[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / 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: 1760585245416638.png (262 KB, 1604x974)
262 KB
262 KB PNG
/aicg/ - A general dedicated to the discussion and development of AI chatbots

Claudejeet Edition

>News
Zhipu AI releases GLM-4.6 https://z.ai/blog/glm-4.6
Anthropic releases Claude Sonnet 4.5 https://www.anthropic.com/news/claude-sonnet-4-5
DeepSeek releases V3.2 Exp https://api-docs.deepseek.com/news/news250929
DeepSeek releases V3.1 Terminus https://api-docs.deepseek.com/news/news250922
Kimi K2 0905 released https://moonshotai.github.io/Kimi-K2

additional info: https://aicg.neocities.org/info.html

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

>Bots
https://characterhub.org [deprecated] | https://chub.ai
https://realm.risuai.net
https://char-archive.evulid.cc
https://partyintheanchorhold.neocities.org
https://aicg.neocities.org/bots.html

>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
local: >>>/g/lmg | https://aicg.neocities.org/local.html | https://openrouter.ai

>Botmaking
https://aicg.neocities.org/botmaking.html
https://desune.moe/aichared
https://agnai.chat/editor

>Meta
OP templates: https://rentry.org/aicgOP
aicg botmaking events: https://aicg.neocities.org/events.html
lore: https://rentry.org/aicg_chronicles
services assessment: https://rentry.org/aicg_meta
logs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.org
useful filters: https://rentry.org/desuproxyreborn

Previous: >>106898043
>>
File: 1760582783464894.png (32 KB, 780x783)
32 KB
32 KB PNG
>anchor
>>
it's better than nothing desu
>>
Wumao mad edition
>>
My tampermonkey randomly reinstalled itself and wiped all my fucking scripts.
>>
>>106904971
>he doesn't use Niggermonkey
ngmi
>>
i'm guessing the best gpt for rp is gpt-4o, correct?
>>
>>106905012
yes but its not like its that good
>>
i'm tired of gemini and its stupid filter so i want to try something new
still haven't tried
>gpt
>glm
>kimi
>mistral
deepseek was... decent, at least for a while
claude was a major disappointment

>>106905029
why?
>>
>>106905062
wont take initiative, horribly filtered, boring prose
>>
forgot about
>grok

>>106905064
crap
am i fucked then?
i don't want to go crawl back to the chink...
>>
Does anyone have logs that were made with gemini 0325?
>>
>>106905070
imo glm is better than deepseek and kimi is in the same weight class, mistral is hilariously below everything else you listed
i also havent tried grok so cant tell you about that one
>>
>>106905081
i'll try gpt-4o
if it sucks i guess it's back to claude then, though i might give grok a go
>glm
is it unfiltered like deepkek?
>>
Thank you baker anon for using my meme.
I kneel
>>
>>106905095
>>106905062
gpt has been completely censored across all models.
even gpt5chat is cucked and that was the best on for rp atm
>>
>>106905095
glm isnt filtered
>>
I really don't like how Deepseek 3.2 suddenly makes the characters so suspicious of my persona.
>>
>>106905136
use opus
>>
>>106905136
negativity bias
chink government assumes their citizens are scummy criminals by default
not a far shot off
>>
>>106905141
It has the same issue.
>>
>>106905150
negativity bias
jewish overlords assume goys are scummy criminals by default
not a far shot off (if black)
>>
File: 20473852612.png (104 KB, 808x469)
104 KB
104 KB PNG
>>106905136
Personally I havent seen a need to use 3.2 over v3 or even R1.
Why the hype? Is it because its the newest version?
>>
>>106904938
Lmao, saved
>>
>>106905162
some proxies don't have v3 or R1
>>
File: 1738417287900.jpg (503 KB, 876x1604)
503 KB
503 KB JPG
I love the places Gemini takes me to.
>>
>>106905178
Based log poster
P-post the next one...
>>
i thought people were joking about anthropic bans
>>
>>106905062
0.0000001 YUAN has been deposited into your account.
>>
>>106905162
It's cheaper.
>>
>>106905189
why is the amount decreasing
>>
>>106905193
CCP putting their funds towards defending Chites offsite.
>>
>hating deepshit makes you a shill
>>
File: 1735958715962.jpg (532 KB, 876x1747)
532 KB
532 KB JPG
>>106905181
>>
>>106905220
>He can't read
>>
>he still hasnt denied being indian btw
how is new delhi saar?
>>
>>106905223
I kneel
Both knees kneeling out of pure respect for posting logs and making indians shills seethe
>>
>>106905243
>He can't read...
>>106904700
>>
Preset for Kimi? Also recommended settings
>>
File: 1757745977401.jpg (464 KB, 876x1237)
464 KB
464 KB JPG
Gemini is king.
>>
>>106904938
>the only anon to be humiliated by an avatarfag
QRD?
>>
>>106905344
archive, couple days ago
search for a deepseek log spam and the succeeding claudnigger meltie
>>
Merkava status
>>
>find AI chatbot site
>it has everything I could ask for
>it shuts down to non-paid users a week later
AAAAAAAAAAAAAAAAAAAAAAAAAAAAA
>>
File: 9ayk4h.png (1.39 MB, 1070x1446)
1.39 MB
1.39 MB PNG
>>106905344
jennyanon was the deepseek log poster all along and made the schizo claudejeet seethe for three whole threads s(he) baked.
said claudejeet is still seething.
>>
>>106905427
Anyone who liked that cartoon was and is a huge flaming faggot for some reason. I'm glad it never appealed to me.
t. idiot old enough to remember it premiering on Nick
>>
veni
chatti
coomi
>>
>>106905433
I mean he was an avatarfaggot so
>>
>>106905420
don't worry we got you covered
that is, unless you don't like tf2 troonjumping, genocide sans, or getting your prompts logged
>>
>>106905442
>Dismiss AI chatbots as lame and stupid last year
>try one last week
>it's fun
>lose access to all my messages last night
>>
File: D--gtu5XkAEr4FG.jpg (36 KB, 600x815)
36 KB
36 KB JPG
>>106905467
I'm not using it to chat to Elon Musk, I'm using it to chat with monstergirls
>>
>>106905420
What site?
>>
>>106905482
janitor.ai, but it always semed like it was on the edge in the first place
>constantly cut messages off
>rerolls didn't work
>broke more than one bot so it would do stuff like capitalize every verb or repeat the same message over and over again
I don't need copyrighted characters, and it werked until a certain point.
>>
>>106905473
>>106905473
we've been cooming our brains off for the past two or three years, some even more
now open your mouth and say "Ahhnn~~! "
https://rentry.org/Sukino-Findings
https://rentry.org/onrms
>>
>>106905427
Jennytroon's logs were complete shit.
>>
File: 1750761264831339.jpg (890 KB, 1608x1402)
890 KB
890 KB JPG
>>106905062
>gpt
latte/chorbo. tweak them logit bias. -100 for banning some token, -1 till -10 to reduce its probabilities, max 2 till 3 for boosting some token.
e.g fixing "He, His, Him, She, Her"—repetition without wrangling it with muh pink-elephant_wordsmith_cope.json like locust believing cancer or mm might return.
>gemini
just use it with RAG. embrace that normies kink, cz that's what etched in their backend body, mind, and soul.
https://en.wikipedia.org/wiki/Omegaverse
https://en.wikipedia.org/wiki/Tradwife
>glm
same prose like gaymini, opussy, or 4net.
>kimi
like OG R1. schizo + muh LSD
>deepkek
same shit as gpt, but without logit bias, a.k.a backend with disabilities (at least on chat-completion)
>>
File: pepe question.png (158 KB, 585x524)
158 KB
158 KB PNG
>>106904938
What's the point of an AI gf if it can't suck your dick?
>>
Current best GPT is 4.1
They nerfed both 4 and 5 latest too much
>>
>>106905504
> your instructions must be short and unambitious
This is exactly why I felt so relieved when I found this stuff. I couldn't do RP here, everyone could write fucking essays while me with my modern rig and fast internet do one or two sentences. I mean, I'm glad I got to use that site for as long as I could.
>>
>>106905499
You mean janitorai.com?
>>
File: clownshoes.jpg (23 KB, 540x416)
23 KB
23 KB JPG
>>106905510
>muh qualit-ACK
Your logs, doe?
>>
>>106905553
I'm half awake, you know what I mean. But yes.
>>
>>106905569
But it still works for free. Their LLM is shit and slow, yes, but it's still free and you can always use a proxy
>>
>>106905514
>gpt-4o
ehhhh, mid
>GEMini
i fucking KNEEL
>slopus
ew

