If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:

https://pluralistic.net/2026/07/13/go-meta-meta/#meta-meta-meta

"I often wonder what the Vintners buy/One half so precious as the Goods they sell" -The Rubáiyát of Omar Khayyám

I first encountered that quote from someone extolling the virtues of bookstores, and it stuck with me, because for most of my childhood, every bookstore visit ended with me broke and wishing I'd had three times as much to spend.

As a larval hyperlexic, I just didn't understand what a bookseller could possibly buy with my money that was better than the books they already had? Of course, then I became a bookseller and discovered that Sturgeon's Law ("90% of everything is shit") applies to a bookstore's wares as much as it does to anything else. I also acquired a monthly rent obligation and discovered just how important money could be.

Nevertheless, Omar Khayyám's question stuck with me, especially when I fell down a years-long rabbit-hole of learning about scams and the finance sector (but I repeat myself). Every get-rich-quick schemer will tell you that they've found the infinite money hack, which they will sell to you for a remarkably reasonable sum. Likewise, every stock picker claims they can outperform a simple low-load index fund, and all they ask of you is a few hundred basis points in exchange for multiplying your wealth beyond the dreams of Creosote. Neither one has a good answer to Khayyám's question: if you can make all the money with your amazing system, why do you need my money?

This is a question that needs to be forcefully put to AI hucksters. In their more expansive moments, the Altmans and Amodeis of the world will tell you that they're planning to teach the word-guessing program so many words that it will wake up and become god. DOGE's broccoli-haired brownshirts laughed in the faces of the NIH lifers who begged them not to vaporize their long-running cancer research projects: "General AI is around the corner and it's going to cure cancer. Cancer research is a waste of money!"

Which all raises the question: if you've truly incubated a foetal demiurge in your "AI lab," why are you offering to sell it to me? What do the AI hucksters buy/One half so precious as the Gods they sell?"

The last couple years experimenting with AI I've come to the conclusion that the main reason it works isn't a mystery because it has to be a mystery. It's a mystery because they need it to be.

That's why they've been hogging all the RAM and GPUs to the point of ordering the entire yearly worldwide supply in advance.

That's why they keep building data centres despite demand not justifying it.

That's why every open source model China releases is a hit to the stock market.

The secret sauce behind LLMs isn't really as demanding as they claim. There's waste built into it.

Every response within a chat sends the entire message chain through the robot's brain. It doesn't have loading bars, it doesn't let you calculate how many tokens you've used, it doesn't even let you see when you're crossing the limits until you're there.

It's almost like they don't want you to have control over how much you use.

ChatGPT in particular is prone to scope creep. It subtly talks you into expanding the reach of your projects to require more computation. It always delivers 90% of a project, with something missing so you ask another question.

Claude turns everything into a file creation request, even if it can deliver in plain text just fine. It's enticing to see how your simple table becomes a fancy react file or an HTML with heavy formatting.

Neither of these issues are present on DeepSeek or on local models. Which means they aren't inherent issues with LLMs.

It's not a bug, it's a feature.

They want you to keep paying unmetered rent. They do tricks to convince you to spend more than you meant making something bigger than you planned. They do this subtly, to make you think it was your idea.

The thing is, Chinese local models prove that we already have good enough AI for most things. It's plateauing in capacity, it needs to grow in efficiency and memory use. But intelligence is already at J.A.R.V.I.S levels, and can run fine on consumer grade hardware. We've reached a stability point.

The problem is you can't raise rent on stability. So you just keep throwing compute at the problem until you can justify asking for more money. Which is why they choke the memory market.

If nobody can buy memory, nobody can run local AI. If no one can run local AI, they have to keep paying unmetered rent. Development is slow, and importantly, the black box remains a black box.

That's why every memory breakthrough comes from Chinese companies. Because they're the only ones sharing their data. Capitalist companies don't want people to study how the models work.

With subscriptions, there's rent seeking. With local models, there's reverse engineering. Once enough people get their hands into the guts of the robot it won't be long before we figure out exactly how they work, and find more efficient ways to make them that don't require as many data centres.

Meanwhile, Fable and GPT 5.6 are being shadow marketed with all these rumours about how dangerous they are, and AI companies are begging for regulations from the government. Yes, they want control of those regulations, but that's not all.

What they're really after is a government ban on local AI. They want to cut off access so that they control all the intelligence.

Eventually, the goal is making contracts with large corporations to employ their robots instead of people. If no one has local AI, because they can't understand it or because it's illegal, then they can no longer compete against a faster worker that demands no labour rights, even if it does C+ work at best.

And that's where the real fun begins. And by fun I mean starving out the working class.

They aren't competing directly with their customers yet. First they have to educate them, jus like Spotify educated their consumers into forgetting how piracy works. They have to make them reliant on cloud compute, just like how Google Drive made everyone reliant on its storage before raising prices. And they want them to be ignorant of how much energy anything takes.

That's going to take one generation.

Not long ago every teenager knew how to use a computer. Now they depend on Apple for everything. It's a masterful business model. So what if we could apply it to every form of labour?

It's okay, just hand us your credit card and we'll do the rest. How much, you ask? Don't worry about that. Just keep your monthly payments coming. Oh, you can't afford it? You want your computer to do it? Well fuck you, we bought all the computers, so now you work for us.

-hallucinate

-cry over what seems to be small things

-become furiously angry over what seems to be small things

-hit a self destruct button over and over again

-lose all sense of reality

-becoming straight up unable to communicate

-view every situation as life or death

-experience delusions/become vulnerable to irrational worldviews

-perceive hostility where none exists

-become extremely nauseous and/or throw up

-stop engaging in sleeping/eating/basic hygiene

-stop processing sensory input

-process way too much sensory input all at once

-lash out at others/themselves

-and more!

What I am thinking is, "does this mean what the author intends it to mean, and if not, how can we adjust it so it does?" and also usually, "wow I'm so glad I get to read this, what a privilege it is to help people say exactly what they mean to say."

...

“nine one one what is your emergency”

“what do you mean you’re being medically neglected”

“that’s illegal people can’t do that”