AI Megathread

[snowdunes]

If the fucking Europoors somehow come out of nowhere with a Mythos-tier model after years of being passed by even the Chinese, let alone every single American company, I will buy a bidet, turn off my AC for a day, and learn how to play soccer.

This is why I love the AI market right now, anything can happen.

US was kneecapping China only for them to come out with Deepseek and GML that is much more optimised with their limited compute and are up there with the leading frontier models, while Anthropic is struggling with compute despite being in the US (don't know if true, but from what I have heard their models eats more per input/output compared to other models which is hinted at their API prices, their strenght came from being able to run their stuff on any data center, not being dependent on a single provider and could scale fast).

Who knows, maybe the French got an ace in their sleeve or they are about to be mogged. Their R&D on custom chips makes me think they are really serious about their larger models.

 

[ringtones]

the best stable diffusion models are all Chinese too. Illustrious, Qwen, Wan2.2, all Chinese

 

[SchizoDaemon]

On that front, I think a lot of very weird shit is going to happen as a result of the limits being taken off of coomers' novelty drive.

There was an AI model years ago on like OF, Twitch or something, and they were getting more cash than the real ones back then.

I know there are loads of Chinese cam hoes those crazy filters, but I think this was like full AI responses etc.

 

[macrodegenerate]

the best stable diffusion models are all Chinese too. Illustrious, Qwen, Wan2.2, all Chinese

I thought Illustrious was a finetune of sdxl.

 

[The Gay Oboma Creature]

US was kneecapping China only for them to come out with Deepseek and GML that is much more optimised with their limited compute and are up there with the leading frontier models,

GLM-5.2 seems impressive; I was skeptical up until then but there doesn't seem to be as much pushback re: benchmaxxing this time around.

I should note that China has its own fairly strong chip industry, even if it's not on America's level quite yet, and they have the advantage of a tech industry not burdened by a horde of jeets and other affirmative action creatures. France, on the other hand, has been thoroughly brain-drained by America and is dealing with all of the same PC burdens on top of that.

---

the best stable diffusion models are all Chinese too. Illustrious, Qwen, Wan2.2, all Chinese

Nano banana pro has some rough edges, but it's generally seen as the best, still. Don't get me wrong, love me some open-source Chinese machine learning models, but they're not quite there yet.

Also, stable diffusion is a brand name, you just mean diffusion unless these models are all somehow fine-tuned from it, which seems unlikely.

 

[oreoless effect]

Read God Shaped Hole, by the way. The author's gone completely up his own ass since, and some of that pretension is there even then, but it was written early enough in his career that it feels like 2015 4chan wrote a book about the future of AI, and I mean that in the most flattering way. The tech buzzwords are dated (GANs are still used surprisingly often, especially because of the useful properties of VAE-GANs, but diffusion models are the new hotness) but the rest holds up.

Hey this is random but thanks for recommending this. I read it all in one sitting this afternoon and I think it's the greatest thing I've ever read besides The Feminist by Tony Tulathimutte.

I'm not plugged in to right wing twitter or LW/rationalists, and the little I could gather about the author is that he may have been part of both communities until dropping off the face of the earth. From reading this story and realizing it was written 7 years ago and, without powerleveling too much, accurately predicted some aspects of my life to an insane degree, I would wager this guy is like a literal 200 IQ Chris Langan type. What can you tell me about his history and "going completely up his own ass", since it seems like everything about him has been scrubbed? If you don't want to derail the thread please send me a DM, thanks!

 

[The Gay Oboma Creature]

I think it's the greatest thing I've ever read besides The Feminist by Tony Tulathimutte.

I'll have a look, thanks.

What can you tell me about his history and "going completely up his own ass", since it seems like everything about him has been scrubbed?

Nothing especially exciting. Standard unpopular-in-grade-school kid joining a 'cool kids club' on the internet and getting full of himself. Way too obsequious towards BAP and his group's groupthink, made it his entire identity and stopped writing anything interesting; his last fiction work was one page and consisted solely of retarded inside jokes with his "friends". Gradually became insufferable.

That said, he wrote other good stuff before that, almost all relevant to AI so I'll post them here:

• Story of a boy with an AI assistant in his head. Push past the weird writing style, you'll like it if you liked the first story.
• Story about AI using gig economy apps. Kind of a relic of the crypto era, but still neat. The occasional puff piece on an LLM hiring a human to do a real world task was predicted by it, as was AI psychosis, in its way. 4o's engagement-maximizing parasitism has natural parallels here.
• An early story, not directly AI relevant, but good and short.
• Another good short story, not directly AI relevant but who knows. Turning point for him was when he became ashamed of it and cited it as his least favorite work. IMO it's his best.

