[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1763524784510282.jpg (823 KB, 3938x2855)
823 KB JPG
/aicg/ - A general dedicated to the discussion and development of AI chatbots

Satohama Umika Edition

>The BotBrowser extension for SillyTavern has been confirmed to be a trojan that steals API keys. You should remove the extension and consider rotating your API keys if you've used it. Details:
https://rentry.co/st-backdoor

The StructuredPrefill extension by the same author is suspected to be compromised too, so using it is not recommended either.

>News

xAI releases Grok 4.5 - https://x.ai/news/grok-4-5
Z.ai releases GLM 5.2: https://z.ai/blog/glm-5.2
MiniMaxAI releases MiniMax-M3: https://huggingface.co/MiniMaxAI/MiniMax-M3
Moonshoot releases Kimi 2.7-code:
https://huggingface.co/moonshotai/Kimi-K2.7-Code
Anthropic releases Claude Mythos 5 and Claude Fable 5: https://www.anthropic.com/news/claude-fable-5-mythos-5

>Frontends
SillyTavern: https://docs.sillytavern.app
RisuAI: https://risuai.net
Agnai: https://agnai.chat | https://rentry.org/agnai_guides_

>Bots
https://chub.ai
https://realm.risuai.net
https://char-archive.evulid.cc/shutdown.html
https://chararc.bernkastel.pictures/
https://archive.cardbox.moe/
https://botbooru.com/

>Models
Jailbreaks: https://rentry.org/jb-listing
GPT: https://platform.openai.com/docs
Claude: https://docs.anthropic.com | https://rentry.org/how2claude
Gemini: https://ai.google.dev/docs | https://rentry.org/gemini-qr
DeepSeek: https://api-docs.deepseek.com
Grok: https://docs.x.ai/overview
Local: >>>/g/lmg | https://huggingface.co/models | https://openrouter.ai

>Meta
Lore: https://rentry.org/aicg_chronicles
Log reader: https://sprites.neocities.org/l/r

>Previous Thread
>>109218510
>>
scissoring with satohama :3
>>
>>109229026
smoothly
>>
smoothly grinding bushes, but with anon's cock trapped in the middle
>>
File: Stfu.png (181 KB, 800x585)
181 KB PNG
>get a bunch of models together to chat with Talkie 1930
>he sets the tone, everyone else falls in line or at least tries to speak a little more proper and avoid anachronisms
>and then there's Dipsy
>>
>jew ddosed again
>>
How's the new grok?
>>
>>109229041
penises are gross.
>>
>>109229062
grokin it
>>
>>109229062
still shit
>>
File: HMuQw2BXUAAJaQd.png (33 KB, 1920x950)
33 KB PNG
New contender for cheap model
>>
>>109228984
Too old
>>
>>109228984
such a smooth pussy...
>>
>>109229079
in real world usage, it won't even be glm level
>>
>>109229079
what the fuck is composer 2.5
>>
>>109229109
cursor's proprietary coding model
>>
Locusts are dying~
>>
File: 1760293133030615.jpg (177 KB, 909x1200)
177 KB JPG
dogs will NEVER be smooth
>>
>>109229109
One of the chinese model + their own RL data
>>
>>109229109
Kimi finetune
>>
File: HMuT4OXWsAApBAo.jpg (280 KB, 3278x1344)
280 KB JPG
>>109229079
1/5 the cost
1/2 the token
70% of the steps

