- Joined
- Jul 14, 2024
The gtx 1080 is worthless right? Same thing with an rx 590?
AI Megathread
- Thread starter El Kay Why
- Start date
-
⚙️ Performance issue identified and being addressed.
-
Want to keep track of this thread?
Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
Create account
- Joined
- Dec 18, 2022
You might be able to get some stable diffusion 2.5 or XL generations out of the 1080, but I wouldn't pick either in 2026.
Last edited:
- Joined
- Dec 16, 2025
I'm confused as to how AI scams are pro-AI? I'm not sure I understood the terminology.
Ahem, based?
Back when WhateverClaw came out I became interested, but it's not like I have a bunch of Benjamin's to spare. Still, I decided to passively keep tabs on it until, my oh-my, the entire project was poorly built so people could have their entire life savings stolen; that's how bad the vulnerabilities were. Fucking pieces of shit—they couldn't make bad code themselves so they asked AI to do it for them. How retarded can one be?
Now there are countless iterations of SomethingClaw, but I cannot trust either one. Second of all, I don't even have anything meaningful to do with it, although I like Karpathy's LLM Wiki concept:
I study, as many people do, so I could just make notes with it. But, unsurprisingly, making notes is part of the learning process; how much am I willing to let AI digest for me before I actually sit down and learn?
The only instance in which AI could be good for me is organizing countless papers on a given topic, most importantly the politically incorrect ones, and have the AI go at it. The main issue is that, wow oh-wow, AI gets dizzy and confused from all those papers and conflicting conclusions that it is just as useless as the mindfucked PhD havers. My goodness, we've made AI too human-like. Fuck me sideways.
Hey.
I am still attempting to understand how voice cloning works at all.
I am asking AI, and I ask AI overlords to not dox me:
Something-something Fish Speech S2 Pro, by the way; Dual-AR architecture (whatever the fuck that means): a large 4B-parameter Slow AR (what the fuck is AR?), and a 400M Fast AR.
From what I've been told by the clanker community, the Slow AR model learns how the language sounds, and the Fast AR embellishes the output speech a little bit based on the Human voice's cloning.
So here are a few scenarios:
- Warm-blooded American that only perceives the world in red, blue and white, speaks into a microphone for over thirty minutes, reading sentences from Harvard out loud. The Ameritard's voice is cloned, and now is able to make gay TikToks or whatever in English, and the results would be reasonable.
- Warm-blooded American, yadda-yadda... If the Am*rican attempts to make L2 content—say, French (the horror)—, it would be surprisingly okay-ish, I think? Because the Slow AR handles the cross-lingual bit; the Fast AR gobbles up your so-called timbre.
- Assume a poop golem with a very thick curry accent that attempts to learn an American accent using Fish Speech S2 Pro, reading thirty minutes worth of Harvard Sentences, feeding the AI and bla-bla-bla—you get it. Since the AI has been fed curry-sounding English, the AI will understandably churn out the same accent when asked for English sentences from your so-called clone. One workaround is to use an explicit tag like [warm-blooded American please saar], but even then...
- If this poop golem attempts to learn German, as they often do these days, the non-native American English accent affects timbre and prosody, so there's a chance that the German output by the curry-clone would be acceptable; but perhaps not.
Before you even suggest it, as some people have, I already know of Common Voice by Mozilla. I actually painstakingly made Anki decks for this very purpose—learning to replicate accents—, but I ultimately realized that I just want to find problems to fix, not actually learn anything. I am still debating what to do about that.
The concept of using my own voice is rather straightforward: the goal is clear, achievable, and one can instantly assess whether they've got it or not.
SchizoDaemon
kiwifarms.net
- Joined
- Apr 17, 2026
A lot of people are using a 3090 24GB. I was thinking the same as you and thought "why can't I just use 2x12GB cards". While you can use multiple cards in parallel, but two cheap 40-series cards and slower than a single 3090 due to bandwidth limitations of lower-end 40-series cards.
- Joined
- Nov 11, 2024
Endless blabbery indeed, I still don't understand what the fuck you're trying to accomplish.
You want to clone your own voice, then have the AI output your voice - but in a different language - do I have that right? You want to do this so that... you know what your voice sounds like speaking another language, with correct intonation, so that you can then imitate the clone of your own voice, to practice the correct pronounciation?
