit in the analysis, and the concerns presented about generated data being present in the data makes sense to me (and if in sufficient quantity, would make sense as biasing the models improperly in a rather significant kind of way).
I particularly enjoyed the humor of this line, the tongue-in-cheek nature is very funny/nice to me here:
"Ascertaining whether an autophagous loop has gone MAD or not (recall Definition 2.1) requires that we measure how far the synthesized data distribution Gt has drifted from the true data distribution Pr over the generations t."
I like their use of color in the paper, I saw a similar orange/green color scheme earlier today and enjoyed it very much as an annotation method.
"A fixed real dataset only slows generative model degradation" is again also a natural consequence of Shannon's noisy channel capacity theorem, one can say that with almost nearly perfect certainty that a limited neural network will not be able to perfectly fit the distribution of the data that it is training on, thus it will have bias, variance, or some combination of both, limited ultimately by the model's capacity itself.
This w.r.t. the original dataset is noise, and we can choose between whether we want collapse, or recursively encoding the noise patterns of the previous model (which might happen to have an additive effect, or maybe not! Who knows! I do not know for sure here, I have not yet figured this one out myself yet).
w.r.t. the real data slowing down degradation, if we are sampling I.I.D. of course then proportionately we still should see some degradation as this is the nature of empirical risk minimization over maximum likelihood estimation. It is still good that they have shown this, however, I thinks.
The fresh data loop, I believe, would be an example of actually a kind of noise in and of itself, w.r.t. the original input dataset, and as long as this 'noise' (from the perspective of the model) has a higher SNR than the (potentially slow) collapse of the model's output distribution, then it should (in some kind of proportion at leasts) be constantly-playing 'keep-up' with the fresh data.
"First, we find that—regardless of the performance of early generations—the performance of later generations converges to a point that depends only on the amounts of real and synthetic data in the training loop. " -- there we are (I saw this after making the SNR point, this makes sense within this framework of interpretation, then.
All in all, I found this paper very aware of itself and what it was studying, it was well-laid out and accessible, and while the points are not necessarily earth-shattering (though I still have to read through some of it, I think), having clear empirical evidence about this phenomenon, detailing it, and cutting away through the forest of (at-least-seemingly) untested battlegrounds is one that I appreciated.
Curious to hear what others think about this one. <3 :'))))
There's renewed excitement about synthetic data because of the Deepmind Geometry solving model that appeared yesterday - which was trained on synthetic data. I guess this is why this is being picked up again.
Perhaps this is vaguely adjacent to a slow motion stream of self reflection, where in the absence of additional (sensory) input, a descent into hallucination is all but inevitable. The same thing happens in a sensory deprivation tank.
Indeed. Such closed loop dynamics turns the AI into a sort of iterated function system, which gravitate to basins of attraction, or exhibit chaotic behaviour.
How about humans in solitary, do they also go MAD, or the "brain in a vat" scenario. It is unreasonable to train a model on its own outputs for many iterations.
> with enough fresh real data, the quality and diversity of the generative models do not degrade over generations
When a model gets deployed and is prompted by people it can get fresh data in the prompt - the prompt itself, RAG material, outputs of tools, human responses to its outputs - and this allows for a form of exploration that can go outside the original scope of the model. If you use the model logs to retrain it won't collapse.
How does that work for image generation AI? And even with LLMs the prompts are going to contain extremely limited information because the users are primarily requesting information, not providing it.
AI is only as good as its models. Now that we're flooding the internet with AI generated content with no way to determine if the content was also produced by AI while endlessly shoveling as much content as possible into these AI models to push up those parameter counts to attract investors, it's inevitable the current generation of popular AI systems will just eat themselves.
LLM interactions can be more than a simple question & answer. Long chats allow the human ample opportunity to steer the model in a direction outside its normal scope.
If they also use RAG or other tools like code execution there is even more ample chance to steer the model outside its original experience.
Then we can use these sessions as training data to re-tune the base model. It's much cheaper than training a new model, can be done on normal computers.
LLMs are not beholden to their training set. They can ingest new information any time, or relate two things that were previously separate to derive new insights.
53 comments
[ 4.0 ms ] story [ 111 ms ] threadIf I give a pig something to eat and he throws up I'm surely not touching that stuff myself
If you interact with people / fresh data regularly you'll avoid that.
I'm not sure how much is down to evolving requirements and data formats vs degradation of the training data. The idea sounds common-sense but it's also triggering a bullshit warning for me, because I'd expect more consistency per AI model, and differences to occur on new model interations - regardless of the dataset.
Or to put it in your analogy, If I feed my cat milk, they'll throw up. That doesn't mean milk is unfit for human consumption.
If AI generated stuff is so bad you cannot train AI with it why would humans use it to train themselves (aka learn stuff)?
a more relevant analogy for llms is a bullshiter that don't know shit about anything but like talk about everything and they're good at that (talking like they know), I'm sure we all at least know or knew in the past someone like that.
at a low enough percentage these bullshiters can thrive and get praise from other people but when there a lot of them and not enough of the real deal, well everyone is in trouble including the bullshiters since they no longer have who to mimic.
This space is where the Venn diagram of marketing and genocides overlap.
At least this is what our specilist at the hospital explained to us.
Obviously ethically we probably should not normalize eating people though.
Just because it doesn't work with AI now, doesn't mean it's something that wouldn't ever work.
Training AI on AI output is less like organised education systems and the less formal written & oral traditions that predate them, and more like dark reddit/chan/other communities eating each other's rhetoric, or some political¹ and religious groups in the wider world. These relatively insular (but sometimes large) groups often descend into a pit they have difficulty reasoning themselves out of.