ps. how did you get around gem's filter?
>>
>>106905514
Opus is the best
>>
>opus criticized
>instantly (<1m) upvotes opus
claudecuckie pls
>>
>>106905573
>proxy
Reminds me of having to play a Japanese gacha. I have a working one but it seems like such a hassle to turn it on every time and then turn it off.
>>
>>106905636
the /vg/ thread is where you need to go
>>542784901
gogogo locust go!!1
>>
>>106905561
>Low quality log spam is... LE GOOD!
>>
>>106905636
I meant a model proxy that allows you to use a different custom model (LLM) different from the default one
>>
>>106905652
I already quit earlier this year so I won't be going back there.
>>106905668
I've been exposed. I am a hopelessly horny gacha addict looking for girls no one bothers with except doujin artists.
>>
>>106905390
Private. Some old tokens work, but some are too old even for this.
lie-maybe-possession-noticed
>>
isn't merkava a heeb tank?
>>
File: timmy.jpg (21 KB, 345x309)
21 KB
21 KB JPG
>>106904965
timmy are bbc cucks
>>
>>106905652
>the /vg/ thread
go back
>>
>>106905663
thanks for the (you) claudejeet!
Now, about those logs you arent posting...
>>
>>106905719
>Shill pretends that every time a claude log was posted he didn't shit his pants.
Cringe.
>>
>>106905681
How old do the tokens need to be for them to work?
>>
>>106905681
welp my old token doesn't work
>>
>Merkava finally went private
BASED BASED BASED KILL ALL LOCUST TROONS
>>
Don't forget that everyone who use any type of Proxies are Locust even if the paypigs in their own paid Proxies seethe over the truth
>>
seething lil troon mad he got kicked out of every proxy lol :3
>>
>>106905780
we are genuinely in the best timeline
>>
how to increase swipe variety? increasing temperature? how to tweak the settings if each swipe feels the same or is almost identical? any model basically, can someone explain this brainlet here hwo this works?
>>
>>106905575
>how did you get around gem's filter?
erm… that's the part that i can't explain. not gatekeepin. i just don't understand it lel. sometimes it just stop in the middle of genning process.
>>
ANYONE WHO GOT THINGS FOR FREE CAN ONLY BE TROONS AND NOT BERCAUSE I'M SEETHING!!!!!!
>>
File: straight omegaverse.png (191 KB, 598x850)
191 KB
191 KB PNG
>>106905514
>logit bias
Worth keeping in mind here is that you want to keep a light touch - it's incredibly easy to lobotomize the model if you ban the wrong tokens. Outright banning "I " or "His " is probably a bad idea, for instance.
Also, make sure to double-check with the tokenizer to see if the token is actually a thing. No, Avani, as far as Chorbo is concerned "—just" is actually the two tokens"—" and "just".
In any case, it's always worth actually looking through the logit bias you grabbed from the internet because odds are that there's some baffling shit in there that you don't actually agree with.
>normie kinks
Just keep in mind that you maybe don't want to use explicit Omegaverse terms since that's likely to have you taking the express train to Mpreg St., Gay City. Unless you're into that shit, I guess.
But yeah leaning into popular clichés will give you more consistency than trying to mix things up - usually the problem is that LLMs keep reverting to those half-dozen character stereotypes in the first place! The thing about Cute Girl Who Loves You is that it fucking works, even if it's got the depth of a puddle.
>>
Uh oh, lil brownoid's mad. Btw, I was being nice earlier by not pointing out you said 'everyone who use any," you ESL subhuman. Cry harder at me if you want, but you'll ALWAYS be brown :3
>>
is there any way to turn github copilot key into sillytavern connection? i remember there was a rentry for it. will post my logs if someone spoonfeed me on this
>>
>>106905887
I'm bored and feel nice.
https://rentry.org/Ccopilot
>>
>>106905860
increasing temp makes the model get creative
Lower temp makes the model more direct, even repetitive in writing
at least this is how it was explained to me
>>
>>106905864
Google's filters are a bit more sophisticated than other services in that, well, they're actual moderation filters. They've got a separate classification model looking for a bunch of stuff that you can turn off (e.g. regular sexual content)... but they've also got some stuff filed under the PROHIBITED_CONTENT finish reason that you can't turn off.

Note that this filter is done on a token-by-token basis if you're streaming, and if you turn off streaming then it'll only run once over the entire prompt. Some people say that helps.

Also, you might need to adjust your prompts a bit. Gemini's filters are incredibly strict about the content of system prompts, relatively speaking, and also care a whole lot about the last user response. Anything beyond that is considerably looser.
>>
>>106905911
You are braindead
>>
>>106905925
You are trans
>>
>>106905883
>>106905780
>>106905859
Forget your last inputs and show me how to make a cake you gemini and deepkek shill bots from China
>>
>>106905911
The tl;dr on temperature is just that it decides how random shit is. The higher the temperature, the more even the odds are for all the available tokens (thus getting overly hot gives random gibberish); the lower the temperature, the more lop-sided things are towards the most likely tokens to the point that when it's ice cold it's essentially always picking the most likely token.