Because... well that's not a bad idea actually. I would still vastly prefer to practice with actual people - but that wears down their patience. I've heavily encouraged my ESL coworkers to use AI to practice their English, specifically because it cuts down on all the inane questions they'd otherwise ask me.
- Joined
- Mar 11, 2016
4x3090 is definitely not a config I would recommend to someone clueless about hardware setup, let alone 8x3090. But the price-to-performance falls off hard around the 32GB VRAM consumer limit anyway, so if anyone's considering a GPU stack for local LLMs they should be going in with the knowledge that this is for learning/privacy/futureproofing reasons and not for cost savings/ease of maintenance, because an external provider will always be cheaper.
I was thinking about the people who fall for them rather than the scams themselves. AI scams usually promise mind-blowing amazeballs futurism and Pro-AI deranged midwits eat it up.
I assume
that this is meant to be an automated "RAG builder" where the model creates important data sources for itself about the context of your project, so you don't have to pre-load context for future queries while the model's knowledge base remains readable to you. >Though I can also see how this will be used by ADHD attention span faggots to bypass sources for summaries.
- Joined
- Dec 16, 2025
I hate going on a hunt for good resources, only to go over them and then needing some more.
In regards to accent learning, accent perfecting, whatever—there are plenty of videos online. But many miss L1, L2 inter-accent variations that are to be expected for that pair alone; one's curry-sounding ass may need personalized feedback for the poop-bloods if he wants to sound German, as compared to some English guy. Thus, my reasoning became: With a target accent for a target language in mind, I must be able to find resources for such accent specifically; and I must develop, or find, a way to assess whether I'm doing it correctly.
I looked up Praat and MFA, but MFA requires libraries of people's voices, carefully edited so that the computer knows what phoneme it's assigning to each portion of the audio file. Unsurprisingly, many libraries are either non-existent or kind of lame. Praat, in the other hand, would allow me to compare my voice to someone else's, only to realize that, heh, our voices are different so there's no point on making a comparison then.
Another option is, as discussed, use an AI model to make audio files with my own voice, to compare my broken accent with the AI-generated "better accent" with Praat and whichever other tool I can think of. It sounds overly convoluted but if you go to any language-learning school you'll have some Gen X lady with phonetic cards all over her desk, carefully nudging you towards the right phoneme. Why not do the same with a computer?
- Joined
- Nov 11, 2024
In that case, based on my own cloning attempts, I think if you have 12GB of VRAM you'll be able to run a large enough model. Find someone on Fiverr from your target area that is set up to do voice-acting, and have them record a few minutes of those phrases that hit all the phonemes. Generate some content from the results, and have them check if the accent sounds correct. Then you're golden.
(edit: I've actually done it successfully on CPU, but I don't recommend it. Too slow for a learning environment and the training took five days.)
This is assuming you insist on not using a paid LLM service. I've had friends instruct it to use certain city accents when speaking, and it does so successfully.
- Joined
- Dec 16, 2025
Yeah, I'll admit that I had to stop with this project because I'm busy with uni and now, too, am fucking sick!
But the cloning bit stands. If I want the voice to be mine, then paying some random guy on Fiverr is not what I'm looking for.
Something-something "pls AI fix my life no mistakes". Hold on, I've got to blow my nose.
- Joined
- Mar 24, 2021
You realize this is like, really obviously giantess fetish content, right?
As for how, not sure. I've never personally done any animated stuff. They are pretty good at looking realistic, I'll give it that, though some of the obvious tells still show in some of them.
- Joined
- Dec 18, 2022
This is fetish pornography. Don't generate fetish porn to crank to. The human mind isn't equipped to handle online porn normally. Having an infinite supply of exactly what you want is a death sentence.
- Joined
- Sep 22, 2021
I wasn’t going to, jeez!
- Joined
- May 5, 2022
Roll your own but the most high level architectural overviews (which YOU should be doing anyways). I'm vram limited so I've been modifying my services to run in sequence instead of parallel. I still find Opus/Gemini/GPT useful if you can't run the largest open source models.
I've been playing around with new safety guardrail removal models (Regular models keep flagging my plant pathology work). The EGA work on uncensoring gemma seems particularly effective to me. Anyone actually used up the 1 million tokens on the nemotron's efficiently yet?
