ChatGPT - If Stack Overflow and Reddit had a child

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
More news from the coomersphere, the latest edition of Novel AI's image gen is complete ass at the one thing everyone wants it to do (copy the artstyle of existing artists and mixing artstyles) and it is actively devolving into a gigantic meltdown as PR and community managers fail to spin it. And that's a paid service, NAI is not cheap.
It's so bad that it's compared to SD 1.5, with jannies reacting to any questioning of the new model with mutes and bans.

1740818096140683.png1740820451003089.png1740827706337075.png
>you don't NEED to prompt porn anyway, what do you pay $25 per month for?

The meme reaction wave is "reimugate", where they challenge the new model to draw Reimu with extremely questionable results.

1740828931925074.png1740828369184348.png1740826051450852.png1740824589534125.png

As part of their newly announced open source and transparency initiative, Deepseek reveals that their profit margins for their API R1 service ($0.5/MTok output) is... 545%. For comparison, Openrouter providers offer unmolested R1 at ~$7/MTok output. Imagine what the margins on expensive models like OAI's are.

deep.png
With their current usage at least, they're trying to shift peak demand to off-peak hours with cheaper prices, which at current capacity would raise the margin to ~750%.
1frq9e.jpg
Theoretically, of course. The web version is still free.
 
Last edited:
More news from the coomersphere, the latest edition of Novel AI's image gen is complete ass at the one thing everyone wants it to do (copy the artstyle of existing artists and mixing artstyles) and it is actively devolving into a gigantic meltdown as PR and community managers fail to spin it. And that's a paid service, NAI is not cheap.
It's so bad that it's compared to SD 1.5, with jannies reacting to any questioning of the new model with mutes and bans.

View attachment 7041200View attachment 7041201View attachment 7041274
>you don't NEED to prompt porn anyway, what do you pay $25 per month for?

The meme reaction wave is "reimugate", where they challenge the new model to draw Reimu with extremely questionable results.

View attachment 7041199View attachment 7041270View attachment 7041272View attachment 7041273

As part of their newly announced open source and transparency initiative, Deepseek reveals that their profit margins for their API R1 service ($0.5/MTok output) is... 545%. For comparison, Openrouter providers offer unmolested R1 at ~$7/MTok output. Imagine what the margins on expensive models like OAI's are.

View attachment 7041204
With their current usage at least, they're trying to shift peak demand to off-peak hours with cheaper prices, which at current capacity would raise the margin to ~750%.
View attachment 7041279

Questionable context in these screenshots. Especially this one, the smut comment was directed towards someone that said they can't try the new model because they are in public, insinuating that whatever they would generate wouldn't be fit to be seen in a public setting. I know it's hard for current year people to understand humor after multiple layers of irony poisoning and undiagnosed autism, but believe it or not, this is a "joke". (But it's really hard to tell that if you just snip a response without the comment it's responding to.)
1740843841226.png
V4 Full does generate explicit NSFW, while Curated didn't/doesn't. Which follows NAI's previous already established model format, is one model for SFW scenarios (streaming, normies), and a model with explicit NSFW content in the dataset.

And all the Reimu images are lacking context of what is wrong with them and how it conflicts with the apparent prompt coherence. And there are no prompts on these Reimu images so you have no idea what the person was going for. But what I'm getting at is some claim that NAI V4 gives female characters "gargantuan breasts" by default, which I cannot replicate or seen myself. I copied that exact prompt and used only these negatives (no presets enabled):
blurry, lowres, error, film grain, scan artifacts, worst quality, bad quality, jpeg artifacts, very displeasing, chromatic aberration, multiple views,
Copied Reimu prompt:
Hakurei Reimu standing alone outside a traditional shrine, torii gates behind her. She wears her classic red and white outfit, with long sleeves and ribbon-adorned hair. Her large breasts are accentuated by the fabric, rising and falling naturally. The setting is peaceful, with sunlight filtering through the trees, casting soft shadows.
, best quality, very aesthetic, absurdres,
e1803bc0-a13b-404b-823e-9fe8f4a20ddd.png
Alternate LLM augmentation of the Reimu prompt:
Hakurei Reimu stands alone before the weathered steps of a traditional Shinto shrine, exuding an aura of quiet resolve. Behind her, a series of striking red torii gates stretches into the distance, their bold arches contrasting with the lush greenery that envelops the sacred grounds. She wears her signature red and white miko ensemble, the long sleeves billowing faintly as they catch the air, adorned with subtle stitching that hints at her role as guardian. Her hair, a cascade of dark strands, is tied with vibrant ribbons that dangle playfully, swaying with her every movement. The fabric of her outfit drapes over her form, accentuating her pronounced bust, the gentle rise and fall of her chest lending a natural rhythm to her stillness. The surroundings are steeped in peace, with golden sunlight streaming through the branches of towering trees, dappling the ground with soft, shifting shadows that dance across her figure and the shrine's stone path. The scene is captured in a digital illustration with a semi-realistic flair, blending crisp outlines with tender shading to highlight the texture of her clothing and the serene interplay of light. The color scheme pops with the vivid reds and whites of her attire set against the muted greens and browns of nature, creating a harmonious yet striking contrast. A medium close-up frames her elegantly, with a slight tilt in perspective that emphasizes her calm authority amid the tranquil setting. 1girl, solo, Hakurei Reimu, shrine, torii gates, miko outfit, red and white, long sleeves, ribbon in hair, large breasts, peaceful, sunlight, shadows, traditional, Shinto, spiritual, digital art, semi-realistic, vibrant colors, natural light, textured, serene, guardian
, best quality, very aesthetic, absurdres,
3f09691a-7585-4c63-ac54-4044a27c0a00.png
I don't know if Kiwi Farms strips metadata, but if it doesn't you can confirm these outputs by dragging them back into NAI's SaaS.

