Oct 14, 2026, 3:33 PMPeople who find LLMs useful when "used correctly" don’t understand that they manage to use them correctly because they spent 20 years of there life learning their craft without LLMs.
LLMs is the only tool for which the less you use it, the better you become at using it.
@ploum In Cory Doctorow's words: be a "centaur", not a "reverse-centaur"
> A "centaur" is a human being who is assisted by a machine (a human head on a strong and tireless body). A reverse centaur is a machine that uses a human being as its assistant (a frail and vulnerable person being puppeteered by an uncaring, relentless machine).
https://pluralistic.net/2025/09/11/vulgar-thatcherism/#there-is-an-alternative
Oct 14, 2026, 8:00 PM@Tock @ploum I'm interested in your exact issue with LLMs in this case, because I consider myself a "critic" as well, but agree with Cory Doctorow on his LLM usage as well.
My take is that there is a difference between the huge private and token-gated behind GPU farms AI economy, and the LLM technology itself. The example he uses in his book is the fact you can use whisper on your local laptop to transcribe a meeting/podcast. That has insane value and is a good use of the technology that now belongs to the commons due to open weights.
Or maybe you had something else in mind?
Oct 14, 2026, 8:12 PMOct 14, 2026, 8:11 PM
Oct 14, 2026, 10:02 PM@Tock @bobunny @ploum cory doctorow is a sci-fi hack writer that people seem to have confused with some kind of revolutionary visionary. he is a kept pet of the tech class and his job is to sell you his books and speeches. the dude is manifestly part of the problem, not the solution. the LLM as a technology is an edifice of white-supremacist culture, to wit: https://www.whitesupremacyculture.info/worship-of-written-word.html
@atax1a @Tock @ploum I'm not sure how what you linked relates to the conversation, is it in response to text-based LLMs as a technology that has largely been fed written text for most models?
If it's the first that doesn't change the technology: the math remains the same, and the training/inference process doesn't change and is still useful. Training data and biases is a real issue though everyone agrees. That doesn't disqualify the technology and some usages (being able to "grep" a podcast on my laptop). It does for others.
Or was it that I cited a post? If you think that I parroted some ideas that's a bit reductive of my own personal views and thoughts
Oct 14, 2026, 10:30 PM@bobunny we said what we said (the technology is inherently white-supremacist) and if you have trouble understanding that, it isn't our problem.
Oct 15, 2026, 9:08 AMOct 15, 2026, 1:33 PMOct 15, 2026, 3:03 PM@ploum he is a "celebrity activist" whose purpose is to absorb and defang any real energy. he is a con. his use of slop machines is an example of what we mean. hth
@ploum
La trad. de la conclusion est imbitable..!
Celle de Deepl:
"Les LLM sont les seuls outils pour lesquels moins on les utilise, mieux on apprend à s'en servir."
(on dirait presque du P. Dac !)
@ploum one of the many paradoxes of this technology. another one might be how being more sceptical of the tech makes you more likely to actually do quality control on LLM output, lessening it's negative inpacts. except that's the least likely people to use it.
and even with that, they have this tendency to slowly fry your brain, which is a huge risk to corrode away any notion of "when used correctly" over time, since it's probably fine, it's been fine 98% of the time, right? but of course that's not a thing we (and especially not investors) worry about.
@ploum We are at the beginning of a great unlearning.
@ploum and everyone thinks they use it correctly but others use it wrong
@ploum Which is also kinda the reason all those “learn to do X with AI” courses are basically a scam.
@ploum You think you’re using it ?
It’s using you.
It’s learning, and you’re not.
Oct 14, 2026, 6:51 PM@robtrew : awesome sentence !
Oh - I think most of them understand, because *nobody* uses them correctly at first.
Prompting a LLM to maximize sane output, get what you actually are looking for, and recognize when things have inevitably gone off the rails and how to correct that is a skill, somewhat similar to hostage negotiator.
And you need that skill on top of actually understanding the subject matter.
The two are different skills. You need both. Either one missing and it's going to go poorly.
I can, for the very narrow field I am an expert in, get viable results from the LLM, saving me significant amounts of time with output honestly better than I would make alone because I am human and get bored. It took me 20+ years of systems engineering and programming, and a solid year of using the AI to get there, and it's far from perfect yet.
The correct way for someone without specific domain knowledge to use an AI for that domain is "not to". I am not a doctor - so I don't ask the LLM about medical matters. I am not a lawyer - so I don't ask it about law.
And if you are learning - it's probably counterproductive.
@ploum No, I quite understand this.
@ploum
LLMs are a new tool. That's all.
If you're already a skilled woodworker, a new woodworking tool may make your job a lot easier. You still have to learn the new tool, find out the hard points and the brittle ones. But then you can show to others how much easier your work is now.
That does not mean this tool will make anybody into a skilled woodworker. It does not mean someone who watched how easy your job is can now suddenly do that job. If it's the only tool they've ever used, then there's no "woodworking job" there to be "made easier".
You do, absolutely, become better at using LLMs by using them more - Just like any other tool. If nothing else because they're still shifting around a lot and whatever worked yesterday may not work tomorrow any more.
Oct 15, 2026, 1:50 AM
@ploum Why do you believe we don't understand that it's the expertise that matters?
My worry is where the next generation of expertise will come from.