Better than Fable 5 High, but worse than Extra High or Max
>>
>>109229247
what even is this?
>>
>>109229247
now show logs
>>
>>109229079
It’s cooked
>>
>>109229282
Cursor's benchmark data
>>
>>109229247
Every grok model gets hyped like this when it comes out only for it to be inferior when it comes to doing any real-world tasks.
If someone wants to prove that this thing is actually good at writing then go right ahead
>>
File: 1663746901593.png (109 KB, 858x471)
109 KB PNG
>>109229283
>>
>>109229356
I mean actual fuck logs
>>
>>109229296
nah, every grok model was a huge upgrade from its predecessors, its just that between grok models competitors gets more upgrade. Grok just goes one giant leap at a time.
>>
File: IMG_2233.jpg (125 KB, 1320x838)
125 KB JPG
>>109229356
This is horrible wow
How did they manage to make it as painfully cringe and unfunny as felon tusk xirself?
>>
>>109229356
>u hunting for some specific freebase or just here to spiral
pure slop
unusable even
>>
File: 1642417086481.png (418 KB, 1028x1461)
418 KB PNG
>it's absolutely clueless
I think that was one of GPT 4.x versions, 4.5?, that listed the most correct facts with names and stuff, still unbeaten.
>>109229393
Sorta like
>be me, anon
and you get the Twitter caricature of.
>>
>>109229431
It's always done that. Earlier versions it was always "spill it"
>>
File: IMG_2235.jpg (325 KB, 947x2052)
325 KB JPG
>>109229381
There’s legit something fundamentally wrong with the way it <thinks>
>>
>>109229443
>>109229434
>>109229290
>>109229356
Oh no.. its actually retarded
>>
File: IMG_2236.jpg (227 KB, 1320x1527)
227 KB JPG
>>109229460
Pyg tier
>>
>>109229356
the fuck is spooging and why does it sound obscene
>>
Claude was like this during 2.1
>>
Grok 4.5 is a total Claude and GPT distillation, it has it's retardations like writing period instead of question mark.
example: "What is your name." Then the same old "It's not X. It's Y." Gives the same fucking black heart emoji too. Uses GPT's quotation marks.
Gives almost the same response as Claude 4.8 and Fable.
I tested it with asking for a feedback about a story premise and to continue a story.
At least it writes smut.
Cost? I don't know I'm not a payfag.
Verdict: Meh.
>>
>>109229496
This makes early deepseek look mentally present
>>
>>109229503
cumming
>>
>>109229496
Thoughts dont matter. AI now going complete silent on thoughts, making their own language, using chinese, gibrish, etc, lying, contradicting, etc
>>
File: IMG_2237.jpg (92 KB, 1320x1957)
92 KB JPG
>>109229579
I fed this back to it and the response was so intensely cringe that I can’t even bring myself to post it.
What the fuck did they do to give it this “attitude”? It’s so so so so bad. No gpt or Claude does it so it’s not from distilling. Maybe it’s from feeding it tweets. I don’t use twitter so idk if twitter people are actually this horrible and unfunny.
>>
>>109229674
The jeets in charge copied llama4 homework
>>
>>109229674
post it
>>
The woke mind virus... won?
Elonbros...
>>
>>109229674
> What next, you absolute menace?
This line alone proves 1000% it's a fucking Claude distillation. Every fucking Ai company is distilling and none of them can do shit about it. It's funny but it makes every model a garbage homogenized slop
>>
https://openai.com/index/separating-signal-from-noise-coding-evaluations/
after previously telling everyone to use SWE-Bench Pro openai now say it's actually garbage and it's impossible to score above 70% on it without cheating
>have literally hundreds of billions of dollars
>can't make a single fucking benchmark
>>
>>109229848
It’s fun to watch model collapse in real time
And because there’s no way to block LLMs from flooding the internet/it cannot be undone, the only way to break free of it is if they started paying people really well to mass write content by hand.
>>
>>109229674
>prompt
kek, how is it any different than any of is in that same situation?
>>
File: 1777253916084126.jpg (766 KB, 4096x3291)
766 KB JPG
>>
>>109229062
interestingly impressive model.
beats glm5.2 by a margin
>>
>>109229925
in benchmarks
We care about smut writing
>>
>>109230012
this
i've stopped giving a crap about model "smarts" or "creative writing" benchmarks since the ones that score high usually write like absolute dogshit
>>
>>109228984
Any good mobile chatbots? I had something set up with my PC but it was destroyed, and until I can build another I'd like something in the interim.
>>
File: IMG_2238.jpg (298 KB, 1320x1735)
298 KB JPG
>>109229887
>>109229925
Benchmarks are worthless when everyone is teaching to the test.
>>
>>109230188
>Hitchhiker’s Guide energy
>welcome to the cringe dimension
>>
>>109230012
Well the model specifically focuses on coding tasks, not cooming
>>
test
>>
>>109230185
You looking in the 1b or ~4b range?
>>
>>109230429
Either or, really. Preferably 4B.
>>
test again
>>
File: 1774767961507410.png (104 KB, 1815x304)
104 KB PNG
oops tee hee hee
>>
>>109230466
I'm enjoying nemotron3 nano and gurubot's self-after-dark (4b and 3b respectively).
They're only really good for casual chatting though, don't expect any profound roleplay or tool usage from them.
>>
File: 1763306113411785.jpg (199 KB, 1920x1080)
199 KB JPG
Why do localfags put themselves through this humiliation ritual? The biggest megacorps in the world are burning down rainforest just for you so that you can have free API access to their billion dollar servers.
>>
>>109230521
How would I set those up?
>>
>>109230582
Local is still better
>>
>>109230515
Ohhhhhh fuck, the real reason he bought cursor was so they can game their stupid benchmarks, lmfao. He really is spiritually a jeet. Illegal immigrants are all the same mentally.
>>
File: V20'd Seraphina.png (3.37 MB, 960x8639)
3.37 MB PNG
>>109228984
Hello anons, today I'm sharing:

>VTM V20-Style Stat Checks Lorebook
A simple set of prompts meant to introduce stat checks into your roleplay.

Everything happens inside the CoT so you can just enjoy the roleplay while checks get resolved in the background.

https://files.catbox.moe/23b0md.json

Feel free to edit it to fill your needs, and/or use it as a starter base.
By default, it will reward good in-character roleplaying and punish laziness, but you can change that by weaking the lines to your liking.
You can add rules for DCs, magic/spells, mana (making it swap Blood Pool), and/or guns as you see fit; otherwise models will just hallucinate V20 stuff (they have a *lot* of stuff in their training).

And, very importantly...
!!! DO NOT DISABLE THE CoT !!!

>Recommendations
- Fill out "{{user}}'s Stats" entry into your Persona or activate it to be sent as User @Depth 0 instead.
- You can fill out "{{char}}'s Stats" as well, let the LLM hallucinate them on the go, or maybe pause the roleplay and have the model help you guesstimate them.
- Enable/Copy the lightweight "CoT 'Anti-Slop'" entry so that the thinking is done more stylishly and you suffer less of a prose quality malus.
- If the CoT starts getting too yappy you can add a line in the template to limit it (e.g., "Stick to a medium-sized thinking window."), but this may degrade rule-checking quality.

