Comment by simonw
2 years ago
> We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models
The key word there is "indiscriminate". All of the big AI labs have been training on synthetic data for at least a year at this point, but they're doing so deliberately.
I don't think the "model collapse" problem is particularly important these days. The people training models seem to have that well under control.
The question (which I raised in a top-level comment before reading your post) is whether there is any such thing as "discriminate" use of web data. Synthetic data created in the same lab as the LLM is discriminate, but what the authors of the paper are saying (if I read it correctly) is that scraping the web is not currently done in a discriminate way. And it's not at all clear to me that there is a discriminate way to use web scraping, because you can't know for sure what's human-generated and what's LLM-generated.
I get the impression that scraping the web isn't nearly as important a source of LLM training data as it used to be.
Everyone is trimming down their training data based on quality - there are plenty of hints about that in the Llama 3.1 paper and Mistral Large 2 announcement.
OpenAI are licensing data from sources like the Associated Press.
Andrej Karpathy said this: https://twitter.com/karpathy/status/1797313173449764933
> Turns out that LLMs learn a lot better and faster from educational content as well. This is partly because the average Common Crawl article (internet pages) is not of very high value and distracts the training, packing in too much irrelevant information. The average webpage on the internet is so random and terrible it's not even clear how prior LLMs learn anything at all.
> The average webpage on the internet is so random and terrible it's not even clear how prior LLMs learn anything at all.
Perhaps we should stop exposing humans to them, as well?
3 replies →
You trim, yes, but AI content surely invades (all?) areas of written material. People are increasingly using AI to assist their writing. Even it if's for slight editing, word choice suggestions.
Even AP doesn't ban the use of LLMs, its standards prohibit direct publishing of AI-generated content. I'm sure its writers leverage LLMs in some ways in their workflow, though. They would probably continue to use these even if AP attempted to ban LLMs (human incentives).
4 replies →
I think this is roughly correct. My 2c is that folks used the initial web data to cold start and bootstrap the first few models, but so much of the performance increase we have seen at smaller sizes is a shift towards more conscientious data creation/purchase/curation/preparation and more refined evaluation datasets. I think the idea of scraping random text except maybe for the initial language understanding pre-training phase will be diminished over time.
This is understood in the academic literature as well, as people months/years ago were writing papers that a smaller amount of high quality data, is worth more than a large amount of low quality data (which tracks with what you can pick up from an ML 101 education/training).
This is a similar problem to what was observed in Diffusion models going "MAD" when trained on synthetic data. https://arxiv.org/abs/2307.01850 . Therefore, going forward AI companies will find it increasingly difficult to get their data by scraping the. web, because web will be full of synthetically generated data.
They needed to deal with degenerate data on the Web anyway. It's always been full of trash and spam.
I agree with you when it comes to training, but at the same time, I think that's also the power we get with the web. You can have a voice, even if others don't agree with you. I don't think that should be taken away unless you are inciting violence.
At least some of the LLM generated content will be vetted/selected for by a human being though.
Read the paper, the problem is each generation forgets information. Starting at the tails of the distribution they learn. No amount of filtering/selecting would help here. People would need to fill in missing information without AI help. If they are just filtering, it does nothing to stop model collapse.
1 reply →
Or to consider the inverse of indiscriminate, selection.
Mutation = bad.
Mutation + selection = good.
(given enough iterations)
wow this is such a good point! Evolution is just that!
How do you "discriminate" data gathering at web-scale, though? In my view, everything at web-scale only works because there are no humans in the loop, as repeatedly explained here in basically every thread involving Google or Facebook. Yes, since it's a scientific paper they should have defined their usage of the word, but I see nothing wrong with the basic premise that automation at large-scale implies indiscrimate use of content.
You can use LLMs to vet the relevancy of the content, so you only select the most useful data. I believe most labs are doing this today.
> I don't think the "model collapse" problem is particularly important these days. The people training models seem to have that well under control.
And you base this on what? Vibes?
Basically yes. Vibes based on reading between the lines of various papers, blog announcements and tweets from people better informed than I am.
