File: 1694883790890.jpg (1.1 MB)

1.1 MB JPG

>ANCHOR

File: 1775255463438.png (1.1 MB)

1.1 MB PNG

>>Noteworthy AI News<<

>IMPORTANT FOR AMERIMUTTS<
AI safety website found that categorizes AI systems based on arbitrary safety scores (Its just a website gauging how close your state is to total GoodGoyID™ surveillance. Its been found the (((advisory))) board member is apart of "8 Degrees East", a ((("""boutique"""))) corporate solutions and intelligence partner based in Switzerland, specializing in strategic alliances, risk management, and market-entry strategies for businesses.)
>https://teenaegis.com/intelligence/ai-danger-index

Trump administration vows crackdown on Chinese firms 'exploiting' U.S. AI models (use opus cuckie)
>https://www.npr.org/2026/04/24/g-s1-118582/administration-crackdown-on-chinese-firms-exploiting-u-s-ai-models

Meta says it will cut 8,000 jobs as AI spending soars
>https://www.bbc.com/news/articles/crm1y89vek8o

China’s DeepSeek releases preview of long-awaited V4 model as AI race intensifies (can it compete against opus?)
>https://www.cnbc.com/2026/04/24/deepseek-v4-llm-preview-open-source-ai-competition-china.html

Grok tells researchers pretending to be delusional ‘drive an iron nail through the mirror while reciting Psalm 91 backwards’ (lol)
>https://www.theguardian.com/technology/2026/apr/24/musk-grok-x-ai-researchers-delusional-advice-inputs

deepseek has fallen

how is it jennyfag holds total dominion over this thread, damn

>>108675650
Probably because nobody else gives a shit about this general anymore.

>>108675657
That's because of the lying jewsucker constantly trolling and shitposting, and the mods not their jobs.

>>108675611
https://rentry.org/aicgOP
>https://rentry.org/aiccgOP
not my bake
OP has be compromised
Impressive you took some old images though

When will skillgod add kimi 2.6?

>>108675611
Hi I'm new. Who is this little robot character based on, or what do they represent? What is the lore?

File: nope.png (7.2 KB)

7.2 KB PNG

>>108675695
Yup
Multiple url shorteners in the molested OP
https://ghostarchive.org/archive/9DHkn

LMAOOO

It's funny when Deepseek reasons in Chinese but responds in English

>>108675763
funnily enough most polyglots still think in their native language

>>108675636
>>108675702
just use opus

is chinkseek multimodal?

>>108675703
Hi, you're not. But anyway: It's XJ-9 (aka Jenny Wakeman) from the Nickelodeon animated series My Life as a Teenage Robot. Bland calart stolen from Jetsons under the guise of future-retro. Every episode is the same. Jenny constantly whines about wanting to be a normal teen but she's a robot (how original!). Her mom says no, you have a duty. Jenny sneaks out. Shit goes wrong. Missiles deploy from her hair to defeat a Poison Ivy ripoff. Status quo restored. Roll credits. The series ended on an unresolved cliffhanger because nobody cared. It's an unfinished, abandoned, half-digested piece of animation roadkill. The show won a grand total of jack shit in major awards during its run. The 80's Smurfette had more substance and was more attractive than this calarts garbage. What's hilarious is that Nickelodeon would rather produce seventeen more seasons of SpongeBob spinoffs than spend five cents wrapping up Jenny's story. Jenny's a discount Astro Boy on a show that deserves to stay dead.

>>108675789
>just use o-
{{user}} thrusts his penis into {{char}}'s O-shaped mouth with enough force to shatter teeth and bones

Anyone try the roleplay mode for v4?

https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/README.md

>>108675801
HOLY TRUTHNVKE

Incoming hours long self bumping.
In all fields

>>108675814
t. poorfag coping that his inferior models can't compete against GODPVS

>>108675839
>GODPVS
that is the sound you made when my penis interrupted your mouth

say "january 5 2026" out loud
this kills the slopusfag

The new DeepSeek version is shit for lewds. What's the best model on OpenRouter for porn nowadays? Preferably one that won't balk at incest, loli, bestiality and other stuff.

>>108675801
>calarts
kek, it's you again

>>108675871
read >>108675789

>>108675775
Fake

Only whites use Opus

Tested v4 a bit (Pro, but Flash seems ok, didn't test a lot).
>>108675796
No, but they could extend it someday.