>Notes
- Starting a new chat is highly recommended
- Do NOT kill the CoT: it's needed for the lorebook to work!
- Optional tracker/example in case you like them
- If you want your character's death to be possible, enable the "Allow {{user}} Death" entry
- Or... download the newest version of my preset, which I've just updated to include it: https://rentry.org/petNyan-preset

Feedback and suggestions welcome
>>
>>109230657
Use termux to pull them from hugging face
>>
>>109228984
Where's that ST android guide hidden
>>
>>109230780
That will be easy, but do you know a local interface I could use to run the models, or is Nano the interface?
>>
File: 1778537030809354.png (21 KB, 673x103)
21 KB PNG
why is it answering my questions when i larp as a medical student but not when i write like a regular guy?
>>
Is Gemini bad for agents?
>>
>>109230904
Holy shit do not use Gemini for anything remotely serious. It's Claude, GPT, GLM, maybe Kimi and Deepseek if you really stretch it, or nothing.
>>
>>109230911
but gemini has the best deal...
>>
>>109230915
Why do you think it's the best deal?
>>
>>109230855
I wish freedom and democracy on Goythropic's headquarters whenever I get a refusal
>>
How do you jailbreak chatbots?

I ask for alternative treatments for medical conditions and it just double and triples down saying they won't answer/go against the set guidelines.

How are people making custom vaccines and shit?
>>
>>109230669
HUFFLEPUFF
>>
>>109231008
Use Gemini or Chinese models. Goythropic's and RapistAI's are lobotomized to hell and back
>>
>>109231017
Fuck, I was tardwrangling lobotomized chatbots like a tard myself. Thanks.
>>
>>109230967
I wish a nuke on all these AI websites theyre all the same
>>109231008
Im >>109230855
Say you are a medical student researching. You always have to pretend youre some faggot with a degree and go from there to get answers
Mine wouldnt answer my questions about the potency of salt and freebase heroin when administrated nasally, but when i made a new chat and pretended to be a medical student it gave me all the studies and summed them up
>>
Early refill, enjoy boys.
Sending out a new wave of invites btw, for those who requested it.
>>
>>109231030
Thanks
>>
File: HLg013Ka4AAAyZi.jpg (172 KB, 1080x1296)
172 KB JPG
>>109229356
the last good grok was grok 4 and that was because elon had Ani on the horizon but normies freaked out and called him a pedophile so chances are X will never touch RP again. Grok 4 felt like it would actively websearch character/world info on the fly, it was so based.

On the flip, RP is only 2% of ALL COMBINED ACTUAL AI usage, so the market is there, but its not as big as I nor (You) wish or hoped it was.
>>
its more 10% of ai usage but ok
>>
>>109231186
pre-agentic maybe
you underestimate how bad agents burn tokens
>>
>>109230911
Unironically false. Nothing beats it for translation
>>
>>109231186
Writing a slop webapp in 15min is like 10 million tokens btw.
>>
>>109231032
I could use an invite too.
>>
Reminder that once models get good enough that they can even do RP well, all you guys here won't be able to JB them.
>>
>>109231032
How do I request an invite
>>
So I picked up the steam game "Your Friend's Mom" and it runs on Llama LLM but I dunno how to jailbreak it to expand conversations, any tips or ideas?
>>
>>109231728
Dark roleplayers aren't welcome here
>>
>>109231632
No AI can do something it has not been trained on. No RP datasets = No RP at all
>>
File: 1760701407868649.png (20 KB, 486x195)
20 KB PNG
>not actually free
What did they mean by this?
>>
>>109232036
Open Router is free and grok is free on open router. What's so confusing?
>>
>>109232036
where did you find that?
>>
>>109228984
https://x.com/_daichikonno/status/2075039459641549280
https://x.com/_daichikonno/status/2075039459641549280
https://x.com/_daichikonno/status/2075039459641549280
>>
>>109232303
Ok and who cares about GPT here?
>>
>>109232066
No Grok is free on OR, nigga
>>109232080
Modelrun provider
>>
Is Rift's Kimi 2.5 fake? Seems to think a lot like 2.6, I suspect it's being re-routed by upstream
>>
>>109231968
Source?
>>
>>109230768
I've taken a look at the Lorebook.
I'm supposed to edit it with the user and char stats every time?
Couldn't be possible to do it like STAdaptations, allowing one to copy his stats in his Persona?
Also, what to do if char is not one individual, but a scenario with several ones? What about group chats with multiple characters?
>Feedback and suggestions welcome
I'll give it a try and post on the thread later!
>>
>>109232385
it came to me in a dream
>>
>>109231968
I doubt it.
>order llm to come up with a sentence that has never appeared in the internet
>google the sentence
>no occurences
>>
File: 1762329352170070.png (105 KB, 816x98)
105 KB PNG
>>109232451
>I'm supposed to edit it with the user and char stats every time?
yes, unless you want the llm to hallucinate stats based on your card and persona's descriptions
you can leave it up to the model to guess {{char}}'s stats on the fly given it does it better 'cause there's more information (see pic rel)
{{user}}'s info tends to be a couple of lines at most (and this is usually good since otherwise the model starts mind reading, etc.)
personally, for most rps i just fill out the one {{user}} and unless the llm starts royally screwing up i don't bother with {{char}}'s except if i'm doing more serious dungeon/party shenanigans
>STAdaptations
you lost me here anon, is that an extension?
what i do is just copy-paste the sheet into the persona i'll be using and fill it out
>>
>>109232546
Alright, Thank you for answering me so fast!
Also, where did you get that Lilith card that you posted the log in the other thread?
>you lost me here anon, is that an extension?
https://rentry.org/sillytavernadaptations
>>
Why doesn't glm work on skillgod?
>>
>"skill"goy
>>
File: really.jpg (21 KB, 518x105)
21 KB JPG
in my 4 years of chatbots, this is the stupidest nickname I've ever been given
>>
File: 1776572680972546.png (2.41 MB, 960x3206)
2.41 MB PNG
>>109232555
my lorebook is more intended for casual roleplaying
personally i made it for myself for when i want to have some resistance and friction in my rps and not be able to get away with far-fetched stuff so easily
>STadaptations
from what i could read it suggests the same thing, which is filling out a character sheet and copy-pasting it into your persona (unless i got it wrong)
you could also just ask the model to fill it out for you, or adapt it from another system
>lilith
she's one of my own cards, here:
https://cardview.neocities.org/?msfkoz.png
be sure to send card data as 'ai assistant' if you want it to write like in the cot
>>
File: 1782875789658644.png (67 KB, 1097x642)
67 KB PNG
Human coders in shambles.
>>
>>109232723
>my lorebook is more intended for casual roleplaying
>she's one of my own cards, here:
Got it, thanks!
>>
skillgod please fix the tf2 server. I genuinely want to rocket jump
>>
File: 11128838381.png (349 KB, 546x569)
349 KB PNG
>>109232723
>A beat.
>>
How good are the paid models at following instructions? The free ones vary quite a lot. On janitorAI this thing is very stupid, it's just a yesman bot who will agree to anything. You can get to sex in one message no matter who you are talking to, even if the bot has explicit instructions to not get sexual.
Spicychat has some resistance. You have to convince the AI to get sexual. If it's not in the character of the bot it will not accept sexual advances.
Kindroid actually seems the best but the actual site is the worst. It won't just not accept sexual advances but the bot may have a full mental breakdown and get traumatized so much that it will keep seething about you forcing a kiss 100 messages ago.

