> but about the originality of the content itself
Your metric is too ill-defined. Here, have some highly unique content
gZbDrttzP6mQC5PoKXY2JNd9VIIxBUsV
ClRF73KITgz5DVnSO0YUxMB6o7P9gh8I
1ttcQiNdQuIs4axdAJvjaFXXkxq0EvGq
Pd0qwVWgSvaPw8volLA0SWltnqcCNJiy
If we need unique valid human language outputs I'll still disagree. Most human output is garbage. Good luck on your two tasks: 1) searching for high quality content 2) de-duplicating. Both are still open problems and we're pretty bad at both. De-duping images is still a tough task, before we even begin to address the problem of semantic de-duplication.
The idea is to let humans be humans, make a mess, debate, have their opinions, and AI comes after that and removes the herp derp from the useful parts.
As a test of concept copy paste this whole page, put it in a LLM and ask for an article. It will come out without junk, but will reflect a greater diversity of opinion, more arguments, will do debunking, and generally have better grounding in our positions than the original content.
So it's careful synthesis over human chats that is the end value. Humans provide that novelty and lived experience LLMs lack, LLMs provide consistent formatting and synthesis. The companies that understand that users are the source of entropy and novelty will stop trying to own the model and start trying to host the question.
Wondering why reddit doesn't generate thousands of articles per day from comment pages. It would crush traditional media in both diversity and quality. It would follow the interesting topics naturally.
This is why LLMs are so helpful in writing research proposals. If I’m writing a proposal, I basically need to vibe on a concept without worrying about it sounding entirely professional / buttoned-up and concise - otherwise I get bogged down in word smithing. So I tell it about the concept I’m trying to get across, and it converts that into a more concise and punchy paragraph, while also providing critical technical feedback (if I ask) and making sure I’m using terms of art correctly. As you put so eloquently, it removes the herp derp.
And then, kind of like how LLMs are good for boilerplate code but get less helpful as you get to more specific problems, this approach works for the first few paragraphs but deteriorates the deeper in you get.
The idea is to use the LLM as a word smith, it doesn't conceptualize novel ideas from itself, that is your job. I would normally chat the topic until it gets it, and after that I ask it to write it up as an article.
Here is a feed I collected over time: https://old.reddit.com/r/VisargaPersonal/ I'm just using this subreddit as a blog.