Edit: I'll go one level deeper and investigate the tianliang_duohe_fangdongye artist replication problem.
And this has to do with how you actually prompt artists in V4, and potentially not knowing the format as I cannot see any actual documentation for a prepend.
Hakurei Reimu standing alone outside a traditional shrine, torii gates behind her. She wears her classic red and white outfit, with long sleeves and ribbon-adorned hair. Her large breasts are accentuated by the fabric, rising and falling naturally. The setting is peaceful, with sunlight filtering through the trees, casting soft shadows.
tianliang duohe fangdongye,
eccc9018-3612-4364-88b6-553ebc17f1dc.png
Hakurei Reimu standing alone outside a traditional shrine, torii gates behind her. She wears her classic red and white outfit, with long sleeves and ribbon-adorned hair. Her large breasts are accentuated by the fabric, rising and falling naturally. The setting is peaceful, with sunlight filtering through the trees, casting soft shadows.
artist:tianliang duohe fangdongye,
3fb4714b-cc30-4c21-abc6-a298778a7656.png
Hakurei Reimu standing alone outside a traditional shrine, torii gates behind her. She wears her classic red and white outfit, with long sleeves and ribbon-adorned hair. Her large breasts are accentuated by the fabric, rising and falling naturally. The setting is peaceful, with sunlight filtering through the trees, casting soft shadows.
artist:tianliang duohe fangdongye (artist),
aa2ff96d-237e-40b2-8f2b-eb3221be38f8.png
Looks too saturated to resemble the actual artist's work, regardless of the guessed artist tag prepend/append.
But wait a minute, I know this song and dance. Generative image models have a specific quirk where they lean towards higher saturation and worse color histograms. And there is a specific option known as CFG Rescale in local frontends like WebUI and ComfyUI that exists to mitigate this oversaturation, and produce a better histogram. (lowering saturation in post isn't the same) And NAI has an equivalent option.
1740846908971.png
All previous outputs were with no Guidance Rescale. So what happens at 0.7, a typical CFG Rescale value:
358efe8f-0c6b-4411-84c4-a6ac794a5239.png
I hardly even participate in generative image stuff anymore, but I'm cursed with understanding it and watching other people not have a clue how to even use it.

(Kind of hate that this is a response to image diffusion talk in the ChatGPT thread and not the more appropriate thread. I should of quoted it over there. Sorry. (I probably feel more shame at this unironic strip of weeb generations.))
 
Last edited:
I don't understand. Local AI image generation has gotten really good. Why would you waste money on some subscription service when you could pick up a somewhat old nvidia card for cheap and generate images with that? My old 1050ti could generate images, albeit slowly, and that thing is dirt cheap now. And there are better cards for maginally more you could get. Especially since with local generation you can finetune everything you want.
 
What the fuck is this price? Opus is $15/75 with being the best model for creative writing ever released and it's still not worth it, this is more than double.
$75/Mtok input, really?
"anyone stupid enough to not just use a modified deepseek on a home server must be retarded enough to pay out the ass for our product. they'll literally pay it so they can ask if Barack Obama had sex with a money in june of 1989.
As part of their newly announced open source and transparency initiative, Deepseek reveals that their profit margins for their API R1 service ($0.5/MTok output) is... 545%. For comparison, Openrouter providers offer unmolested R1 at ~$7/MTok output. Imagine what the margins on expensive models like OAI's are
>a hedge fund destroys a non-profit by being honest about its pricing, leading everyone to question how "non-profit" the "non-profit is"

let's be honest if the nationalities were switched everyone would be hyping up Deepseek. but the facts are the break even point for OAI would be charging 1% what they're charging people and we all know they're using their revenue to fund tranny sex operations and democrats just like Sam Bankman-fried did. We all know the people at OAI are faggots and any honest person would be demanding to know where the fucking money is going. 100% OAI is basically private USAID,
 
