So this is how you foil a forest fire!
No doubt about tit, they're not just good, they're the breast in Texas.
No. LLMs and AI image models are fundamentally, a different technology. They use completely different mathematical models, data models, and target different platform spaces.
As a simple example, to train an LLM you need to feed it lots of text and not much else. To train an image model, you need to feed it billions of tagged images. That is, images that have every stupid little thing in the image "tagged" with a bounding box and text describing what's inside said bounding box.
Also, these terms might not mean anything to you, but LLMs use transformer decoders (autoregressive causal retention) while image models use diffusion networks. The way they work is different at a fundamental level. Text is 1-dimensional and an LLM acts as a word prediction machine whereas image models start with chaos (random garbage pixels) and refine it over and over again in n steps until the chaos starts looking like the prompt.
The way prompting works is also different.
Someone cut above the rest‽
Grimy already pointed out that this article is non-scientific nonsense but there's something else it's missing: The efficiency of AI over time.
Qwen-3.8-27B is due out next week and it's supposed to be on par with Claude (Sonnet, latest). That runs on a regular PC GPU.
Big AI isn't the only game in town. They're just what makes the non-tech news.
The open weights stuff gets drastic improvements every three months or so and there's breakthroughs in efficiency every six months (ish). We're only at the baby steps of AI tech and the advancement is behaving like a chaotic Moore's Law that leapfrogs itself regularly, then stagnates for a bit, then there's another big shift.
If this keeps up, the efficiency of AI will become something that makes using huge data centers for anything but training a waste.
Yes, but on the other hand, they get the mail before anyone else and those guys deliver!
You are correct! Absolutely 100% accurate!
That is not how LLMs learn anatomy!
What does that have to do with AI image generation models though?
It's the code generation and automation uses that are growing. I don't know about casual use or using AI for office tasks.
AI providers that offer API access don't have enough hardware to support the demand. They keep increasing prices (the roundabout way by slowing down requests, reducing the amount of included tokens in tiers, and replacing models with new ones that use more tokens) and that's still not enough to reduce demand.
Using AI to code and automate stuff is probably its most effective use. But most people don't see it. Is all business/enterprise stuff happening behind the scenes.
AI image generation is such a weird thing. When you self-host it (run it on your own GPU), you get vastly superior capabilities but at the same time you get all the awful results too.
Example: When you use ChatGPT or Gemini or similar "Big AI" service to generate an image, you get one image. However, in the background, those services were actually generating a whole bunch of images and at the end, using AI to pick the best one.
When you DIY, you tell it to generate say, 50 images, hoping you get lucky (haha) and that's when you see the weird shit. Sure, 10 out of those 50 will be OK-ish—something to work with (using additional tools to say, fix those wonky hands), but in "the bad 40", there will be a random naked person.
Like... Why‽ I mean, I actually know why (AI image models only work well if they're trained on nude people! Why? Because that's how they learn proper human anatomy/ratios! Just like a doctor, actually), but it's mildly disturbing when I'm trying to visualize a scene in a novel where a girl is holding a doll and it mixes up the doll and the girl and somehow the girl ends up naked (looking like a doll). Ugh! W T F! I didn't ask for that!
Can't delete shit like that fast enough!
Note that "Big AI" has that problem on occasion too! It's basically impossible to fix with 100% accuracy, which is why AI companies are fighting laws like this tooth and nail.
Also, the detection algorithms/AI don't work very well! They know this. I know this. We all know this. That's why these public Big AI models are regularly generating nudity despite having system prompts that are supposed to stop it.
So what can we do? Many people would say, "just ban AI image generators" or "make it so they can only generate cartoon/non-realistic images". That's called "banning a core technology" and it never works if said tech doesn't have some sort of rare supply chain ingredient (e.g. uranium).
I don't know. On the one hand, Big AI and smaller AI companies are definitely offering a depraved service. On the other hand, it's all fake!
All these "nudify" tools and apps are generating 100% totally made up bullshit. Why do we need to ban fake, digital things?
Think about it from a tech standpoint: Let's say we ban using AI to "generate nudes of people without their permission." All that users of these services will do is generate generic nudes and then use a different model to place the person's face on the body. There's nothing at all you can do to stop that because the user could just crop the image so that only the face and neck are showing and then just paste the result back into the original image.
This is a people problem. That is: People are using tools in ways we don't want them used.
You can't solve people problems with technology. It's one of those "laws" as old as tech.
IMHO, the best we can probably do is to hold people accountable when they behave inappropriately, but the anonymous nature of the Internet makes that extremely hard to do. So... 🤷
[removed by mod]
Nah. If there really is a crash and a sudden glut of AI GPUs and related hardware, it'll drop the prices of everything PC-related real fast.
The hardware will absolutely, 100% continue to be used for AI stuff. Nobody is going to buy that stuff in bulk just to start desoldering chips. That would be stupid, because...
The demand for AI is just too damned high! If anyone can bid on OpenAI's bankruptcy hardware sale, either one BIG player is going to take it all (read: Microsoft... Which was the plan all along I swear!) or it'll be sold in batches and loads of small AI companies (including startups) will get it cheap.
Small companies are already reselling tokens on their tiny little hardware farms using open models like GLM and Qwen. Anyone with the hardware can do so, and there's loads of money to be made doing it. So why would businesses bother trying to use that memory for anything else? They wouldn't.
However, a huge sale will create a glut regardless, which will almost immediately drop prices by some huge amount. The demand will slowly go back up—because AI demand is still growing—but we'll all get relief for probably six months to a year.
Listen to me: If that happens, that is when you want to buy a new GPU/PC! Wait like 2-3 months after OpenAI's bankruptcy announcement (save that money), then spend it on the best hardware you can get. Because the demand for AI isn't going anywhere. It's just going to drop back down into reality: Where they only build data centers where it makes sense. As in, where there's actually plenty of power available (or they can just install a zillion solar panels).
I did the math and these are the approximate real numbers:
- US: ~30 million
- World: ~105 million
That's active Linux desktops. That is, they're being used to browse the Internet regularly enough to make it into StatCounter.