When it comes to something like a LinkedIn post however who cares. Everyone uses AI cause the content itself is mercenary anyway right. Thought leadership.

But I have been using Sonnet to write some Youtube narrations and I think I've discovered a kind of value that is not 'AI slop'. I already posted one of them -- about the giant pink hologram in Blade Runner 2049 -- and it's been getting thousands of views and a dozens of likes when clipped into shorts. [https://www.youtube.com/watch?v=sK-TsLQkIos]

The text is like: "The neon is pink. Her hair is blue. Her face is light. She smiles because she was made to smile, and somewhere beneath the awareness of that fact is the older, simpler, more animal truth: a face is smiling at you. Someone said your name. Someone said “you look lonely” and meant it as an invitation rather than a wound."

To me that is a kind of LLM writing that's "value above replacement". Like if I lost that Sonnet 4.6 script I wouldn't be able to get it again it was just a particular permutation of the neural network manifold that emerged from our discussion about the movie and the style I wanted etc.

So, yes, humans can also expand text in a mechanical way. I just don’t want to read expanded text. So really the question is “is this text expanded?”. Pre-LLM very little was. Post-LLM much is and has certain distinct flavours that are detectable.

But I’m happy to say that I also don’t like human expanded text and the content in OP certainly reads expanded. It’s like inlining functions everywhere and then writing unrelated comments around them. Not interesting.

After all, if I just want to read text there are reams of content marketing blog posts and SEO keyword concatenations that I could go read. Or to put it in a different way: SaaS dropshipping shemale singles near you this one mother found a puppy buy crypto toys cheap car loans zero down furniture sale foreclosure mesothelioma lawyers car Trump USAID ethereum get rich quick

LLMs are fundamentally different from other writing tools because they attempt to construct meaning in a way that is not done by the author. A main purpose of writing is to convey meaning, so LLM writing defeats that purpose.

This isn’t the common case. How many of the thoughts produced while working on the words have much novelty or high quality may be an open question, but it’s rare to type or handwrite 3000 words and not have any thoughts.

> Another can dictate an idea, argue with a language model for an hour, reject six drafts, rearrange the whole piece, rewrite half of it, and publish something they understand down to the final comma

Also probably not the common case. And so far my experience is that models make decent to good editors/critics, but average originators (though with remarkable breadth and sometimes outlier originations).

When I get materials that are clearly AI generated, I look at them as if there was less effort put into it.

But that is irrational. If it was verified for accuracy, and edited for conciseness, and addresses the task at hand, does it matter?

There is SO MUCH STUFF that you could read. You could read all day every day and not keep up with the human-written output in just a single sub-niche of writing within a niche sector. That being the case, you must use some heuristics to determine what should be read. Whether the entity publishing a piece felt it was worth human time to write that piece seems like a fine heuristic.

The privilege argument is so bad! We aren't doing this any more! The tyranny of edge cases is over. It ended some time in 2023. I am not going to feel obligated to read every piece of slop out there because 0.01% of authors have lost the use of their hands. Stephen Hawking somehow managed to write without AI.

The craft of writing is part of the interesting part. The words you choose are important,same as brush strokes in a painting. Offloading the act of writing to llms takes away that whole part.

Although there are some niche uses where LLMs assist or enhance the creative process, the vast majority of machine writing exists specifically because it takes zero effort and allows you to spam human cognition at an unprecedented scale. There is no redeeming quality to 99%+ of AI-generated LinkedIn posts, AI-generated books on Amazon, and so on.

The style-based heuristics we previously had at our disposal to filter zero-effort content no longer work here, so detecting LLM text is the next fallback.

Alright, compare the following degrees of autonomy:

- I wrote it, made sure everything's correct, let the LLM proofread it/translate it to English, and checked it afterwards

- I had brief scattered notes with no coherent idea, then threw it at an LLM to make it an article in whatever way it feels best because I don't care

- I had a three word prompt "Write about X", then a deep research agent spent ungodly amount of tokens and search and scraping requests, and put up an article (at least some grounding and it can depend on the harness quality)

- I had a three word prompt "Write about X", then a model did it in a single call with no real-life grounding at all

You're looking at my clearly generated article. Can you tell what is mine, what is collected from the web by the agent, and what is sourced from the model's own knowledge? Sourcing is exactly what the generated text obfuscates.

1. https://news.ycombinator.com/item?id=49042354

But when they do that, you can tell.

I get this is tongue in cheek, but it's so wrong that it needs pointing out. Even if that is a thought in their head, they clearly have to have others in order to produce the text.

But in general, checking that a human wrote it is economic as much as anything else: attention is currency limited by time and energy, AI effort is abundant and therefore low value, human effort is less abundant and therefore more valued. Pyrite can look pretty but it's not what people want.