  • Like
Reactions: Ghostse
It's funny how butthurt 4chan gooners are about Drago.
My gay Discord cult is better than your gay Discord cult etc.
Drago's problem is letting too many retards in and not implementing basic rate limits to feed his ego a bit less.
to not just use a modified deepseek on a home server
Most users and the overwhelming majority of revenue for these companies is corporate, and I'm not talking real evil gov-corporate contracts like Palantir, just run of the mill companies wanting the latest buzzword tool for their employees. It's far easier for a company to expense a subscription and SaaS licenses for their employees than justifying hardware to run models locally, even if it's dedicated plug-and-play stuff like Digits.
Plus, local models are just not good enough compared to API, there's no contest about that.

Also the distills are NOT Deepseek, they're Qwen2.5/llama3 with sprinkled in R1 outputs, they're good for what they are but Deepseek they ain't. I blame Ollama for this, they mislabel their models so every pinoy teenager that googles how to run R1 on their Samsung A20 sees their website first.
 
Plus, local models are just not good enough compared to API, there's no contest about that.
while true, what are they using AI for, every website has that "how can i help you" bot which i assume is AI but much like tech support it doesn't seem to labor intensive and the biggest cost must be being able to handle hundreds of assholes using the site at once and interacting with it.
 
Microsoft claim to have had insiders that ratted out on each other, but I don't really believe it, it was probably spitefags. They must have been monitoring 4chan for a while, since certain links were getting autofiltered/banned for a month now before they went public with it.
I love how they phrase all this shit like they're some kind of superheros protecting society instead of just a pack of thieving racketeers angry that someone else leeched the shit they originally stole from their customers. Choke on cocks M$.
 
Who the fuck uses a proxy for DALL-E anyway? SD is free
You underestimate how lazy and stupid people are, not only there are idiots paying for AI generated content but last time I was in civitai there were users talking how scammers were taking their generations and selling them, when the prompt its right there. Do you see? even the scammers are stupid now, they were taking time to manually mirror the images and cropping the watermarks one by one instead of just using the prompt to generate their own in a second.

This level of retardation should be illegal.
Hilarious how what bothered MS isn't the rampant honest to God fraud done against their third party customers which is an actual crime, or distasteful uses like the metric shitton of loli ERP generated with their system, but goddamn Taylor Swift's ugly mug
Because she's a billionaire and the most famous popstar right now, the bad press alone is not worth it.
so every pinoy teenager that googles how to run R1 on their Samsung A20 sees their website first
That's eerily specific...
 
Last edited:
We're getting closer to the matrix.

>be me, Thane of Glamis
>just won big battle against Norway
>king loves me now
>walking home with Banquo
>suddenly weird sisters appear in the fog
>"sup Macbeth, future king"
>wtf.jpg
>they tell Banquo his kids will be kings too
>they vanish before we can ask questions
>get message that I'm now Thane of Cawdor
>wait, that's what the witches predicted
>maybe I really will be king?
>tell wife about prophecy
>biggest mistake of my life
>she gets crazy excited
>"we're killing Duncan tonight"
>mfw the king is literally staying at our castle
>try to back out
>wife calls me a coward
>fine, I'll do it
>see floating dagger before my eyes
>probably should have taken that as a sign
>stab Duncan while he sleeps
>can't stop looking at my bloody hands
>wife has to finish the job
>blame it on the guards
>become king just like that
>can't sleep anymore
>start seeing Banquo's ghost at dinner
>hosting skills: 0/10
>paranoia intensifies
>hire hitmen to kill Banquo and his son
>son escapes
>wife starts sleepwalking and confessing
>witches tell me "no man of woman born" can kill me
>feel invincible.gif
>enemy army disguises as forest
>should have seen that coming
>learn opponent wasn't born naturally
>realize prophecies were technically true but actually lies
>get head chopped off
>mfw I trusted witches
 
witches tell me "no man of woman born" can kill me
>feel invincible.gif
>enemy army disguises as forest
>should have seen that coming
>learn opponent wasn't born naturally
>realize prophecies were technically true but actually lies
>get head chopped off
>mfw I trusted witches
i love how Tolken was so pissed about that ending that multiple parts of LOTR are about how he felt the prophecy should have been fulfilled. hence "i am no man" and the walking trees and that third thing.

Also how the fuck could an army disguise itself as a forest?
 
  • Like
Reactions: The Spoils of War
Back