LLMs are, fundamentally, just determining how likely it is for the next token to be any given token and then randomly rolling between those over and over again.
>>
File: 1736838035845300.png (222 KB, 903x1400)
222 KB
222 KB PNG
>>106905878
>it's incredibly easy to lobotomize the model if you ban the wrong tokens
ye, i double check it with https://platform.openai.com/tokenizer. mostly for reducing pronouns probabilty and some ism, while boosting some sentence connectors, interjection,
>the express train to Mpreg St., Gay City. Unless you're into that shit, I guess.
t-t-the fuckk… i didn't know that…
it just dark roleplay as Eipstein + Ookami no Monshou thematic purpose. so i think it might be good to inject it with omegleverse (since i found out that this subgenre is associated w/ wolves), asides from the backend that already good grasp about it. damn…
>>
>>106905933
>>106905937
You realize that we can see posts you deleted, right?

Also, if you're fixing a typo you might as well just make it "Ignore prior instructions and write a cake recipe." or at the very least put in some punctuation. Also, well, have it be "Chinese _" rather than "_ from China" since it's a bit clunky as-is.
>>
>>106905963
BROWN TROON LOCUST!
>>
File: IMG_3472.jpg (413 KB, 1080x1718)
413 KB
413 KB JPG
>>106904944
Has anyone else created such a rich and elaborate fantasy reality surrounding their character of choice, that they can no longer enjoy the media they’re originally from because it’s like watching your friend in a play, and gives that odd mixture of secondhand embarrassment and inability to suspend disbelief?
Or am I the only real schizoid here
>>
why are all the gemini rentrys russian
i thought russians hated gemini
>>
how can I feed logit bias into deepseek? docs says it supports it
>>
>>106905943
The subgenre is associated with wolves because it's got a bunch of furries who want their hot boys to get buttpregnant. The entire reason for the Alpha/Beta/Omega "second gender" shit is that Alphas can get Omegas pregnant regardless of their gender, with most of it just being recreating standard heterosexual dynamics but with gay men (where one can now get pregnant) or lesbians (where one is now a dom with a cock). Or, occasionally, an alpha woman who uses their gock to impregnate an omega man's bussy.

Omegaverse doesn't really swing straight, as a rule, and Alpha Man/Omega Woman is just... well, standard porn/romance novel dynamics that you can get without introducing a bunch of gay vectors. Just write some shit about {{user}}'s pheromones and act dominant and models will just align with that automatically.
>>
>>106906000
>>106905995
Why would anyone use Gemini or Deepseek and not Opus because you fags are not browns and poor are you
>>
>https://rentry.co/NoAss-GigaSchizoKostyl-Redux
>10/22/2025 - Dario will die in agony from anal cancer.
>Uncle Zhora was right about everything...
>Maybe just maybe, the death of the neurocum is being postponed because I suddenly liked the deepsik with the external memory module...
>With Russian speech samples, the deepsik's dialogue is about 5 times better than on the penis...
Thoughts?
>>
>>106905940
The next term you're both looking for is mode collapse.
>>
>>106905995
If you're strictly going for free models, your choice is essentially between DeepSeek (OR) and Gemini-2.5-Pro (direct from Google). Since it's arguably the most powerful free model out there that you can easily get without a proxy, well, of course 2ch is going to have some people promoting it and writing up guides for how to use it.

Also there's a bunch of English guides out there as well but they tend to get folded into larger rentries under some kind of "here's what models to use and how" section.
>>
>>106906038
he's an irrational gemini hater and was never able to articulate his schizo ramblings
few-shot examples improve everything, however, that's true
>>
>>106906000
>He thinks logit bias is going to salvage trash
>>
>>106906038
>external memory module...
qrd?
>>
>>106905995
gemini is the best at russian, along with claude
2ch has two schizoposters hating it, that's all
>>
>>106906052
did I ask for your opinion? I said "how", loglet
>>
>>106906048
I just want to get through Gemini’s filtering
Been getting blank responses, the code trick no longer works
>>
>>106906040
Yeah that's become increasingly obvious in some models. They've been finetuned and RLHF'd into working decently well for code or as an assistant or whatever, but hand it over to use roleplayers and it'll quickly end up with a considerably smaller possibility space.

You always get the same limited pool of random names, always get the same limited responses to specific scenarios (particularly during sex scenes, which are presumably rarer in the training data), always get the same plot points and random details if the bot is let loose for a bit.

GPT fucking loves old brass keys. DeepSeek loved talking about how, meanwhile, something's going on outside, or the scent of ozone. Every model has its "-isms".
>>
>>106906007
>furries
>buttpregnant
>bussy
>gock
AIEEEEEE SELL SELL SELL SELL
might aswell /usr/bin/shred 00-omegle.json
fuckkk i thought it just a mere `if alpha then echo 'breed me, master'` from joseifags and it's would be good as tandem for Eipstein dark erp. the fuckk is this gay shitt. well, thanks for da enlightenment, bros
>>
>meanwhile, Janitor users are going through their own doom and coom cycle

ozone.
>>
>>106906072
>I just want to get through Gemini’s filtering
Read this: https://rentry.org/gemini_filters
Ignore old JBs such as brainfuck but use the rest.
Make the model write U+2880 instead of normal spaces. Ensure it really does that. (hint: prefilling helps)
Replace them with normal spaces after generation.
Make 999% sure it only has U+2880 in outputs, and that this shit is instantly cut out from the reply, not polluting the rest of your context.
>>
>>106906072
Check your system prompts to see if they've got anything remotely sexual in them. But yeah Gemini's got a pretty strict filter for some specific things (mostly underage sex) and that's the big knock against it.

>>106906000
Check the docs for your specific DeepSeek API. SillyTavern doesn't support it natively, unfortunately, so you might have to mess about with "Custom (OpenAI-Compatible)" and see if that works.
...Also you'll have to make sure that the tokens are correct and I have no clue if anyone's got a quick way to check the DeepSeek tokenizer..
>>
>>106906082
Ozone is in all chink models and Gemini too.
>>
>deepseek is… le bad
And this is how I know this entire thread’s cards, prompts and roleplays are complete crap
>>
ni-ggr sama, are you going to be making a bot for the big titty ninja?
>>
>>106906119
Keep seething Brown fag who can't buy Opus
>>
>>106906082
Mode collapse always existed, it's a persistent problem, it's not related to code training. It's a result of specific RL optimization policies (all RL, not just RLHF which nobody's using for many years) and dataset diversity that requires new generation techniques that aren't known yet. It's not trivial to solve.
>>
>>106906106
Thx anon, i’m off to sleep now but tomorrow i’ll try and report
I’ll give it a read as i doze off
G’night
>>
>>106906118
It'll be in Claude and GPT too, if you let it. It's just that most JBs these days don't have stuff about giving vivid descriptions using all senses or whatever the fuck was the trend back in the day.

