- Joined
- Dec 28, 2014
Is anyone making or even contemplating ASICs (or even FPGAs) specifically tailored to AI tasks?
Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump
- Thread starter Near
- Start date
- Joined
- Jan 28, 2018
Oh many people, I think the most known ones are groq and their LPUs (language processing units). You'd need quite beastly FPGAs for (smaller) LLMs, at that point the GPU hardware is cheaper (even though latency and power efficency would be better on the FPGA). groq are the only ones I'm aware of that actually have an already presentable product you can even use via their API.
There's currently a lot of noise around bitnet (1.5 bit quantization method, basically ternary models) and banishing matrix multiplications from the process, which would make efficient, bespoke hardware for LLMs more viable and would be a complete paradigm shift in hardware requirements for even the largest models. It actually would even get pretty viable to run big LLMs on normal CPUs. It's too early to tell if these actually will lead anywhere though. A bitnet 70b parameter model would fit into ~20 gb of ram. Downside with bitnet is that the model needs to be made for it from the get go, it's not something you can do to existing models. The original bitnet paper people haven't released their models but Nous managed to reproduce them back in March.
That's what I meant with my earlier post, it's a bit early to tell which direction this will go.
- Joined
- Dec 3, 2020
The integrated GPU on your CPU which uses system RAM, you can force your browser or any other program to only use that and leave your real GPU alone. Is what I been doing for years when playing a game on one monitor while still having my work stuff on the other monitor.
List them, sharing is caring.
Which is...?
Don't all those sites for stable diffusion "art" (pr0n mostly) and DIY LLM-roleplay count as consumer tools? no idea if any are making real money or even breaking even tho.
Yeah AFAIK one company called Etched is making transformer ASICs, the numbers they claim are almost too good to be true and they just got 120mil to make the first chips codename 'sohu' so its too early to tell.
As for groq, IDK if their LPUs count as ASICs tho, plus no idea if they even plan to sell those or just set their own datacenters, anyone knows?
Heard of it, didn't read much but there was talk about precision issues vs matrix-based?
- Joined
- Aug 28, 2015
Having worked a long time in networking, now having flashbacks to ternary memory noooooo
- Joined
- Nov 21, 2020
I stick to CPUs that are CPUs. iGPUs are for laptops.
Yes, I know the next generations I won't likely have a choice.
- Joined
- Nov 15, 2021
That's the consumer side of AI. The production side is the development and calibration of the model. Nobody's planning to make tools to create your own LLM that are accessible to your average retard with a Chromebook and zero knowledge of programming.
- Joined
- Dec 3, 2020
Practically every CPU has an iGPU now because normies can be bothered with getting a GPU or can't afford it now that even a midend GPU costs almost as much as the rest of the computer.
Also useful for when your videocard decides to shit the bed and you don't have a spare at hand.
Should've said "prosumer" since normies aren't using SD but bing as the former is too complicated for them.
- Joined
- Oct 12, 2015
As promised, here's KiloNegative and seven more goodies:
KiloNegative LoRA
Nightcore
Global Scripts
MiniNegative
Reaverbots
Vintage Shojo Manga
Anime Promo VHS Tape
Monsho
KiloNegative LoRA
Nightcore
Global Scripts
MiniNegative
Reaverbots
Vintage Shojo Manga
Anime Promo VHS Tape
Monsho
- Joined
- Dec 18, 2022
Monsho and Anime Promo VHS Tape are gorgeous.
- Joined
- Dec 3, 2020
No SDXL/Pony version?
BTW what are some alternatives to Civitai? seems every LoRA comes from there now.
- Joined
- Mar 21, 2019
I know it's not the point but I feel pure disdain for this concept that your gender can be defined by you dressing up like a doll and putting on obnoxious make up. Particularly when your absurd costume is hyper feminine and it somehow makes you express being non-feminine because you have a septum piercing. On topic, though, I'm mostly just embarrassed by that AI avatar. Making a whole ad campaign to show off your tech and then having an uncanny valley glassy-eyed avatar flail about unconvincingly on screen (which required a full 3D body scan and a bunch of data input from her tiktoks) is just...it's a marketing choice, I guess, but to who? The crowd who appreciates the body positivity will either hate AI or be informed enough to know that's pathetic, and the people who aren't informed won't approve of the message.
- Joined
- Oct 20, 2019
No, he was correct in his terms. Services like Civitai's generative UI may be focused towards the unwashed masses who are "consumers", but the people building the service themselves are professionals because they're building and running that service. Right now there is overlap in hardware for both because AI research opportunistically built on existing GPU tech because it was there and widespread. But who is to say that this alignment of tech between two different audiences isn't some anomaly. The driving factor in hardware is professional use right now because it's a seller's market - nobody can produce enough hardware to keep up with demand so manufacturers are chasing those with the most money.