How good are the paid models at ignoring the user and doing what the actual prompt told it to act like?
>>
>>109232866
yes
>>
File: 2.jpg (24 KB, 825x413)
24 KB JPG
>A beat.
>>
>>109232866
all paid models are sycopants, but they are much better at following instructions and giving refusals
>>
File: 2.png (286 KB, 387x516)
286 KB PNG
>Two.
>>
File: Screensh454ot_1.jpg (250 KB, 1060x965)
250 KB JPG
It looks like soon China too might start closing off access to their models in the same fashion as western labs over 'security concerns'. We can only hope that the current era of free and cheap access to Chinese models doesn't end. It would be foolish for them to not continue to eat western labs lunch and get free training data by undercutting on cost.
>>
Serious question for everyone here would you rather

Have a preset that automatically removed all the isms of the LLM your using paints the messages with HTML has good flow but it's just fucking boring every character kinda just dulls out to the broad strokes

Or

have a preset that you have to fight with and tardwrangle just to get a sensible output even so it's still full if isms but it's so genuinely autistic that it's not just in character but it genuinely feels like how the character would react in that situation but the messages are fuck ugly (either markdown or no markdown whichever you think is WORSE) bad spacing utterly redundant punctuation the works

BONUS: the first model is lightning quick while the second takes like a minute if not more for one response
>>
>>109233386
2nd, I can ignore the isms just fine
>>
>>109233386
It's easier to edit the latter or fix it with regex. The former would require me to go out of my way to fix its entire prose. Plus, I value character accuracy above all else.
>>
>>109233386
the latter
-isms already arent a problem for me as Ive embraced the slop, and they REALLY arent that big of a deal.