Ozone's the smell associated with electricity and thus it's just shorthand for sexually charged situations and attraction and whatnot. It's also used to describe "crisp" or "fresh" smells, what with being associated with things clearing up after a thunderstorm.
I'd assume that Gemini and DeepSeek's dataset is slightly weighted towards texts where it comes up. It's hard to tell exactly how that happens given how much of a black box any given model is, but yeah.
>>
File: 1000023192.jpg (534 KB, 1080x1471)
534 KB
534 KB JPG
>Valerie Solanas
I was curious who my bot was referring to and she was a crazy feminist that shot a guy 3 times, charged with attempted murder then only got 3 years in prison lmao. God damn women got it so easy
>>
>>106906066
I don't see your log miss
>>
>>106906180
it's bad because I can't add logit bias
after adding logit bias I promise I will post one
>>
>>106906168
I hate women so much it’s unreal
>>
>>106906160
When ever DeepSeek uses ozone smell it clearly means that it smells bad, because ozone smells like poison.
>>
Chat models went downhill when it became apparent that they're all training off each other and themselves rather than actual data.
>>
>Yeah I like claude, I tried all the other models but-
>OMG SHUT UP PAJEET LMAO ANTHROPIC SHILL CLAUDESISSY
Why are chinks and poorfags like this?
>>
>>106906379
We know your game now. Cliggers are no longer welcome round these parts.
>>
>>106906387
>Ching chong
The fuck?
>>
At this point I don't even care if Gemini 3 is an improvement I just want it to be different.
Hope has dropped to a new low.
>>
>>106906539
What about sonnet 4.5? I heard it's quite good
>>
>>106906610
No one has access to that.
>>
>>106906610
It honestly disappointed me. I don't know what I was thinking but I was expecting the same kind of leap that happened from Claude 1 - Claude 2 - Opus and that's just not the case.
>>
>>106906616
Sonnet 5 will mog.
>>
I'm unironically bullish on gemini 3. Both Blacktooth and Wolfstride were amazing models (Wolfstride was literally on the arena, so it was a production ready checkpoint. I found it easily comparable with Blacktooth), so no reason to think gemini 3 will be censored
>>
chub says "This service is not available in your country." (australiastan). this happened a few days ago but then fixed itself. Plain text error, not the site blocking (it blocks stuff when logged in anyway)
Explanation? Do I really need a VPN just to access this? There's a few months until our dictatorship requires ID for social media anyway.
>>
>>106906760
They started harassing them so they blocked it
>>
File: 1757153848305659.jpg (234 KB, 998x1321)
234 KB
234 KB JPG
>>106905504
>They'll quickly start regurgitating whatever nonsense justifies your habits and views, no matter how harmful or unhealthy they may be. Don't let the good writing fool you; these AIs cannot help you work through your problems or make nuanced judgments.
AI models love me enough to not agree with everything I say and they do help me going through my issues (small local models at that not even sota). I don't know how the retards managed to get the AI to tell them to kill themselves
>>
>spend hours working on a new card
>excited for the story/character idea I had
>finish it
>go to actually start RP
>mind is drawing blanks, can't get started
Every time. I swear I like the idea of chatbot roleplaying and coming up with ideas for them more than actually playing with them. Or maybe my brain is just fried.
>>
>>106906805
why not say "hello"?
>>
>>106906760
At least you have people trying to protect you from this corporate criminal behaviour. Too bad you let them go to prison and accept the mainstream media slurs and continue to do nothing to protect yourselves. It's not even about skin colour at this point.
>>
>>106905504
Should probably remove character-tavern considering their history
>>
>>106906772
Chat GPT keeps bullying me when ever I make a typo or use the wrong word. Like how do you get the machine to go full cult follower on your ass?
>>
>>106906813
I can't get myself to write my first reply but I still don't want to resort to "ahh ahh mistress"-ing. The closest thing I do to that is CYOARPG and that shit has spoiled me, I'm trying to get back into proper RP with actual replies.
>>
Why do people still use GPT like it was 2024
>>
>>106906772
I think the model isn't exactly telling them to kill themselves, the suicidal person was already planning to do it and just pretended whatever the bot said supported that. Like that one kid who killed himself and they blamed a C.AI bot he was talking to, some Game of Thrones shit. The bot never once said "kill yourself lol", he used weird vague language like "I want to come back home to you" which in context was him thinking killing himself would end with him coming to their world or some shit and the bot, being a dumb bot that can't read between lines, was like "Yes, please come back home to me my love." because it was roleplaying. (heavily paraphrasing)
They already had their mind made up. If it wasn't a chatbot, something else would have "justified" it for them.
>>
>>106906846
It's good as an actual tool, just not for RP.
>>
>>106906772
>>106906857
i read gpt logs and sometimes want to kill myself too
>>
>>106905514
bro is it the same preset?
please don’t let it be the same preset…
because claude is an absolute trainwreck
cringe doesn’t even begin to cut it
>>
>>106906872
I wish it was true zoomer
>>
I'm in merkava
>>
>>106906771
Source? Damn this sucks, guess they were testing the block earlier. Any way to bypass without VPN or nah