" The people training models seem to have that well under control." The people training the models are not the C-Suite, and that is an element of entropy there is currently zero accounting for.
The paper is interesting, but it seems to focus on iteratively training models on synthetic copies of the same data. Obviously, this is going to cause problems.
They did not address what happens if the model is trained on synthetic data that is distinct from the source corpus.
They make it clear in the paper that their primary "real-world" concern is that it's difficult to distinguish synthetic data from real human interaction when scraping data from the web. This will only get worse over time with our current way of doing things.
How are they supposed to deliberately train on synthetic data when they don't know whether it is (synthetic) or not?
Also, do you not feel that it is presumptuous to dismiss a body of work in a few sentences with a "seems fine to me"?
In this case I wasn't reacting to this specific paper so much as to the widespread idea (at least that I've observed among AI skeptics) that "model collapse" is a huge problem.
> All of the big AI labs have been training on synthetic data for at least a year at this point
Curious how you know this and the actual extent of such training.
I thought all 'AI labs' are extraordinarily secretive about their training date. Do you have any inside connections to ' All of the big AI labs' ?
I find nothing wrong with your statement. I am curious about the paper's use of "indiscriminate." I read this as "just feed the AI more AI output without care" which one can indeed do deliberately.
Seems to me that deliberate discriminate use should yield better against expectations.
Came here to say the same. "indiscriminate" doesn't really make sense. It's very deliberate.
However, there is one scenario: Scraping of web data. In that case, AI labs might know what is model generated.
> I don't think the "model collapse" problem is particularly important these days.
I think you might misunderstand what model collapse is. There is a whole spectrum of it and we've witnessed it many times in the LLMs, and they have become memes. A fairly recent example is the Golden Gate Claude[0]. This is mode{,l} collapse. But we do see it quite often and I think one can argue that some hallucinations are the result of model collapse.
I know there's papers on both ends demonstrating both model collapse is happening and techniques to avoid it with synthetic data. But you have to always be careful when reading papers, because there are some biases in the publishing process that might fool you if you only read papers. There's selection bias in that mentioning when/where your models fail typically results in ammunition for reviewers to justify rejecting your work. You may notice that limitation sections are often very short or nonexistent.[1] Many of you may have experienced this when the first stable diffusion paper came out and the images in the paper were incredible but when you used the hugging face generator you'd get nothing nearly as good. Hell, try even now[2]. Can you do better than I did? Sure! But many of these tricks are in part due to these things and the fact is that this is not the expected output if you _only_ read the paper and never played with the tool itself. That there's a big difference between these.
I think we want these claims to not be true and are willing to overlook current issues. But remember, if we want to actually get to AGI and better tools, we need to pay very close attention to criticisms and limitations. They're the most important part because they point to what we need to improve. Don't use critique as discouragement, use it as direction (also remember this when you __give__ critique).
[0] https://arxiv.org/abs/2112.10752
I don't think Golden Gate Claude was related to model collapse. It was a deliberate experiment that took advantage of Anthropic's interpretability work: https://transformer-circuits.pub/2024/scaling-monosemanticit...
My previous [0] points to [0] I just used the HN link instead.
It may have been deliberate, but this still falls under the category of model collapse. Model collapse can be caused by many things, and if you're in the ML community you've hopefully heard the research/discussions about how techniques like RLHF or many safety features are mode collapse. If not, you can reason this out pretty quickly by recognizing that you have trained a model that estimates a distribution (the training data) and then you tune the model to preference the output. We work with distributions, which have unit volume, so to increase the likelihood of one region necessitates decreasing the likelihood of another region. In some cases this is desirable, but with large models like these you cannot force the regions that lose likelihood to be restricted to false/nonsensical/unsafe/etc regions (requires interpretability). You just do your best and hope. Added complexity depending on the distributions you use or estimate.
If you don't believe me, read the comments to the thread. That's actually why I linked the HN comments. Here, I'll save you the work of clicking the link
[0] https://www.anthropic.com/news/golden-gate-claude
2 replies →
How do you envision thwsw companies aee discriminaying ans how many manhours goes into disxriminating an hour?
I fine your optimism here delusional at best.