>>108675816
I saw this, it's basically something you put after the first message to tell it to think in-character, or refuse to think in character and be fully analytical.
The guy that posted that (unofficial) claim their dataset had some of that prompt in training so it might work, and that it should work even on the free API, no system prompt needed.
I haven't really tried it yet, from my initial tests, it does think in character 90% of the time already, so is this even needed? Might be useful if you want it to think autistically like old R1 did, which is fine too. Something to use to taste.
I'd be surprised if that prompt is really needed, but could be tried. Will it do well with it even in a translated version or does it have to be in Chinese?
I've once fucked up the turns (I use a custom client, not ST, and made a mistake and sent my message as it was from the assistant role) and it started thinking in Chinese (just once), but output was good anyway. I rerolled (swiped) the turn and it generated normally.

For RP: I tested on some loli yuri imouto prompt for now. Mostly seems fine, output is cute, not very slopped, just tiny bit (noticed maybe 1-2 words I consider "slop" in 15 turns so far). Clever enough. More creativity due to larger size compared to smaller models. Output length is shorter than some competitors, but got longer with more turns, you can of course instruct it to write more. It's enjoyable, but I lack the time to finish it now or try other "cards".
I will test something less cute/more dark another day to see if there's any positivity bias to it (R1 used to ace this well).

Continues

>>108676001

For code: I tested a tricky, but "known" math/physics/engineering/coding problem, but not an entirely novel one, compared to 3.2 it used a few extra clever tricks and optimizations (in both the math and the code), approaching what you'd get from a professionally written program to solve it (essentially a simplified physics simulation, but having enough features to be usable for real world engineering).

Recently I've tested a number of models with this task, models pre-o1/r1 were not capable of even approaching solving it, but nowadays even something local and tiny like Gemma4 can approach solving it, just do far worse than bigger models like DS3.2 which itself does worse than DS4. Both 3.2 and 4 solve it, but the quality for 4 is better, it had to think like 18k tokens to solve it all, which is reasonable (the problem's solution is long and not easy, most textbooks that try to do something similar often take at least 10+ pages for just the math without the code).

>>108675871
So, what's wrong with it? Be more specific. I find its writing fine, but I also find old R1, DS3.2 and so on okay too.
Even fucking Gemma4 is usable for ERP, but being small it is more prone to slop than big models, however it's not stupid so you can do 40 turns with it keeping track of the plot well.
It's just that you get used to the writing too soon.

>>108676007
It barely listens to my custom prompt, writing is emotionless and flat, posts are far shorter than they were before, it seems to have trouble controlling multiple characters at once (defaulting to speaking for one character per post instead of acting for every character involved in a scene), it struggles heavily to keep the story moving, and in general it is a big step down from the previous version. This is all using the pro version, which is like five times more expensive than what I had before, despite being worse in every conceivable metric.
>>108675905
I'm trying this now. Any good custom prompts for it?

Tested Claude. Twenty chat messages cost me nearly two dollars. Fuck that shit. That's way too expensive.

>>108676027
Is this on Flash or Pro? Set model to deepseek-v4-pro.
Writing was quite cute here and thinking was in character (didn't prompt for this). Replies started short, I could try to tell it to write longer though, they got longer by turn 8.
> it struggles heavily to keep the story moving
Pacing seemed slower, yes, R1 would try to satisfy my request really hard, such that I could finish the story even in 10-15 turns, maybe too fast. It does feel more mellow, but I think it'll be fine for slow burns. Maybe could try to prompt it to go faster. I will try other stuff tomorrow, no time today.

>write 50 messages with Deepseek
>one (1) caches
but why

What's the meta at the moment? Coming up on my three year anniversary of botting with you tards and haven't posted in here in months. What's skillgod's trycloudflare atm? Is he even up? Mystery never returned? What model is king?

>>108676001
>tested on some loli yuri imouto prompt for now
how much "is it safe for my safety safe rules" (outside of arbitrary ones you'd give) does it waste doing by default in the thinking?

>>108675650
because he forces threads early so he can force his retard uggobot on everyone he wont stop fixating on in an unhealthy manner and forcing it on everyone else like a retard autism tyrant. I wish a person would just go fuck rules and permaban this subhuman broken autism chuckle fuck who thinks what he wants and obsesses over is what everyone will.

he's the worst of the various tiers of autism

>>108676335
you lost btw tranny

>>108676303
I had some very minimal "jailbreak" in the system prompt that I use for many LLMs, just instructing it to be uncensored and write thing explicitly.
Never have gotten any refusals or "thinking" about safetycucking, but I never got refusals with DS3, R1 and maybe once I remember getting a refusals with one of the intermediate DS3s which seemed like something they picked up from gemini synthdata rather than legit "safety" shit.
I'll see if I get any refusals, but so far none, meanwhile I remember first Kimi2 would refuse every single turn for loli, but I could just reroll or use inline instruction to bypass it even when they didn't have prefills on the API. Meanwhile, I never needed anything but a simple instruction in the system prompt that worked on all DS models.

