Hacker News Reader

Comment by jerpint | Hacker News Reader

ChatGPT is fine tuned on a curated, relatively small dataset, which is of very high value and would mitigate some of the attacks

View on HN

Have you come across any specific details on their input datasets? I’ve been assuming CommonCrawl and Pile but if you have anything specific it would be very interesting what else was used.

All the major LLM vendors are vague on datasets because it's quite certainly copyrighted text, (Remember when Google was scanning all the world's books a decade ago?) and they don't want to give the inevitable lawsuits a jump start.

You can just ask ChatGPT what data it was trained on.

> The news groups' concerns arose when the computational journalist Francesco Marconi posted a tweet last week saying their work was being used to train ChatGPT. Marconi said he asked the chatbot for a list of news sources it was trained on and received a response naming 20 outlets including the WSJ, New York Times, Bloomberg, Associated Press, Reuters, CNN and TechCrunch.

https://www.bloomberg.com/news/articles/2023-02-17/openai-is...

Wouldn't ChatGPT name a source even if it did not use that source or if the claim could not be sourced?

Can ChatGPT be trusted on these kind of matters, though? If doesn't really know how not to know.

Would ChatGPT snitch on ChatGPT

>Remember when Google was scanning all the world's books a decade ago?

Yes, and I remember that they won that lawsuit.

https://www.google.com/amp/s/blog.google/products/search/ove...

This works fine for now. I am however worried that politics will make this much harder in the future. We're already seeing that a certain crowd wishes that AI responses reflect the "average" without any kinds of filtering. If they were to pass a law in this direction (and, let's face it, considering all the other overstepping laws passed in the direction of book banning etc. this is sure to come) we might not have any choice or options left.

Reply on news.ycombinator.com