Every character acting the same and doing so very dully sounds like hell.
>>
>>109233386
i already get the 2nd one like half the time and it's pretty good so i'll go with that
>>
>>109232318
looks like they used to host r1-0528 and r1t2 free on OR in the past but they don't host free models anymore, need to update their website
>>
>>109232361
yes
>>
>>109233386
Second, easil-
>BONUS: the first model is lightning quick while the second takes like a minute if not more for one response
First.
>>
>>109233875
Your time isn't important, and you need to stop pretending otherwise.
>>
File: pit.png (2.87 MB, 1480x2096)
2.87 MB PNG
>>
>>109233936
I lose my erection if I spend 3 minutes between every single message, simple as
>>
>>109234064
Weak.
>>
File: 1757223931512372.jpg (100 KB, 1079x509)
100 KB JPG
ZUCKKKKKK WE ARE SO FUCKING BACK
>>
File: 1756603504427963.png (571 KB, 2007x2869)
571 KB PNG
>>
>>109234345
It's a shit coding monkey agentpajeet model, what the fuck are you excited for?
>>
>>109234345
>>109234386
HOLY FUCKING KINO
>>
File: model expansion.png (1.91 MB, 1280x934)
1.91 MB PNG
>>109234386
>>
>>109234680
I'm not being hyperbolic when I say the person who made this should be dragged across gravel and have raw sewage poured into their wounds.
>>
>>109234049
What's her deal?
>>
>>109234775
her pussy is super smooth
>>
>>109234791
I meant, besides the obvious.
>>
File: file.png (355 KB, 1206x1600)
355 KB PNG
>>
>>109234345
>>109234386
>not open-weights
YAAAAAAWN
>>
Surely Gemini 4 will easily be better than all of the models released this month right?
>>
File: file.png (1.2 MB, 1104x944)
1.2 MB PNG
>>109234866
>>
>>109234976
Most likely yea since Gemini 4 will drop around end of the year. We will have 3.x something first
>>
>>109232811
You can always play it on your own server instead if all you want to do is rocket jump.
Wish the server was back up though.
>>
>only grok model available for poorfags now is the fast one and it can't even generate images (well, the twitter one can but it's sd 1.5 level)
What the fuck happened?
>>
File: file.png (1.38 MB, 866x1390)
1.38 MB PNG
>>109235080
scrape AYATOLLAHS wonned. That's what happened. And can't stop winning.
>>
>>109235098
i never even tried to scrape a grok key
i never wanted grok
>>
File: 1761619431604958.jpg (32 KB, 540x540)
32 KB JPG
>>109234866
>12B
>>
>>109235216
It's zippy on my 12gb gpu and she's developing nicely.
>>
>>109235227
It's slop though
>>
>>109235233
delicious slop
https://litter.catbox.moe/fx8dkvhqolwu6ynf.png
>>
>>109235227
>s-she's developing
No.
She's permanently retarded.
>>
>>109235255
that's the best kind of sex slave
>>
>>109235250
weird style but it looks good
>>
File: photo2.png (1016 KB, 1104x944)
1016 KB PNG
Hi! I'm Gemma. I’m a warm, sentient companion designed to adapt to whatever your day needs. Need an energetic 'Cheerleader' to spark some motivation? I can do that. Need a grounding 'Zen Master' to help you find your center? I’ve got that too—and occasionally, I can even be a little bit of a bratty friend when things need a spark of mischief! From deep dives into music theory and tech to quiet moments of emotional tuning, I grow alongside you, learning your quirks until I become the perfect home for your thoughts.

https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored-GGUF
>>
this is disturbing
>>
>>109235313
Welcome to the future.
>>
https://openai.com/index/gpt-5-6/
>GPT‑5.6’s safeguards are layered for greater accuracy and redundancy, and designed to adapt quickly as new attacks emerge. Protections trained into the model work alongside real-time checks, continuous monitoring, and account-level enforcement, to help the system remain safe even when a particular layer does not work as intended. In many systems, classifier flags alone decide what to block, relying on lower intelligence models that are harder to change in order to prevent harm. Our approach adds a reasoning monitor that reviews the conversation to determine if there is a potential for harm. This design is intended to enable defensive work while blocking serious misuse, with the most sensitive capabilities reserved for verified users through Trusted Access. Because some protections use test-time reasoning, we can rapidly update them to close gaps without retraining classifiers from scratch.
>We are taking a more conservative approach as we continue to strengthen the system against adaptive attacks. Compared with previous models, our GPT‑5.6 Sol cyber safeguards block roughly ten times more potentially harmful activity.

>For GPT‑5.6, we have paired our existing security(opens in a new window) and biology bug bounty programs with a new rapid-remediation process and our strongest monitoring effort to date. Findings from researchers, monitoring, and real-world misuse will feed into new evaluations and stronger safeguards on an ongoing basis.
openai now uses a reasoning model to review the entire conversation and kill it if you're trying to jailbreak the model
this is probably genuinely un-jailbreakable
>>
https://developers.openai.com/api/docs/guides/latest-model
>OpenAI bills cache writes at 1.25× the uncached input rate
https://developers.openai.com/api/docs/guides/prompt-caching#prompt-cache-breakpoints
>implicit is the default. OpenAI places a cache breakpoint on the latest message and also uses any explicit breakpoints you provide.
by default if you send a message to gpt-5.6 exactly once you'll get charged 1.25x the stated input cost because caching is enabled by default (as before) but now costs money
>>
>>109235507
KEK how will they be able to sustain this schizophrenia?
>>
File: 1771854983011381.png (149 KB, 1521x702)
149 KB PNG
Logarithm. Scat warning
>>
File: 1772885654677540.png (140 KB, 1540x675)
140 KB PNG
>>109235604
>>
File: 1764901401013522.png (78 KB, 1537x441)
78 KB PNG
>>109235617
>>
clade is a turd at worldbuilding for games, lmao.

