this post was submitted on 01 Nov 2025
145 points (93.4% liked)
Technology
76558 readers
2477 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I'd generally agree, the industry itself seems to be high risk for exploitation because of the very nature of it.
Realistically though this isn't going to help that much, most AI generated content that tries to look real ends up quite uncanny. So this is probably more likely to cannibalise some of the least problematic parts of the industry like digital art.
You haven't seen the newest Flux/Qwen/Wan workflows, much less all the new video models coming out.
Perfect, no, but porn is one of those industries where 'good enough' is good enough for most, when combined with convenience. See: potato quality porn sites that still get tons on visitors
It's just all locked behind 'enthusiast' tooling and hardware now, but that won't last long.
Not even that. You just need a relatively recent GPU if you want to run locally. There are plenty of sites that will give you free credits or charge to run your prompts, no local computation needed.
TBH most 'average' people don't have GPUs or even know what that is, but they do have smartphones.
And they can already run these models pretty well too, albeit suboptimally. That's my point. But the software needs to catch up.
Honestly, there is a lot of potential room for substantial improvements.
Gaining the ability to identify edges of the model that are not-particularly-relevant relevant to the current problem and unloading them. That could bring down memory requirements a lot.
I don't think
though I haven't been following the area
that current models are optimized for being clustered. Hell, the software running them isn't either. There's some guy, Jeff Geerling, who was working on clustering Framework Desktops a couple months back, because they're a relatively-inexpensive way to get a ton of VRAM attached to parallel processing capability. You can have multiple instances of the software active on the hardware, and you can offload different layers to different APUs, but currently, it's basically running sequentially
no more than one APU is doing compute presently. I'm pretty sure that that's something that can be eliminated (if it hasn't already been). Then the problem
which he also discusses
is that you need to move a fair bit of data from APU to APU, so you want high-speed interconnects. Okay, so that's true, if what you want is to just run very models designed for very expensive, beefy hardware on a lot of clustered, inexpensive hardware...but you could also train models to optimize for this, like use a network of neural nets that have extremely-sparse interconnections between them, and denser connections internal to them. Each APU only runs one neural net.
I am sure that we are nowhere near being optimal just for the tasks that we're currently doing, even using the existing models.
It's probably possible to tie non-neural-net code in to produce very large increases in capability. To make up a simple example, LLMs are, as people have pointed out, not very good at giving answers to arithmetic questions. But...it should be perfectly viable to add a "math unit" that some of the nodes on the neural net interfaces with and train it to make use of that math unit. And suddenly, because you've just effectively built a CPU into the thing's brain, it becomes far better than any human at arithmetic...and potentially at things that makes use of that capability. There are lots of things that we have very good software for today. A human can use software for some of those things, through their fingers and eyes
not a very high rate of data interchange, but we can do it. There are people like Musk's Neuralink crowd that are trying to build computer-brain interfaces. But we can just build that software directly into the brain of a neural net, have the thing interface with it at the full bandwidth that the brain can operate at. If you build software to do image or audio processing in to help extract information that is likely "more useful" but expensive for a neural net to compute, they might get a whole lot more efficient.
Whooa nellie.
I don't care to get into that, really all I meant was 'they need some low level work'
Popular models and tools/augmentations needed to be quantized better and ported from CUDA to MLX/CoreML... that's it.
That's all, really.
They'd run many times faster and fit in RAM then, as opposed to the 'hacked in' PyTorch frameworks meant for research they run on now. And all Apple needs to do is sick a few engineers on it.
I dunno about Android. That situation is much more complicated, and I'm not sure what the 'best' Vulkan runtime to port to is these days.
I think that a lot of people who say this have looked at a combination of material produced by early models and operated by humans who haven't spent time adapting to any limitations that can't be addressed on the software side. And, yeah, they had limitations ("generative AI can't do fingers!") but those have rapidly been getting ironed out.
I remember posting one of the first images I generated with Flux to a community here, a jaguar lying next to a white cat. This was me just playing around. I wouldn't have been able to tell you that it wasn't a photograph. And that was some time back, and I'm not a full-time user, professionally-aimed at trying to make use of the stuff.
kagis
Yeah, here we are.
https://sh.itjust.works/post/27441182
"Cats"
https://lemmy.today/pictrs/image/b97e6455-2c37-4343-bdc4-5907e26b1b5d.png
I could not distinguish between that and a photograph. It doesn't have the kind of artifacts that I could identify. At the time, I was shocked, because I hadn't realized that the Flux people had been doing the kind of computer vision processing on their images as part of the training process required to do that kind of lighting work at generation time. That's using a model that's over a year old
forever, at the rate things are changing
from a non-expert on just local hardware, and was just a first-pass, not a "generate 100 and pick the best", or something that had any tweaking involved.
Flux was not especially amenable, as diffusion models go, to the generation of pornography last I looked, but I am quite certain that there will be photography-oriented and real-video oriented models that will be very much aimed at pornography.
And that was done with the limited resources available in the past. There is now a lot of capital going towards advancing the field, and a lot of scale coming.
I mean, that looks AI generated to me. In particular it looks like a 'smooth skin' and shiny FLUX image, which is kinda that model's signature.
It's not bad though.
It's like a lot of AI content where if you're just scrolling past, or not scrutinizing, it looks real enough. I'm sure soon it will take lots of scrutiny to distinguish.