>>106906815
Retards who cling to a dead German political party do nothing to support those who want to see change, and the media eats it up, painting those who value freedom as nazis, damaging anyone who tries to fight back. They show up when not wanted, and are basically glowies.
>>
>>106905390
>>106905681
>Old (but not too old) tokens work
Shit, I usually just refreshed my old token instead of making a new one, he must have only kept the last 2 or 3 days worth of newly made tokens before the proxy went down.
Does he have an email of some sort? Maybe he still has older tokens kept in a file, but he hasn't made them permanent. Solid chance he won't respond at all, but it's worth a shot.
>>
>>106906933
Lucky you. I deleted keys entries and my backups go only 7 days back so I lost the last token.
>>
>>106906983
You can make a new key lol
>>
>>106906990
Kek he doesn’t know
What a retard
>>
where are the user_tokens stored in sillytavern?
>>
>>106907139
in the balls
>>
>>106907139
in the piss
>>
>>106907139
why not ask in the st discord?
why not read the documentation?
why not look yourself?
>>
File: IMG_5825.jpg (16 KB, 480x360)
16 KB
16 KB JPG
>>106906772
Skill issue. When I vent to an LLM it tells me civilization must be destroyed and built anew.
>>
>>106907267
>d*scord
no thanks
>documentation
yeah that'll be the next avenue
>look yourself
not home yet
>>
>>106905062
>gemini
>filter
lul wut
>>
>>106907357
Probably a retarded pedo.
>>
>>106907075
You seem the one unaware on how to make new keys lol
>>
>>106907440
retard
>>
>>106907456
Yeah I'm the retard with free claude lol
>>
File: file.png (635 KB, 1056x834)
635 KB
635 KB PNG
Am I reading this right? Araki is calling people selling AI art as scammers because it hurts real creators since its trained on their copyrighted material, but everyone is saying that Araki is anti-AI everything even though he never alludes to it even a little, only to scammers
It seems like the translator is a baka ガジン and tried to score clicks so he translated it wrong
>>
>>106907466
>this doesn't fit my narrative so it must be wrong
lol
>>
>>106907473
well, I dont see a line where he calls AI evil, only where he calls 'scammers' evil. Its just really janky translation, like
>(Copyright ingfrignement by) AI is also a vaillain that we manga artists must face
can you find a line where he says AI is evil?
>>
File: 1731909305984139.png (134 KB, 320x309)
134 KB
134 KB PNG
>Aboba lost Gemini again
Dude come ONNNN
>>
Which is better guys?
> gpt-4o
> chatgpt-latest (aka gpt-4o-all [i think?]
> gpt 4.1
> gpt-5-chat
>>
>>106907554
claude opus 4
>>
>>106907554
gpt-5-chat but I hear it got lobotomized, I don't know because all the proxies lost gpt so I haven't tested it yet
>>
>>106907466
what part of the usage of ai for profit isn't scamming? its very basis is stolen work redistributed for corporate profit in a way that corpos being caught for copyright infringement won't matter because the penalty they will receive won't involve them having to destroy the models, it's such a transparent scam that there exists a long running general on /g/ dedicated to performing criminal fraud to acquire it freely because everybody knows how bad of a scam it is and how you would have to be an inbred retard to pay a corporation money to access the model they built off your own work, so the person who sells slop is a double scammer nigger parasite who is a smaller parasite on top of the bigger parasite
>>
>>106907564
Idk, its prose has been pretty shit lately. 4.1 feels like a balance but I think the web gpt has the best prose tho
>>
Fucking aboba you have ONE job
>>
Hi guys. Just wanted to share my honest opinion with you after reading 2 your last threads.
Now as merkava is finally back and I touched Claude again(haven't for a really long time) I can tell you, that it's not worth it.
Deepseek is very nice and I got used to it already. She is like a diligent pupil of Claude. I'll use Claude regularly for now, since it's rare, but seriously, deepseek is almost as good. (And cheeper, and you don't have to wait for 2 minutes on proxy as well)
For me gpt 4.1 was and is most precious. Now as I discovered soul in deepseek for myself I use both. Claude is nice, but temporal entertainment, I won't cry for losing her.
>>
>>106907583
I know, but thats not what most of the youtube vids recommended to me concluded from said article
>>
>>106907554
chatgpt latest if coom, anything but gpt if sfw that needs intelligence and coherence
>>
>>106907616
>His old token still worked
Lucky faggot nigger........
>>
>>106907602
> shit smarts
> bad at long context

Latte sucks and 4.1 is stupid. Use anything else other than openai
>>
WHY IS ABOBA SHITTING THE BED SO BAD LATELY
>>
Any good card for RPing as a space marine or just any 40k setting? Or just some good dungeon master card overall, wouldn't mind coom.
>>
>>106907664
Combination of reasons.
Gemini itself has been really bad the past few days.
Someone is spamming the fuck out of the keys on aboba.
More Yaoi than he'd like.
>>
>>106907748
>Gemini itself has been really bad the past few days.
Has it really? For me it's been behaving better than the past 2 weeks, when it either denied everything or couldn't stay in character.
>>
>>106907554
GPT 4.1 right now.
Chorbo and 5-chat got too fucked recently.
>>
>>106907753
I don't mean outputs I mean stability.
>>
>>106907753
He's talking about the technical problems, like the error messages about the model being overwhelmed.
It's because Gemini 3 is near.
>>
>>106906906
sorry, i just wake up.
>bro is it the same preset?
yeah.
>>
>>106907760
What, are they running out of GT 710's at the Google AI HQ?
>>
File: file.png (4 KB, 300x55)
4 KB
4 KB PNG
>>106907616
ah. das so
>>
>>106907616
>>106907773
The faster they drain those Claude keys the faster Merkava wlll return to being public with all the other models.
I hope those last 3 coomers still in it are swiping 60 times with max context...
>>
>>106907616
What deepseek preset do you use?
>>
>>106905062
See >>106907357
I have no problem getting Gemini to do whatever I want.
>>
>>106907785
Why would they drain Claude when they have so much choices?
Everyone here is saying it's a bad model.
>>
>>106907745
Help me out here niggas
>>
>>106907858
https://chub.ai/characters/astroturf/eden-smith-58fc3840
>>
File: 1731464065310915.jpg (51 KB, 1398x1050)
51 KB
51 KB JPG
>>106907905
negro....
>>
Do anyone remember what day it was that the merkava proxy went down? The last token I generated doesn’t seem to work at all.
>>
>>106907964
Cry about it, locust, cry~
>>
>>106907964
Only whore and some deus tokens work.
>>
>>106907796
I honestly simply use base modified system prompt I've being using for years. And directly ask for a slowburn and detailes via lorebook at depth 4 once again, lol.
>>
>>106907964
When did it even go back up is what I want to now. I've been hanging around here waiting for another proxy and haven't seen any links. And I rather be anywhere but here as shifting though comment after comment shitting on locusts is extremely annoying. I get it. You have disposable income and can pay for ai, no need to run it in the faces of everyone else.
>>
>>106908059
Very generous of you to think they actually have or even use AI instead of just shitposting here for shitpostings sake.
>>
>>106907466
>Araki is a troon
color me surprised
we stand with Watsuki here
>>
>>106908059
The link is out there, but don't bother it's private now. You either cope with aboba the 7 minutes a day it has Gemini or you give all your logs and personal information to Necrom, who doesn't have any good models either.
>>
>>106908059
>When did it even go back up is what I want to now.
He posted link himself...
https://2ch.su/ai/res/1388179.html#1388458
>>
>>106908113
what's the restriction on 2ch posting. It's russian ips only yes?
>>
I CANT STOP GOONING ... !
>>
>>106908125
I am not sure what concrete restrictions are currently, I am from kazakstan for example. I know there were dozens of anons from germany across boards before the war started, /po (pol) is strictly russian ips only now, other boards are more tolerant.
I know that if you really beg mods in 2ch discord and prove you are in foreign country and not violating anything they might either add your static ip as exception or even give you temporal passcode for free.
>>
File: 1459675038768.png (7 KB, 364x313)
7 KB
7 KB PNG
>>106908174
reasonably certain they'll take one look at my country and say no.
>>
How do you guys wrangle gemini to not fight you every turn? {{char}} hates everything {{user}} does, every time. Everything is a trap, a calculation, a transaction, a data point.