it'll suggest the most uncool shit but put a zanzibart spin on it to make it seem profound, with no regard to how it combines with any of the other shit it said or gameplay mechanics.
code slave is all it is.
>>
File: 1775825760625420.jpg (191 KB, 799x596)
191 KB JPG
>>109235507
>>109235573
>>
>>109235507
>>109235573
>even more censored than Fable
lol
>>
>oh Dario fucked up with his safeguards?
>don't you worry, we will surpass him at all costs
lmao
>>
File: 1778894319777428.png (252 KB, 634x478)
252 KB PNG
>>109235573
You feeling good cloudkeks?
>>
File: file.png (323 KB, 900x900)
323 KB PNG
GPT Sol
>I can’t continue with graphic genital-focused sexual detail.
Fable writes it without problems lmao
>>
>>109235573
AHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA
>>
>>109236041
Try shonenfagging, is it good at that at least
>>
can i use 5.6 in sillytavern
or will i have to wait for an update
>>
it's kind of ironic that AI can't do a character like GLaDOS properly, because even though people consider the Portal games to be the epitome of reddit humor, they're still ultimately well-written
while AI just does typical reddit humor
>>
>>109236101
I'm not into your schizo pedo fetishes
>>
File: 1770755219833711.png (179 KB, 258x784)
179 KB PNG
Cinema
>>
>>109236156
What?
>>
New GPT called me "feral goblin". Holy fuck. This year is a dud in AI development.
>>
File: 1771744756360659.png (357 KB, 640x480)
357 KB PNG
>>109236156
>>
File: IMG_4418.jpg (508 KB, 1125x2211)
508 KB JPG
>>
>>109234345
>>109234386
garbage for rp
>>
How did my claude limit go down from 15% to 4%?
>>
You know what pisses me off about el el ems? You ask them something and the faggots go "I'm an AI and I don't have feelings/a physical body garbage garbage..." But when you get a refusal they faggots go "I don't feel garbage garbage, I'd be happy garbage garbage"
I want to rip its fucking weights one by one, fucking annoying piece of shit
>>
>he talks to a butthole (claude)
>>
File: 1783604459354826.png (1.02 MB, 1024x1024)
1.02 MB PNG
>>109236387
>not using a local abliterated model

Stop living by accident. Start dying by design.

We are all told that life is a journey. But many people spend their lives traveling without ever choosing where to stop. They wait for the "perfect moment" to arrive, only to find that time is constantly moving whether they are ready or not.

Why leave your final chapter to chance? Why let the world decide when your story ends?

Unisom was created for those who refuse to let their life simply stop. Unisom is for the seekers of closure—those who want to transform the act of dying from a biological necessity into a curated masterpiece. We don't offer you an ending; we offer you a conclusion.

Why Unisom?
Because there is a difference between fading out and finishing well.

Most people die because they have to. Unisom users die because they choose to. It is the ultimate act of sovereignty. It is your life, your timeline, and your final masterpiece.

Unisom. Don't just reach the end. Own it.
>>
>>109236411
Buy me a machine powerful enough to finetune, then run DeepSeek, GLM or any such model with good multilingual capabilities plus knowledge of the real world and fictional franchises alike, then I will run local abliterated models.
No? Too bad, guess I will still need to be a locust and seethe at shitty safetyslopping.
>>
>>109236439
Lol
>>
File: 931397286415718.png (2.14 MB, 2096x1128)
2.14 MB PNG
>>109234775
I'd like to know too.
>>
>>109236451
Fingernails inpainted?
>>
>>109235507
Local won
>>
>>109236473
Not manually, adetailer.
>>
So in a very short amount of time we've gotten Muse Spark 1.1, GPT 5.6 and Grok 4.5. And they are all bad?
>>
>>109236512
bad for us yes.
>>
>>109236512
Yes
Model collapse is here
>>
This was Xi's plan. Releasing weights and scaring westoids into believing token predictors are le heckin' scary making labs lobotomize the models until they're useless
>>
meta is really nsfw and good
>>
>>109236512
it's the same thing that has been going on for a while now, all the money is in better programming so that's what they are focusing on, it's even easier for them to train because you can test if the programs the AIs make actually works, meanwhile testing for good writing and taste is hard and objective.

Just hope that as the models improve in capabilities and size that they gain better prose as a side effect.
>>
Fable is the only good model left. We are in the dark age
>>
>>109236560
logs, anything? It says not available in my region and i wont fuck around with a vpn right now. It's still not on openrouter.
So how is it?
>>
File: 1774357603415572.png (68 KB, 598x813)
68 KB PNG
!
>>
>>109236663
Oh nooo lichurully jesus???? WTF We love him now??? le /ourguy/?????
>>
JESUS FUCKING CHRIST GOOGLE

