A simulation of me: fine-tuning an LLM on 240k text messages

Post by krainboltgreene » Tue, Jan 09, 2024, 5:08 PM UTC

valine wrote: post #140

krainboltgreene wrote: post #139

Earlier quoted context omitted.

> The idea that reading a piece of text constitutes copyright infringement is ridiculous. No man, it's not ridiculous. If I write a program that copies someone's book and try to sell it I'm infringing on that copyright. I cannot sell a zipped version of the Harry Potter books. I feel like there's so many people weighing in on this discussion who haven't actually done any real world copyright related stuff.

I see the source of your confusion. LLMs are not actually zips of the training dataset.

Post by krainboltgreene » Wed, Jan 10, 2024, 5:16 PM UTC

valine wrote: post #142

krainboltgreene wrote: post #141

Earlier quoted context omitted.

It is an incredibly common refrain amongst experts in the field that it is a compression of the dataset. But that doesn't matter, because you clearly didn't understand what I was writing which doesn't shock me considering your position on LLMs.

> It is an incredibly common refrain amongst experts in the field that it is a compression of the dataset This was a common idea three years ago. No one in the field seriously believes this today.

Post by valine » Mon, Jan 22, 2024, 3:55 AM UTC

krainboltgreene wrote: post #143

valine wrote: post #142

Earlier quoted context omitted.

> It is an incredibly common refrain amongst experts in the field that it is a compression of the dataset This was a common idea three years ago. No one in the field seriously believes this today.

You're going to have to let a lot of scientists know that, because they're still publishing papers with that understanding. I guess they should have consulted you first.