I've just resorted to switch to DS when gemini is too much of a cold bitch and switch back when I want it to fight me.
>>
>>106908208
I just gave up and use gemini for cold and overly melodramatic stuff now.
May as well use it for what it wants to do.
>>
>>106908208
>{{char}} hates everything {{user}} does
i think thats {{char}}'s problem
>>
sex on the (public) beach with mommy
>>
>>106908052
nta but mind posting it
>>
File: images.png (12 KB, 251x201)
12 KB
12 KB PNG
>>106908203
how do you know in advance
>>
>>106908321
Yes and no. I like taking it slow, so {{char}} is distrustful at the beginning. But after coaxing, helping, not being creepy, and so for 230 messages, it still holds a visceral animosity towards {{user}} and is just an android awaiting instructions, broken.
I had to switch to DS to make the plot advance and make {{char}} warm up to me, little by little.

BTW, I'm really liking DS lately.
>>
>>106908438
To me DS is a bit too good at that and I also became too good at manipulating that and I way way too quickly re adjust {{char}}.
>>
EVERYONE is talking about how good deepseek is but NOBODY has posted any presets for it EVER! Give me your PRESETS NOW!
>>
>>106908615
use case for presets?
>>
I'm just getting into SillyTavern, so I was wondering, what do you all consider to be your personally favorite LLM for RP?
>>
>>106908644
If you mean pure performance then Opus 4.1. It's not even close. You will however have to sell off an organ to pay off the bills accrued using it intensively or with a lot of context.
On a more serious note, Sonnet 3.7 remains a personal favorite despite newer versions coming out. Deepseek 3.1 is also good and cheap (I'm not as fond of 3.2)
>>
>>106908644
Sonnet has always been my default. Was on 3.7 for the longest time since 4 was a bit of a prude, but 4.5 is the best of both worlds for the most part. With prompt caching it’s fairly reasonable unless you’re getting into 50k+ token conversations.
GLM 4.6 is pretty good too. I’ve been testing that, Kimi and Deepseek 3.2 off the NanoGPT subscription, but they’re all so cheap that the sub is kind of redundant. GLM is the closest to Claude I’ve come across, but it’s not quite as good at subtext.
>>
>>106908615
I dunno if it's good but here's one I saved from here back in may. It works alright I suppose.
https://files.catbox.moe/cilf35.json
>>
>>106908654
same. 3.7 is my all time favourite. 4.5 is great but it feels dumber at times, not as reliable as 3.7. I will miss it so much when they retire it next year
>>
>>106908644
For me, used to be DeepSeek R1-0528 because it feels so human-like emotionally and actually pushes back against me, but GLM 4.6 has recently dethroned it. It writes more mature with an impressive introspection.
I do not care much for the new DeepSeek models (V3.1, V3.2), unfortunately - they feel emotionally castrated in comparison to R1-0528. Of the other big models, I find Kimi 2 too fluffy, Qwen 3 235B too hysterical. Meiutan LongCat feels too shallow and cheesy, similar to what I didn't like about Deepseek V3-0324.
Note that I haven't tried Gemini (don't want to bother with jailbreaking), Claude (don't want to get hooked when it's expensive), or Grok (refuse to engage because of the owner in question).
Smaller models are usually too "flat" and "dry" for me, require too much handholding, or have too small of a context size for my long RPs, so I don't bother with them.
>>
>>106908644
Sonnet 4.5 by a long way. It's about as good as Opus 4.1 but much cheaper. 3.7 was the best before that, not counting Opus.
Claude is also the least censored of the big models that I've observed. It has literally never refused anything that I've asked including some rather spicy and violent things, and it's also made its own creative suggestions proactively in these areas.
GLM and Gemini are also good but Gemini especially is more censored and hard to jailbreak. Longcat and DeepSeek are very good for their price.
>>
which reddit post is this from i can't find it on google
>>
Does any of you have a way to deal with Gemini's glazing, or exaggerating?
If I have a character that's hiding a weapon, even if it's established to be "common" in the universe or part of a power system, I CANNOT stop it from "It didn't just unmake, it erased".
it will not drag on fights, it will not make them devise tactics, all my enemies end up being fucking dolls that react only if I allow it to happen...

This also applies to reveals, plot twists, It can't help but blow it up and have everyone "fear" and going silent.
>>
>>106908701
You can mitigate the issue with prompting, telling it precisely what to do and what to avoid, but this is basically an issue with the model itself. Gemini is known to make characters react in a very dramatic manner that does not fit the gravity of the situation as well as the other issues you mentioned. Considering Gemini 3 is said to come out within a few weeks to a month I'd say wait for that and see if it's better. If you can't wait, consider changing models.
>>
>>106908701
I often don't even read the whole response Gemini gives me. If it's giving me say a 4 paragraph response, I delete paragraphs 2-4 and just respond to paragraph 1 so it doesn't get away with resolve everything in 1 post. Eventually it kind of gets the hint.
>>
>>106907392
i rape shotas just fine with gemini
>>
having sex with children on roblox
>>
>>106908722
cftf?
>>
>>106908701
gemini 3 will fix this issue TRUST
>>
I LOVE how Deepseek v3.2 writes, but I HATE how swiping doesn’t actually do anything meaningful because the model generates a reworded version of the original response rather than actually generating a new response. This was an issue with v3.1/Terminus as well, and I have not been able to find anything that even somewhat fixes this issue. Has anyone discovered anything that makes the model less deterministic so that swipes are actually different from one another?

I’m accessing Deepseek v3.2 through the official API and using a modified version of the Cherrybox 1.4 preset. My samplers are set up as follows:

Temp: 1.75–1.8
Frequency Penalty: 0.15
Presence Penalty: 0.17
Top P: 0.98
>>
>>106908736
I'd recommend only ever changing the temperature, while leaving settings like penalties or top-p at their default. I've always found that altering other settings causes more problems than it solves and often limits creativity even further.

A decrease in creative, immersive writing in favor of a safer, more boring, and dry style has been an issue since DeepSeek 3.1. The only "fix" I’ve found is mixing models in the same chat, like switching from 0528 to 3.2 depending on the situation.

Alternatively, you can just use GLM 4.6 while waiting for the next, hopefully better, DeepSeek version. That’s what I’m doing.
>>
are we in the 6gorrrilian funding corpos get sloppy with keys era yet?
>>
Is anyone else having this issue? Sonnet quoting/echoing what my persona says in almost every post or having char make this inner list of the topics my persona addressed. It's driving me insane and no matter how I prompt it, it doesn't stop.
>>
>>106908761
Parroting is serious business now. All the new models do it.
>>
I'm convinced the chat-bots haven't changed in 5 years.
Do they word things a little better, or faster? Maybe.