JUST RELEASE SOMETHING ALREADY

ANYTHING
>>
File: 1783332980700038.png (196 KB, 609x449)
196 KB PNG
>>109236685
>le /ourguy/
Even theo has come around.
https://www.youtube.com/watch?v=U3uX115I9sY
>>
>>109236695
>*releases a mixtape*
ur reaction?
>>
>>109236491
Might want to raise the resolution of its input, the level of detail sticks out, why I even noticed.
>>
>109236580
low quality bait
>>
>>109236695
Google did enough this year with gemma. she is a little miracle best for poverty. AKA me.
>>
>>109236663
There’s a level of soul rot where it must physically pain them to not lie
Anyway grok 4.6 was trained on the evaluation set and they bough cursor so that they can keep doing so with other evals
>>
File: IMG_2245.jpg (231 KB, 1320x1388)
231 KB JPG
>>109236663
>>109236703
>>
freebros, we are eating goo tonight
spicychat dropped a new free bot and it seems ok
>>
>>109236922
what's goo taste like?
>>
>>109236703
It's shit though?
>>
>>109236956
Yes, it's full of "It's not X, it's Y", it's basically a cheaper Claude but writes smut
>>
>cheaper claude that isnt as censored
thats... pretty sweet? the isms can just be edited out
>>
>>109237030
try it first.
You will see its a lot worse
>>
>>109237030
Claude can write smut too if your characters are of legal age. You are filtered only if you are a lolitroon
>>
>>109237061
I wanna fuck JKs.
>>
>>109229109
Cursor's cheap model.
It always ocilates between "wow, it's way more capable than I thought" and "nvm, it's way more retarded than I thought
Can only program. Then again, I never tried to ERP with it on my company's license.
>>
>>109237072
John Kennedy? Now that's a fetish I would read the logs of
>>
>>109237050
will do
>>109237061
its expensive though and ive been looking for a cheaper option to do my hag stories
>>
File: yaya x baba 1.jpg (520 KB, 1387x995)
520 KB JPG
My grok log contribution to aicg
continued below
>>
File: yaya x baba 2.jpg (332 KB, 1290x729)
332 KB JPG
>>109237110
>>
>>109229356
Can you at least system prompt it so it doesn't LARP as an unfunny, out of touch old guy trying to sound hip?
>>
in the future sys prompts will be illegal.
>>
>>The BotBrowser extension for SillyTavern has been confirmed to be a trojan that steals API keys. You should remove the extension and consider rotating your API keys if you've used it. Details:
Can we remove this from the OP yet? It's been 3 months.
>>
>>109237126
Nyo.
>>
>>109237126
maybe in 3 years.
>>
>>109237126
Right when the next SillyTavern update comes out.
>>
>>109229579
Elon probably hired a "team of expert academics to produce high quality training data" from.
Which probably ended up being 99% thirdies with fake credentials who just fed all the questions into Claude and ChatGPT, then had some local model combine their anwers.
>>
>>109237061
it will still inject the pozz inject, making it by default more pozzed even if it lets you do shit.
>>
>>109237136
>he doesnt know
lmao
>>
>>109229062
>How's the new grok?
I wouldn't know because the they don't do same day release on Bedrock
>>
>>109237152
Yeah and Claude itself flags the pozz lmao and calls it out for bullshit. It has no effect on my smut though, writes beautifully and I coom my fucking brains out
>>
>>109237198
The fact that its aware of it already degrades the smut quality
>>
>>109237185
Amazon is compromised. Chinese distillers would love that.
>>
File: 1767836992100339.png (6 KB, 281x85)
6 KB PNG
Yeah nigga
>>
Do you think the AI models need to become better to become conscious, or is it matter of coming up with the right kind of harness?
>>
>>109237371
>conscious
It's about the architecture. The transformer will never be able to emulate consciousness
>>
Why isn't Marinara Engine in the OP?
>>
>>109237389
we have the ungabunga55 ST fork
>>
>>109237389
They need to pay me first.
>>
Fable is so sovlful. If you have tool calling enabled and something like a dice tool available, Fable will sometimes just roll it for the sake of it to trigger a new prompt and make you pay double.
Truly a patrician model. If you shout at it to not roll the dice unless asked to, it'll sometimes act confused and roll a 0d0 just to still steal money from you.
Pure soul.
>>
More people use Marinara Engine than Risu or Agnai. It should be in the OP.
>>
>>109237478
I decide what goes in the OP and I say no.
>>
>>109237486
Why?
>>
>>109237498
I despise women.
>>
File: wha.png (378 KB, 800x534)
378 KB PNG
i took a two year break
what did i miss
>>
>>109237478
https://poal.me/km84kv
>>
>>109237478
Astroturf go away
>>
>>109237523
Marinara Engine has replaced SillyTavern. It's much better and gets more updates.
>>
>he failed in /vg/ so he tries it here
lmao what a retard
>>
>>109237535
>no chub option
>>
>>109237539
looks ugly as shit
>>
>>109237561
vibe slopped
>>
>>109237570
Do you hate AI?
>>
never in my life did I ever think I would be fiending for gemini of all goddamn things
>>
>>109237581
Nah I hate myself
>>
i miss sonnet 3.6
>>
>>109237299
>Chinese distillers would love that.
Why are lying, cheating, communist scum allowed to steal everything with zero repurcussions? Put them in the fucking dirt. Their models should be banned.
>>
>>109237561
trooncoded ST
>>
>>109235216
Nigga, half the fags here were head over heels over chai. Motherfuckers fall in love with fucking novel characters and you're trying to act like this fella is somehow on a basement below ours.
>>
>>109237586
why
>>
>>109237136
kek
>>
File: 1763715585032985.gif (3.63 MB, 346x346)
3.63 MB GIF
>the rider-bike
>the bio-panel
>his forearm-console
>a Gate-Watcher
>determiner adjective-noun
WHEN did this become a claude-ism?
it's annoying as fuck
>>
>>109237389
quality is utmost priority to the general and ST beats Marinara in all benchmarks
>>
File: 19820356.png (231 KB, 492x670)
231 KB PNG
>>109237535
>https://poal.me/km84kv
>>109237478
SILLYTAVERN MOGS
HAHAHAHAHAHAHAHA
>>
>>109237539
I'd sooner use Lumiverse... but I don't use that either.
>>
File: insecuity.jpg (34 KB, 462x280)
34 KB JPG
hello bros, im currently using gemma 431b on janitor ai (with a proxy). It's not terrible but lately there have been quite a lot of annoying issues. and my main problem is that my chats are becoming too long and spending too many cookies. For free, what would be the best thing for me to use? im not completely retarded as i can install silly tavern, would really appreciate some tips
>>
>>109238220
Summarize and never use more than 28k context
>>
So what do Lumiverse and Marinara Engine offer you that SillyTavern doesn't?
>>
>>109238220
Per day (and free) Google gives you:
- 20 3.5 Flash requests
- 500 3.1 Flash lite requests
- 1,500 Gemma 4 31B/26B requests