Maybe the answer for AI is to work out what sort of external mixing helps to keep humans on track, and if it can be emulated in the training models used, so their reasoning continues to grow instead of falling into this pothole.
This might be a crap idea (too little sleep, caffeine has just made me tired and jittery) but perhaps the scale of the information we pile into the process and the fact the retraining at least partly on AI output seems inevitable, means we need to look at training an AI more like training a small population of humans rather than a single unit human.
----
[1] mostly the far right, but that might in part be because they tend to be loud so noticed – far anything is a problem
AlphaZero/MuZero game AIs do very well feeding themselves. Often better than previous AIs with hardcoded rules and training based on human games.
We know that there is more than one AI software but there is not as much variation as for humans.
And as someone else already wrote here, humans have something like a ratcheting mechanism. I would paraphrase it as "Humans have learnt to stand on shoulders of giants". AI does not have this.
Whether it's an AI eating its own training data, a human stuck in an empty cell and forced to live with just his/her own thoughts, or a culture believing its own myths and aggressively punishing insiders and outsiders who challenge them - it's all solitary confinement. And we know how that goes.
(/s of course - although the self-referentiality of memes is balanced by the mad scramble for novelty.)
A pig (or AI) throwing up on an input should increase your priors that it may be bad for humans, but it does not prove the matter either way.
It's not unqualified bad. It is bad if you do it for many times in a loop, and without any external inputs. But you can put fresh material right in the prompt and it won't suffer from retraining on synthetic data.
If you train a small model from scratch on purely synthetic data, like the Microsoft Phi models, it comes out competent and 5x more efficient than models trained on web text. So there's the flip side. You can do it in bad way, and you can do it in a good way.
There is a difference to talking only to yourself and talking with other people.
I find that catchy and descriptive on more than one level.
It is not at all necessarily surprising, I think, from a purely high-level perspective, but I do personally think that I find that it is good to have the analysis. From a purely professional standpoint, I do not believe it is unique or distinctive enough as an individual method to need its own separate name for day-to-day use. From a personal perspective, however, I thought the mad cow disease reference was hilarious and applaud whoever came up with the acronym.
I find the benefit in the analysis, and the concerns presented about generated data being present in the data makes sense to me (and if in sufficient quantity, would make sense as biasing the models improperly in a rather significant kind of way).
I particularly enjoyed the humor of this line, the tongue-in-cheek nature is very funny/nice to me here:
"Ascertaining whether an autophagous loop has gone MAD or not (recall Definition 2.1) requires that we measure how far the synthesized data distribution Gt has drifted from the true data distribution Pr over the generations t."
I like their use of color in the paper, I saw a similar orange/green color scheme earlier today and enjoyed it very much as an annotation method.
"A fixed real dataset only slows generative model degradation" is again also a natural consequence of Shannon's noisy channel capacity theorem, one can say that with almost nearly perfect certainty that a limited neural network will not be able to perfectly fit the distribution of the data that it is training on, thus it will have bias, variance, or some combination of both, limited ultimately by the model's capacity itself.
This w.r.t. the original dataset is noise, and we can choose between whether we want collapse, or recursively encoding the noise patterns of the previous model (which might happen to have an additive effect, or maybe not! Who knows! I do not know for sure here, I have not yet figured this one out myself yet).
w.r.t. the real data slowing down degradation, if we are sampling I.I.D. of course then proportionately we still should see some degradation as this is the nature of empirical risk minimization over maximum likelihood estimation. It is still good that they have shown this, however, I thinks.
The fresh data loop, I believe, would be an example of actually a kind of noise in and of itself, w.r.t. the original input dataset, and as long as this 'noise' (from the perspective of the model) has a higher SNR than the (potentially slow) collapse of the model's output distribution, then it should (in some kind of proportion at leasts) be constantly-playing 'keep-up' with the fresh data.
"First, we find that—regardless of the performance of early generations—the performance of later generations converges to a point that depends only on the amounts of real and synthetic data in the training loop. " -- there we are (I saw this after making the SNR point, this makes sense within this framework of interpretation, then.
All in all, I found this paper very aware of itself and what it was studying, it was well-laid out and accessible, and while the points are not necessarily earth-shattering (though I still have to read through some of it, I think), having clear empirical evidence about this phenomenon, detailing it, and cutting away through the forest of (at-least-seemingly) untested battlegrounds is one that I appreciated.
Curious to hear what others think about this one. <3 :'))))
At least AI hasn't reinvented Mutually Assured Destruction!
> with enough fresh real data, the quality and diversity of the generative models do not degrade over generations
When a model gets deployed and is prompted by people it can get fresh data in the prompt - the prompt itself, RAG material, outputs of tools, human responses to its outputs - and this allows for a form of exploration that can go outside the original scope of the model. If you use the model logs to retrain it won't collapse.
AI is only as good as its models. Now that we're flooding the internet with AI generated content with no way to determine if the content was also produced by AI while endlessly shoveling as much content as possible into these AI models to push up those parameter counts to attract investors, it's inevitable the current generation of popular AI systems will just eat themselves.
If they also use RAG or other tools like code execution there is even more ample chance to steer the model outside its original experience.
Then we can use these sessions as training data to re-tune the base model. It's much cheaper than training a new model, can be done on normal computers.
LLMs are not beholden to their training set. They can ingest new information any time, or relate two things that were previously separate to derive new insights.
Luckily we don't hear too much about Mad Cow Disease these days. Let's hope it doesn't come back.