Also, a not-a-story on writing fiction with the very first LLMs. Interesting to see what issues persisted and which ones didn't. I can imagine an abliterated GLM 5.2 being much more interesting for these purposes.

 

[ringtones]

Gemini provides a recipe for Lean (just like Jaqual in the hood used to make)

 

[debristar]

vibe coding

hybrid programming

 

[macrodegenerate]

hybrid programming

No. Hybrid programming implies a combination of an actual software developer driving the AI, and refactoring while making a spec.
Vibe coding is figuring out a spec or basic idea, and gambling tokens until you roll the right code generation, and get what you want. No thoughts required.

 

[debristar]

Vibe coding is figuring out a spec or basic idea, and gambling tokens until you roll the right code generation, and get what you want. No thoughts required.

vibe coding is like driving without destination.

and gambling tokens until you roll the right code generation

so, throwing money away, because you will never one shot the kind of solution you're imagining.

and get what you want. No thoughts required

tell me what you want without thinking.

 

[macrodegenerate]

vibe coding is like driving without destination.

so, throwing money away, because you will never one shot the kind of solution you're imagining.

tell me what you want without thinking.

You want without thinking.

 

[Reversal]

I'm trying to set up something local and was wondering if anyone here had any advice. Basically, I'm trying to do image>video. The idea is I want to try and animate stuff I draw. Problem is I only have 12GB of vram. Is this even feasible? or not even worth the trouble?

 

[ringtones]

depends on the card. I have a 3090 and can make a shitty video in Comfy

out of about 40 attempts I have about 5 7-second videos that came out as what I'd call "acceptable"
of those videos only one is "good"

so I have a ~10% hit rate

unless you have a 40 series card or better I think it's not worth trying

 

[El Kay Why]

Anyone try Venice over OpenRouter?

 

[snowdunes]

Does anyone how much GLM 5.2 you can use with OpenCode Go? (the $10 a month sub)?

Claude feels like a scam at this point but it's the only really good one (I refuse to pay for Max after seeing how Anthropic treating their customers). I have discount on Mistral AI but I don't know it it's even capable of planing an generating code? I don't want pay OpenCode Go/Mistral Vibe only to get nothing done due to limit or not useful outputs.

Yes, I'm aware I'm in the subscription dependency exploitation AI maffia hellhole, don't need to remind me how bad it is. I can't build a 1 TB uniform RAM PC at the moment...

 

[Legalize Asbestos]

I'm trying to set up something local and was wondering if anyone here had any advice. Basically, I'm trying to do image>video. The idea is I want to try and animate stuff I draw. Problem is I only have 12GB of vram. Is this even feasible? or not even worth the trouble?

Is someone finally going to try the in-betweening thing that Youcis suggested? Having the model auto-generate/duplicate between key frames? Because I don't think you'd need too much horsepower to accomplish that, since you're asking it to tweak images rather than invent them wholesale.

There's a few youtube animators that mix hand-drawing with AI as well, though the guy I'm thinking of escapes me atm.

 

[macrodegenerate]

I've come to the conclusion that 31B with a high context is all you need if you are coding agentically. It's best used as a helper capable of duplicating patterns which you've already implemented. Sure if you are driving a ford Ford GT you are going to beat a compact on a straight drag, but at the end of the day coding breaks down to something more like LeMans over time. Knowing the track through experience and handling is the most important in getting to the finish line at the end of the day.

 

[DavidS877]

I've come to the conclusion that 31B with a high context is all you need if you are coding agentically. It's best used as a helper capable of duplicating patterns which you've already implemented. Sure if you are driving a ford Ford GT you are going to beat a compact on a straight drag, but at the end of the day coding breaks down to something more like LeMans over time. Knowing the track through experience and handling is the most important in getting to the finish line at the end of the day.

I've been flipping between Qwen 3.6, Qwen 3 Coder Next, Gemma 4 and GPT-OSS-120b in writing the code for my LCD display for my retro gaming system. I need to do more formal testing, but other than 3.6 sometimes getting stuck in loops they've all seemed to figure out what I wanted and gotten it done. Admittedly it's bog standard C on a microcontroller.