“Revise generically: [manifesto]” (certainly not the best prompt)

Folks seem to like Ollama per HackerNews threads: in a coding context here:

Not using Codestral (yet) but check out Continue.dev[1] with Ollama[2] running llama3:latest and starcoder2:3b. It gives you a locally running chat and edit via llama3 and autocomplete via starcoder2.

It's not perfect but it's getting better and better.

[1] https://www.continue.dev/ [2] https://ollama.com/

Please no unabombing though

Oh wow he wrote a 35k word manifesto… feel like that’s so rare you’d still stand a solid chance at being identified somehow.

You could, but even then you need to put some thought on how to prompt and review/edit the output.

I've noticed from usage that LLMs are extremely prone to repeat verbatim words and expressions from the prompt. So if you ask something like "explain why civilisation is bad from the point of view of a cool-headed logician", you're likely outing yourself already.

A lot of the times the output will have "good enough" synonyms. That you could replace with more accurate words... and then you're outing yourself already. Or simply how you fix it so it sounds like a person instead of a chatbot, we all have writing quirks and you might end leaking them into the review.

And more importantly you need to aware that it is an issue, and that you can be tracked based on how and what you write.

I did simplify the whole thing, as you noticed; but note that his SIL and brother identifying him is another example of the same process, David knew that expressions that Ted used like "cool-headed logicians" were highly unusual, not too unlike what the socio- and forensic linguists did there.

But if the author does not have a large body of writing that is known to come from that individual

Such as a Lemmy or Facebook account? Or any other online account associated with your writing, really; we produce far more text in the internet than ourselves realise.

And while a priori, your different accounts through different websites might look completely disconnected, as you connect two of them as coming from the same person, connecting a third one is easier. And a fourth. So goes on.

A small caveat is that while the corpus is bigger, so is the noise introduced by people from the other side of the world that happen to use the same patterns as the person whom you want to identify. Even then, I believe that the ability to bulk process text to find authorship grew considerably faster than the number of potential matches.

Let's see how well this quote ages as subscription services become ubiquitous.

You can’t eat your cake and have it too.

And I can't "magically" know what you're referring to, either - given that you're replying to a rather long comment but providing exactly zero context on what specifically you're replying to.

Quotes, use them.

I don't rule out the possibility of future intelligent text generators, but I don't think that large language models will do it.

I don't know enough about the evolution of cognition in human beings to believe that your second statement is wrong or right.