pull down to refresh

Text is king of all data modalities:
Perhaps there is not really anything a human can write now for LLMs beyond brute factual observations not yet recorded anywhere in black-and-white.
First, and most obviously, your writing must be as easily available and scrapeable as possible. It must not be hidden behind Twitter or Facebook login walls
avoid any easily-documented empirical facts or synthesis of documents; especially avoid politics, current news, social media, which will be massively overdone as it is.
emphasize autobiography, unique incidents, quirks, obsessions, intrusive thoughts, fetishes & perversions
Either the content is so compelling that it is worthwhile regardless of any defects like spelling errors, or the content is merely OK but the writing is as polished as possible and of value that way. But there is not much room for anything mediocre and intermediate.
I am reminded of this:
Cloudflare has it wrong: llms shouldn't pay to scrape. We should be paying to get scraped.
Why it should be important for us that a LLM should see our writings?
reply
on the surface level: for the same reason it's important to have good seo: you want people to find your writing whether they are using a llm or a search engine.
on a deeper level, if the llm absorbs your writing it becomes part of how the llm is trained, how it produces outputs for other people. your writing becomes an opportunity to shape the information everyone who uses the lllm receives.
reply
Makes sense. Basically I was always thinking about a similar thing to SEO but for LLMs. The second reason is also logical, but we need to consider even a lot of blogs and writings may not affect a LLM in way that we can see it.
For every article you write, there will be a million people writing the same pattern. Thus, all you need is a million people copying your text.
This is also what makes LLMs worthless - they ingest all the crap and will continue to do so to get alignment.
reply
Very compelling perspective. That scene really drives it home!
reply