Big AI mostly switched to synthetic training data anyway. The books they’re digitizing are being used to gather knowledge, not writing styles or logic (mostly).
As in, when you ask ChatGPT how long some book is, it can just go check (if it’s in the database). It’s also useful if you ask about that book or about knowledge contained in that book. It’ll even reference books now (if you demand that in your prompt).
It’s not the same as earlier LLM tech which relied on scanned text to figure out how to respond to any given prompt (from a language standpoint). The “language” part of LLMs is a solved problem now (thanks to the synthetic training). At least for English 🤷
The problem is that synthetic data is not fit for that purpose. The more of it you use, the worse at dealing with the cases LLMs get.
Think of it like this… You feed a language model a bunch of genuine human-written content. Great. Now it can produce the most likely text in a lot of cases. Word combinations that rarely appear in written language rarely get generated, so most of its synthetic data lacks those rare - but still valid - combinations.
Train it on this synthetic data, and now more outliers and rare combinations get filed off. Rinse and repeat.
Big AI mostly switched to synthetic training data anyway. The books they’re digitizing are being used to gather knowledge, not writing styles or logic (mostly).
As in, when you ask ChatGPT how long some book is, it can just go check (if it’s in the database). It’s also useful if you ask about that book or about knowledge contained in that book. It’ll even reference books now (if you demand that in your prompt).
It’s not the same as earlier LLM tech which relied on scanned text to figure out how to respond to any given prompt (from a language standpoint). The “language” part of LLMs is a solved problem now (thanks to the synthetic training). At least for English 🤷
The article quotes a post from ISBNdb saying the issue is model collapse from training on synthetic data.
Total collapse is a solution.
It’s also unavoidable.
Yes, what’s the issue?
That’s like saying, “they had some failure modes from the synthetic data, so they should just obviously stop trying forever.”
They’ll just fix the edge cases and move on. Like any programming task.
The problem is that synthetic data is not fit for that purpose. The more of it you use, the worse at dealing with the cases LLMs get.
Think of it like this… You feed a language model a bunch of genuine human-written content. Great. Now it can produce the most likely text in a lot of cases. Word combinations that rarely appear in written language rarely get generated, so most of its synthetic data lacks those rare - but still valid - combinations.
Train it on this synthetic data, and now more outliers and rare combinations get filed off. Rinse and repeat.