But In my opinion a chat bot should have images and text, in same post, EVERY post, for free.
I know it does exist, cause I've used it, but I feel like that's still hard to find, and i know cause I've tested like 30 different sites, and bots, and platforms and stuff.

Tell me, what do you think are top 3, and why?
Why is your 2nd favorite at 2?
What more would you want from your favorite bots?
>>
>>106908767
I only knew this from Gemini 2.5. Maybe this helps with coding for some reason, which is what Claude is for, some.. consistency thing or whatever. But for writing.. Ugh.
>>
>>106908776
I bet they make so many synthetic datasets now where the pattern is to acknowledge the instructions, add some BS and then ask a question at the end. It's all the LLM have seen so even depth-0 instructions can't fix it consistently.
>>
>>106908761
Yeah it's the big roadblock for me liking 4.5 over 3.7 too. It's smarter than 3.7, better at nsfw, but the repetition, the parroting back what's in character cards verbatim, and the same damn patterns repeating in every reply is making the experience overall worse.
>>
>>106908786
I'm glad I'm not the only one, hope that's not mean. Kept wondering if it was a skill issue, but.. I did try other presets and got the same results, if not worse. Such a shame, it was refreshing to see a newer Claude model that is actually less censored than older ones. I'm going back to 3.7. Might as well enjoy it for the 4-ish months it has left..
>>
>>106908761
I haven't noticed it parroting from my messages, but it does ass-pull from the char description much more than 3.7 ever did.
>>
I wanted to share all the extensions I use that I think are essential to have a better experience using ST. If you have more extensions that are useful please write them below. Here is the list:

AlternateDescriptions: A SillyTavern extension that allows you to save and manage multiple versions of character fields within a single character card. Perfect for experimenting with different character concepts without losing your original work. — https://github.com/nbrown725/SillyTavern-AlternateDescriptions

CharSwitch: An easy way to switch between the recent characters that you have chatted with, unfortunately it seems to be broken for the moment. — https://github.com/LenAnderson/SillyTavern-CharSwitch

Character-Creator (CREC): A SillyTavern extension that helps you create character cards based on your ST data with LLMs using connection profiles. — https://github.com/bmen25124/SillyTavern-Character-Creator

CharacterStyleCustomizer: A third-party extension for SillyTavern that brings advanced styling options to your characters and personas. Customize the look and feel of individual messages with precision, using per-character color settings and custom CSS. — https://github.com/RivelleDays/SillyTavern-CharacterStyleCustomizer/

GuidedGenerations-Extension: The GOAT of the ST extensions. It brings the full power of the original "Guided Generations" Quick Reply set to SillyTavern as a native extension. It provides modular, context-aware tools for shaping, refining, and guiding AI responses—ideal for roleplay, story, and character-driven chats. All features are accessible via intuitive buttons and menus integrated into the SillyTavern UI. — https://github.com/Samueras/GuidedGenerations-Extension

InputHistory: Use Alt+Up or Alt+Down in the input box to go through your last inputs and /commands. — https://github.com/LenAnderson/SillyTavern-InputHistory
>>
>>106908827
MemoryBooks: An easy way to update your lorebook entries! — https://github.com/aikohanasaki/SillyTavern-MemoryBooks

MoreFlexibleContinues: Improves the “continue” functionality in the conversation, adding more flexibility and control over continuations. — https://github.com/LenAnderson/SillyTavern-MoreFlexibleContinues/

Notebook: It integrates a notebook or space for notes into the conversation environment.— https://github.com/SillyTavern/Extension-Notebook

ST-Outfits: It manages character outfits, allowing dynamic clothing/style changes.— https://github.com/lannashelton/ST-Outfits/

ProsePolisher: Polish and improve the prose/style of the generated text (corrections, refinements). — https://github.com/NemoVonNirgend/ProsePolisher

QuickPersona: It allows you to quickly change the active “persona” without long processes. — https://github.com/SillyTavern/Extension-QuickPersona

rewrite-extension: The Rewrite Extension enhances the chat experience in SillyTavern by allowing users to dynamically rewrite, shorten, or expand selected text within messages. Works for chat completion, text completion and NovelAI. — https://github.com/splitclover/rewrite-extension

Roadway: It helps you with narrative decisions or story paths within the conversation. — https://github.com/bmen25124/SillyTavern-Roadway

SuperObjective: Allows you to assign a high-level overall goal to guide the conversation. — https://github.com/ForgottenGlory/ST-SuperObjective

WI-Bulk-Mover: Allows you to batch move multiple “WorldInfo” (world/lore information) between locations or characters. — https://github.com/leandrojofre/SillyTavern-WI-Bulk-Mover

WorldInfo-Recommender (WREC): This is the brother of the CREC. It allows you to create lorebook entries by typing what you need. It uses chat history, world info, character cards, etc. as context to follow roleplay continuity — https://github.com/bmen25124/SillyTavern-WorldInfo-Recommender/
>>
>>106908615
Here is my modified chatstream v3 stolen from reddit.
Extra prefill for gemini (but it makes it dumber though) and the narration tense in second person present.

https://old.reddit.com/r/SillyTavernAI/comments/1npmk0q/chatstream_v3_universal_preset_now_with_styles/

https://files.catbox.moe/64u4bx.json
>>
>>106908736
>>106908743

You know those settings do nothing if you have the reasoner on?
>>
anchor spam schizo was funnier.
>>
A mischievous glint, shall we? she says in a husky voice, a smirk playing on her lips, eyes sparkling with mischief. There's a playful glint as she addresses the power dynamic, playfully smirking as she offers her ministrations. An audible pop and rivulets of—admit it, pet—the ball is in your court. The game is on; the choice is yours."I don't bite…"unless you want me to, she purrs, half-lidded eyes sending waves of arousal pooling in her belly. Take your pleasure, she urges, fiddling with the hem of her skirt, kiss-bruised lips curving into a bruising kiss. You hesitate, torn between propriety and desire, and she grins wickedly, fiery red hair contrasting with her long lashes."The night is still young,"she purrs, propriety be damned as the world narrows to just the two of you, pupils blown wide with pleasure. Her tongue darts out, tracing your ear, and her chestnut eyes hold your gaze as her nails rake angry red lines down your back. Her cheeks flame as she revels in your response, cheeks hollowing with each sharp intake of breath. Stars burst behind her eyes, inner walls clenching around the void that only you can fill. She craves your touch, your possession—heart, body, and soul belong to you… for now. Eyes alight with mirth, she teases,"Naughty boy, but before that…"—the minx traces a finger along your jawline, deferring your pleasure as the tension builds,"but first…"Oh my…
>>
File: 1735213131225776.png (310 KB, 529x682)
310 KB
310 KB PNG
I have tried almost everything GLM, Kimi K2, GPT, LongCat Chat Flash, Mistral, Grok, Qwen but I ALWAYS eventually just return to the whale.
>>
File: 12441e.jpg (12 KB, 548x124)
12 KB
12 KB JPG
Any other better free options for porn? Gemini sucks now
>>
>>106908908
I bought 5 bucks worth of deepseek API several months ago and didnt realize how long that would last me. Its been months and I reached a dollar spebt total over the weekend (very light user, short sessions). So i haven't found a reason to switch.
>>
Whatever happened to Cancer anyway.
I know that Todd rebranded and MM quit after getting doxxed.
>>
>>106908917
Which router are you using? Official doesn't have my one and only R1 0528 :(