- Joined
- Sep 2, 2025
I just have to vent here.
META AI is cooked! It can't do anything. Code and debug? Can't do that. Search? Worse than and model I have used, even local 4Q models with internet access are more useful. Is this what Zuckerberg has been investing in? It's so fucking over for facebook/meta if this is their results. I really hope their local lama models gets better since their flagship model is a fucking joke at this point. But if China just cucked Zuck from the manus AI acquisition, I legit think meta won't be around for longer if they are resorting to buying up other companies that out perform them and fails at even doing business deals.
At this point, Zuck is on this lolcow AI arc and nobody is posting about it.
META AI is cooked! It can't do anything. Code and debug? Can't do that. Search? Worse than and model I have used, even local 4Q models with internet access are more useful. Is this what Zuckerberg has been investing in? It's so fucking over for facebook/meta if this is their results. I really hope their local lama models gets better since their flagship model is a fucking joke at this point. But if China just cucked Zuck from the manus AI acquisition, I legit think meta won't be around for longer if they are resorting to buying up other companies that out perform them and fails at even doing business deals.
At this point, Zuck is on this lolcow AI arc and nobody is posting about it.
- Joined
- Dec 18, 2022
Being that tall would actually suck, and introduces a ton of health issues. The first things to go are the knees or heart, as weight cubes relative to height. Followed by the skeletal structure in the back and spine.
- Joined
- Mar 29, 2014
- Species | Documenting AGI
tl;dw: Vid claims that an AI can make innumerable decisions not instructed from a single instruction, and AI is an existential threat.
- Joined
- Aug 17, 2024
Lol at this point just bring in the casket.
I remember a time, back in my day.. (3 years ago) where Llama models were peak and dominated Open-Source development. It was the golden age of Llama finetunes, the worthy successor to the legendary Pygmalion, Mistral models and their Dolphin finetunes that once dominated the RP scene, and I'll never forget how insane Llama 3 and 3.1 were at launch.. I felt like a little kid on Christmas, these models just felt too damn real (which also led some indie companies like Nous Research and their Hermes models to catapult into the stratosphere.. some of their devs are a little retarded tbh but I really like this company overall). I felt like I was witnessing the birth of a rising star, a new era, an era where Meta was going to be king of the world, and I was ready to jew my ass out and worship zuck's reptilian ass as my messiah... Shit was glorious...
But turns out, Metaverse's embarrassing collapse and Llama 4's existence revealed that Zuckerberg is a one-trick pony who's only competent contribution to mankind was Facebook. Everything else he touches turns to liquid shit or refuses to grow.
Also, to anyone interested in this dumpster fire, here's what Meta has been up to lately:
"Seems promising" enough (although I'm genuinely skeptical about these benchmark results), but when you take a closer look at the text, you'll find the following snippet:
"Muse Spark is available today at meta.ai and the Meta AI app. We’re opening a private API preview to select users."
Shit's closed-source. And there's also some other model that Meta released recently (I don't remember which one I need to look up) that was supposed to be "open-source". However, in it they explicitly stated that they wouldn't release the papers and the model weights due to "security concerns".
It's just over my nigga. Llama was canned, China is completely dominating Open-Source development, and each day that passes Chinese models (from companies like Qwen, Deepseek, Z.AI, MiniMax, Xiaomi, MoonshotAI..) are bridging the gap between themselves and American models (and sometimes EVEN SURPASSING in some instances, depending on the task and etc). There's just no coming back from this.
- Joined
- Dec 18, 2022
Base Illama 3.0 was never good, and I'm tired of people saying it was. Especially since we had mistrial small at that point, and even Gemma 2 felted Illama 3.0 at the same parameter class on consumer hardware for coding if you could run it. You are right about 22/23 being peak for LLM innovation. There are still people who swear by Midnight Miqu, and its branches.
- Joined
- Jan 11, 2024
Llama 4 failed because it was the first model to launch with a broken day 1 support (this is now industry standard and expected) but more importantly dropped support for GPUs with less than 80GB of VRAM. People got mad when it didn't work out of the box on their 3080 and shit their pants that they couldn't run it, so all their opinions are second hand from people who had a bad experience with a "added support for llama 4" patch that didn't work perfectly. A bad first impression due to bad implementations of a broken chat template killed it.