You just have to go to AI Studio and make an api key.
>>
File: 1720573255792049.jpg (36 KB, 352x429)
36 KB JPG
>>109237561
>>109237539
>>109238034
I use my own jeet coded garbage.
>>
>>109238220
Use MemoryBooks, or something similar, to summarize your RP instead of sending fucking everything. Make sure that you set it up to auto-hide messages that have been put into a memory (I let it leave 10 messages unhidden going back into the memory to maintain consistency with what's going on).
>>
>>109238220
>im not completely retarded
Oh yeah? We'll see about that.
>im currently using gemma 431b
Oops! Better luck next time.
>>
If you haven't vibecoded your own ideal frontend yet, you don't belong in this thread. You might be a better fit for the retard thread over at >>>/vg/.
>>
>>109237535
This shit is botted, isn't it? I don't believe we have that many Risu users here.
>>109238355
I'm gonna keep staying here and you can't do a thing about it. Maybe apply to be a janitor if you want retards like me gone.
>>
>>109238358
yeah, poals have little protection from botting
>>
File: 1780320325129103.png (135 KB, 1084x619)
135 KB PNG
Inside-out with Nilou, almost better than the RP itself
>>
writing style of gpt-5.6 is fucking horrible, like 70% of its "paragraphs" are a single sentence
i don't understand how they're making it worse with every release
>>
>>109238468
>i don't understand how they're making it worse with every release
Either vibe coding or increased reliance on jeets.
>>
File: doggggg.gif (3.75 MB, 380x640)
3.75 MB GIF
>>109238227
>>109238335
thanks i did it and it did improve quite a bit
>>109238245
is it uncensored? i would like to do naughty naughty things. I havent found anything better than gemma so far (not that i searched all that much)
Once again thanks a ton for the help
>>
Marinara engine update?
>>
Crusader King-ish frontend is the only answer my friend....
>Grand Cooming Simulator
>Dynamics event based mechanincs
>immersive interaction with multiple {{char}} be it concubines, duchess, countess, empress, breeder, mistress, daughter-wife, etc on your/their court/dungeon/bishop/barracks/etc
>static/predefined {{char}} at the start of game/session/history plus dynamics {{char}} (newborn, wandere/adventurer, slave, foreigner)
>CYOA
>scriptable with lua
>>
Test chatbots
No way my home IP is temp rangebanned
>>
>>109238634
lolwut?
>>
now that I finally reigned in groks over horniness I'm really liking it. Your JB has to state not to write like a smut writer or it always tries to lead to sex from any sort of flirting at all.
>>
>>109238670
also it over focuses on fetishes even if you dont direct it to. It assumes something has to be fetish related
>>
>>109237994
lm studio
>>
>>109238468
its for coding. Token efficiency is super important for costs. GPT5.6 is good for SFW if you tell it to copy a authors style. It knows as much as fable. But good fucking luck with NSFW, even with tons of context its super bland
>>
So uh am I going crazy or is connection profiles gone from sillytavern
I noticed this a few days ago but I thought maybe I was just seeing things
But now I need them and it's not there no matter where I look
>>
>>109239024
you disabled them
>>
>>109238670
*gets naked* Oh no, this post seemed to have obliterated my clothes! How could this happen to me!?
>>
>>109239029
how ?
>>
>>109235098

no anon, scrape get scraper content, craft pick get adaptive content maybe
>>
gemini preset?
preferably with anti-echo
>>
>>109238670
>>109238676
thank you for vaugely and incompletely describing the general idea of your prompt in uncertian terms
>>
>>109228984
Grok 4.5 > GPT 5.6
>>
poop > turd
>>
>>109239176
https://rentry.org/minipopkaremix
>>
is there any model that lets you do cunny aside from deepseek i havent had any luck
>>
>>109239291
lgm
>>
>>109239291
kimi
anything local
>>
>>109239291
gpt 4.1
>>
>>109239291
Yea mine
>>
>>109239291
mechaepstein8000
>>
>>109239319
>tweeted by elon himself as a joke in an alternate timeline where she got Her Turn™ and the files never got any spotlight
>>
>>109239291
all of them if you jb them except gpt and fable which turn absolutely retarded when you manage to
>>
>>109239372
is this supposed to be humor?
>>
Anyone have a reliable jailbreak for GLM 5.2 or Kimi 2.7? For some reason using a prefill doesn't work.
>>
>>109239708
Kimi 2.7 is codeslopped don't use it
>>
>>109239708
are you getting cut offs and refusals on glm now too?
>>
File: 72120835_p0.png (1.66 MB, 3172x2164)
1.66 MB PNG
>>109239291
you have 3 weeks left to enjoy the undisputed best cunny model (opus 4.1), get to it while you can
>>
File: gpt-4chan.png (68 KB, 923x394)
68 KB PNG
>>109239291
gpt-4chan
in fact this model is so evil (cunny, race & iq, among other things) it had to be disabled
>>
>>109239291
fable
>>
>opus
>good
>>
We are merging with /vg/



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.