> Will future artificial intelligence systems perform increasingly poorly due to AI-generated material in their training data?
No. Synthetic data is being used to improve LLMs
No. Synthetic data is being used to improve LLMs
That doesn't mean there aren't ways to train a model incorporating synthetic data without seeing model collapse
This line of thought was exacerbated by that one paper that was then parroted (hah!) by every influencer / negativist in the space. It didn't matter that the paper was badly executed, their setup was flawed and that it got rendered moot by the existence of LLama3 models. People still quote that, or the "articles" stemming from it.
A simple example would be chess ai. The core knowledge is rules of the game. We have human generated examples of plays, but we don’t really need them - we can (and we did) synthesize data to train ai.
A similar pattern can be used for all math/physics/programming/reasoning.
No it can't, the pattern for chess worked since it was an invented problem where we have a simple outcome checks, we can't do the same for natural problems where we don't have easily judged outcomes.
So you can do it for arithmetics and similar where you can generate tons of questions and answers, but you can't use this for things that are fuzzier like physics or chemistry or math theorem choices. In the end we don't really know what a good math theorem is like, it has to be useful but how do you judge that? Not just any truthy mathematical statement is seen as a theorem, most statements doesn't lead anywhere.
Once we have a universal automated judge that can judge any kind of human research output then sure your statement is true, then we can train research AI that way. But we don't have that, or science would look very different than it does today. But I'd argue that such a judge needs to be AGI on its own, so its circular.
If you've noticed, most LLM interfaces have a "thumbs up" or "thumbs down" response. The prompt may provide novel data. The text generated is synthetic. You don't need an automated judge, the user is providing sufficient feedback.
Same thing goes for the other disciplines.
You might be interested in some of the details of how AlphaGo (and especially the followup version) works.
Go is a problem where it's very difficult to judge a particular position, but they were still able to write a self-improving AI system that can reach _very_ high quality results starting from nothing, and only using computing power.
There does not appear to me to be any fundamental reason the same sort of techniques can't work for arbitrary problems.
> But I'd argue that such a judge needs to be AGI on its own, so its circular.
But is it circular in a way that means it can't exist, or can it run in circles like AlphaGo and keep improving itself?
I know they're training with synthetic data, I didn't realize that has been done at scalr for long enough to really know if it improved (assuming the metrics its improving are defined well).
LLama3 were post-trained on almost entirely synthetic data. Yes, it works. No, the model doesn't collapse (unless you want it to, of course).
What they did is use Model n-1 to classify, filter and enhance the datasets for Model n.
> almost entirely synthetic data
thing?
edit: found it. The money quote is here, but I really recommend the entire podcast since it's full of great tidbits and insights.
> Thomas [00:33:44]: You mean between supervised fine-tuning like supervised fine-tuning annotation and preference annotation? Yeah. So 100% to RLHF. In fact, that's quite interesting. You start for Llama 2 with a pre-trained model and you have to have an instruction model to chat model. Otherwise, like the model is just like continue finishing sentences. So you need that to start RLHF. So we had to annotate like 10,000 examples. What did we do for Llama 3? You start with a new pre-trained model and then you want, before starting the RLHF, to have now a chat model, which is not too bad. The option one was, let's do human annotation again, like SFT stage. But in fact, by the principle I said before, the annotation would be actually worse than Llama 2. So what we did is that we generated all the data on the prompts with Llama 2 and we applied like basically the last round of Llama 2 we had to kick off and start Llama 3 post-training. So Llama 3 post-training doesn't have any like human written answers there basically, almost. It's just leveraging pure synthetic data from Llama 2.
I have a best fit line. Then I take random data on that line to train a new line.
I pretty much get the same line.
From an intuitive perspective... it doesn't get worse. At worst it stays the same.
Now imagine something a bit more complex. I have a best fit curve that's very close to a line.
I use random data from that curve to train a new best fit line.
I get something different now. Not necessarily worse.
I mean literally just take all your ideas of ML and just imagine it on the 2D plane doing curve fitting. If retraining new lines from generated data doesn't necessarily make things worse.