Perhaps that will change. And perhaps that will lead to the overlap continuing as AI is used in modern games for example. But I believe his terms were correct to his point.
Yeah, it's sad on multiple levels. That somebody would want to interact with their loved ones via a fantasy version of themselves rather than their real self. After all, who is more important to accept you for who you are than your family? That the actual tech demo'd is frankly so... bad. It has all the hallmarks of a marketing idea that didn't turn out as hoped but was rushed out the door anyway. And finally in that Lenovo, a business focused company, thought this was good marketing for their product in the first place.
- Joined
- Dec 3, 2020
- Joined
- Sep 15, 2016
Heya fellas!
Just got a hold of an M3 MBP, so I figured I'd try out AI image/speech generation.
Anyone got some recommendations to which tools to use? Apple's App Store is a dreary wasteland of subscription crap, so I've mostly given up on that angle.
Just got a hold of an M3 MBP, so I figured I'd try out AI image/speech generation.
Anyone got some recommendations to which tools to use? Apple's App Store is a dreary wasteland of subscription crap, so I've mostly given up on that angle.
- Joined
- Oct 20, 2019
- Joined
- Sep 5, 2023
Posting about some funny drama I found.
This repo is a tool that lets you reverse a finished image into a series of line drawings at various stages of progression, from blank canvas to finished product. It was obviously made with the intention of trolling xitter artcels who previously were using their line drawings as proof that they weren't using AI to generate their furry cub porn inflation vore scat art. The release of this tool has been quite controversial, and the artcels have taken to the issues section of the repo to voice their anger. Archive links because GitHub admins are now involved, due to numerous thinly veiled threats and general raging that violates the AUP
Thanks for destroying art
Are they idiots? This is evil
Why would you pretend this is not AIGC if you are proud of AIGC?
AI bros belike "lol I can generate*ARTS* by simply writing prompts" while tryin' hard to mimic the labor done by REAL ARTISTS doing REAL ARTS
What's the actual use case of this?
Dataset and training details ?
After several comments were removed for misbehaving, a new thread was opened to contain all the coping and seething in one place: Mega Thread for Non-Technical Discussions
I feel for artists who are going to lose their cushy jobs. I know I'm probably not long for my own field, with LLMs getting better by the day. But damn if this isn't funny
This repo is a tool that lets you reverse a finished image into a series of line drawings at various stages of progression, from blank canvas to finished product. It was obviously made with the intention of trolling xitter artcels who previously were using their line drawings as proof that they weren't using AI to generate their furry cub porn inflation vore scat art. The release of this tool has been quite controversial, and the artcels have taken to the issues section of the repo to voice their anger. Archive links because GitHub admins are now involved, due to numerous thinly veiled threats and general raging that violates the AUP
Thanks for destroying art
Are they idiots? This is evil
Why would you pretend this is not AIGC if you are proud of AIGC?
AI bros belike "lol I can generate*ARTS* by simply writing prompts" while tryin' hard to mimic the labor done by REAL ARTISTS doing REAL ARTS
What's the actual use case of this?
Dataset and training details ?
After several comments were removed for misbehaving, a new thread was opened to contain all the coping and seething in one place: Mega Thread for Non-Technical Discussions
I feel for artists who are going to lose their cushy jobs. I know I'm probably not long for my own field, with LLMs getting better by the day. But damn if this isn't funny
- Joined
- Nov 14, 2022
I don't.
Everyone who loses a cushy job over LLMs is, by definition, a marginal artist, something that doesn't deserve respect
- Joined
- Jun 17, 2021
@grand larsony In a deeply ironic twist, my frustrations with style training led to me just learning how to draw in the style of the artist I was trying to copy. Wound up making a couple grand from fans of the artist that wanted full-art versions of his MTG cards made.
I think longterm AI's gonna have a decent foothold on the braindead art tasks like advertisements and randomly generating character features, but I don't see the impact on the entertainment industry being anywhere near as big as people hope/fear it will. And my main reason for that assumption is the fact that we're 2 years into GenAI being in the public's hands, yet I haven't seen any good games or animations using the image side of GenAI in a major way. For games all I've seen have had the same level of polish as most asset flip games, and the only animations I'm seeing are either slideshows or uncanny parallax shots with really robotic character motion.
I think as far as the independent/contract artists go, they're probably gonna have to brace another year or two of the suck, but those that are able to wait things out will be finding themselves with plenty of work redoing projects marred by poor use of AI.
Last edited:
- Joined
- Mar 1, 2020
That AI driving that meme generator site is beyond savage.
I kneel. Humanity has been utterly mogged by the machines. We will never fucking recover as a species from this:
- Joined
- Dec 28, 2014
Whoever said that bullshit about vacuum tubes has never had a Mesa Boogie amp.
Oh it's an AI. Confirming that AI isn't even as smart as insects yet.