File: download - 2026-04-14T234823.890.jpg (71.2 KB)

71.2 KB JPG

>>108676340
uh oh another faggot retard runs to his cliche retort of tranny. next you'll lie and make up things and say schitzo and meds and touch grass and the famous claims im big mad when I called shit out. but keep seething bad autism is being called out because it triggers your subhuman malfucntiong autism brain like others. I await your many no u retorts and fabricated lies made up you throw my way or project out at me. it's 4chan thats what subhumans like you do.

bonus play you're stupid enough to bite the bait but do so not knowing it's bait. me see your bait and bite knowing it's bait. stay inferior every step of the way and know i'm mocking you and laughing at you and smoking weed and writing stupid chatbot stories while doing so. your turn subhuman.

>>108676355
thanks anon, I'm glad they didn't go the moralfag route
and yes I've used kimi before, by default it's obsessed with safety

>>108676361
>braaaaaaaaaap
you lost tranny

new deepseek seems alright

give me the news. do i need to rewrite my prompts *another* time just to sate ds4?

>>108676335
It's crazy how your actually obsessed over a namefag.
This is actual insanity, as nothing you've said was true. Is this what schizophrenia looks like??

>>108676070
We're you caching?

>>108675901
Ah, yes. Only one person hates calarts, according to blue-haired they/thems with a nose ring and a Funko Pop collection. Sorry, while uncle-raped HRT recipients may say otherwise, people have been screaming for the death of that style since Steven Universe's faggotry first aired and that shitty style became a meme. SU was the exact moment western animation decided genuine art was oppressive, and every cartoon needed to be a twenty-two-minute therapy session for adults who never grew out of their middle school DeviantArt phase. After that "let's hug it out" show with unironic fatasses dropped, every single CN pitch meeting became a contest of who could cram the most beanmouthed, chinless, noodle-armed characters into a show about diversity feelings. Craig of the Creek. Summer Camp Island. Infinity Train. All of it, variations of the same flaccid house style. The industry didn't evolve, it consolidated into one gigantic beige blob of performative inclusivity and zero craft. But please, keep praising an already-dead style for permanently poisoning western animation.

before D4 drop:
>Chinkseeeeeeek foooour!!!
>airhorns
now it's here:
>crickets
Why? genuine question

>>108676762
wrong thread retard

>>108676783
No I'm pretty sure I'm in the right thread faggot

>>108676762
Kek, it do be like that

>>108676027
I think it's better than 3.2 but otherwise you're right.
>>108676076
I had the opposite where it would sprint through the fucking sex scene because it wasn't listening to my system prompt. I even slid it down to post history and it still barely listens. Might be another case of 'use a 20 token prompt and hope for the best.'

>>108676762
People are too busy using it, retard. The quieter the thread is, the better the model.

File: 1768630970837254.png (70.9 KB)

70.9 KB PNG

Knowledge seems pretty narrow so far compared to opus and gemmy. Otherwise it's pretty good in terms of creative writing. Also more prone to context poisoning which is a step in the right direction (or maybe other models are just too assistantslopped idlk).

>>108677904
DeepCope

>>108675650
I prefer his stupid antics over the pedotroon bakes

>>108676070
>two dollars is too much
Are you from fucking Bangladesh or what

>>108676361
Meltie

File: file.png (53.8 KB)

53.8 KB PNG

my codingboy has been slopping it up

File: file.png (4.1 KB)

4.1 KB PNG

I'm almost out of credits on OpenRouter, what do?

I've mostly been using models like GLM 4.7 and today DeepSeek V4, basically cheaper models but still ok output

Is there anything more efficient than OpenRouter to enjoy these things?
I read people talk about nano gpt subscription but I don't know how it works

>>108680104
>what do?
Get another dollar? Do you live in fucking Rwanda for one dollar to be too much?

>>108680421
>he doesn't know

>>108680442
Britbro?

I've been lazily using ollama for a while to run local models, and I just installed piper TTS to try and verbalize the output, and I can't for the life of me figure out how to get it to work.

>>108679007
Two dollars for twenty messages is way too much, yes. I run through three hundred messages per day on average. Are you from India yourself, marketer-chan?

>>108680104
Nanogpt is good if you are just going to goon to it, don't try it for agent stuff, fails way too much and it is too slow.
But really, if you are just gooning put 10 bucks in open router and that should last you a while

>>108680695
Unless you use Opus, in which case you'll run out of money before you even get to the sex. Fuck Opus.