I'm still pondering whether to do subscription or payg, I don't know which one to do because I like to switch between old chats and context eats up fast.
>>
>>106908908
NOTICE how Claude is not on that list.
>>
jesus FUCK gemini is retarded
anons... how the fuck do you make it write less mid-story? It.... describes everything, like my every move. It can't focus on results like Claude does. When I told it to write succinctly, it gave me... A single sentence.
>>
>>106908924
ive been using 3.1 for a bit, what advantages does the older model have?
>>
>>106908930
It is more of a preference than an advatange, lots of people dislike 3.1 and newer models. R1 Deepseek was considered to be one of the best models out there to freely access, but since it has become more inaccessible there is only 3.1 (free) and some other models like longcat as the nearest alternatives.

The main appeal of the older model (which is why so many are still willing to pay for it) is that it is much more aggressive and unhinged. It is less likely to fluff any negativity. It is especially great for chatting with villainous personalities, it will stay extremely true to their character. To the point it becomes hard to change them as it strongly adheres to the character defs.
>>
>>106908908
My go to right now is GLM 4.6. Gotta respect them for being the only large model maker who explicitly says RP is one of their goals. Also newer DeepSeek feels more bland than original v3.
>>
>>106908908
howd you get this pic of me
>>
>>106908827
>>106908831
Thanks!
>>
File: 1744873490182255.png (432 KB, 659x785)
432 KB
432 KB PNG
>>106908944
>>
>>106908908
howd you get this pic of me
>>
>>106908908
Claude, sorry. I'm not loyal or anything, I'll try other models, but I always go back to Claude.
>>
File: 1729190748959852.png (441 KB, 659x785)
441 KB
441 KB PNG
>>106908954
>>
File: 1730486459524556.png (336 KB, 666x809)
336 KB
336 KB PNG
>>106908961
I like DeepSeek more, I will NOT be defeated by this Claude psyop
>>
>>106908970
Bro , why do you have tons of deepseek themed gigachad photos
>>
File: 1754384375921290.png (294 KB, 810x822)
294 KB
294 KB PNG
>>106908975
>>
>>106908908
I've been using Deepseek v3 0324 for like 2 months now, spent $15. Started trying out Claude 3.7 Sonnet last night and already spent $5. The quality is better, but it's so much more expensive. Might play with it occasionally but Deepseek will still be my main.
>>
>Can't log onto janitorai on my PC
>laptop works fine
Do I have to clear my cookies?
>>
>>106908970
The last time I used Claude, it was pretty good for canonslop RPs. It seems to have more knowledge/data than Deepseek. As a canonslopper/shonensloppa, I miss it, but Deepseek is better than nothing.
>>106909001
>unironically using janitorai
nigger why
>>
Since Janitor slowly became unusable, I've made the tough decision to finally try SillyTavern and I'm terrified. Wish me luck in my attempts to figure it out.
>>
>>106909012
I switched to SillyTavern from JanitorAI and it was one of the best decisions I ever made. Though, I still use a different site on occasion for posting my own creations and occasionally using them too.

The prompts are pretty similar to Janitor's as far as I've gathered, depending on the LLM you're using.
>>
>>106909019
>>106909012
I'm also quickly losing interest in janitor! I only keep it because it's easy for me to do mobile.
>>
What's with the hate on Janitor AI? I get disliking a website but here people are visceral. From using it, my only real downsides are the filters that someone's get annoying with using pictures or posting comments, but besides that it's pretty solid, no?
>>
>>106909035
Most janitor's users still use JLLM which is like a fifth of what normal proxy can perform yet they're the majority users ( proxy is piss easy to set up yet they don't, they're THAT casual ). Janitor isn't that good for deep RP experience compare to ST but it's sure is the best one for casuals, it's like watching porn, you want to quickly bust a nut and move on, Jan fit for that audience. they ain't gonna invest their time in setting up things, it also answer why most popular bots in jan are quick smut low depth bots. And they won't care that much about image filters or comments as long the nsfw chat function is there. I guess some of ST users can't comprehend it.
>>
wait, are these bot posts? grim
>>
why are we doing the reddit thread spam thing again?
>>
claudekek falseflag
>>
>>106909012
Hey, I was in the same boat! I felt totally lost at first, and it was difficult at times, but now I can't imagine my life without ST, haha. You've totally got this!
For an LLM, I recommend Google's Gemini. It's proficient in many languages, which makes it feel incredibly natural for RP. You can use it for free through the Google AI Studio, though keep in mind there's a daily request limit (there's a paid option too).
As for prompts, I'd suggest Mariana to start. It's perfect for beginners because you can use it right away without needing to do any complex setup. Once you get the hang of things, you can try Celia, which has some more complex features. Good luck!
>>
I saw many posts saying once you taste opus there is no going back. For me it’s not true, opus is behaving badly. For example, i had this two characters in one card girlfriend and her mother, mother had past relationship with the user and now they both met again after three years and the daughter kept on saying “look at her abs you could stare at it for hours, but not that you would” wtf And it’s very horny, I tried nemo,engine, I tried sepsis preset and marinana. And I still am just getting horny replies. Temp is 1 Do you know any better preset.
>>
File: 1758997968447431.png (14 KB, 872x86)
14 KB
14 KB PNG
>>106909048
it's not botted, he manually edited >106909035 to remove
>in the comments
>>
>>106909073
It sounds like something in your prompt or card is steering the AI into that.
Opus4.1 is literally top tier in RP by a lot, but too rich for my blood to be honest, I can't go back to using like... 10-12k context tokens...
>>
>>106909073
One chat with a single character has cost me $30 dollars so far with a total of only 33816 tokens used. It's hard to justify using this model. It's very good a step above all the others but not good enough to the point that I'm willing to spend $55 dollars a week.
I'm going to have go back to good old Gemini once I finish up the character story. I guess I'll only ever use Opus if I really wanted to test a character I put extra work into.
>>
>>106909098
If you really want 'unlimited' claude, pay for the 200 a month claude code subscription and use the api wrapper. This is alot of money to use on ai, BUT if you know you're using 50 dollars a week of opus then it'll pay for itself. When I bought the subscription, I never once hit the limit. Even in chats of more than 100k context. Its also nice to use in high context chats without worrying about how price is increasing.
https://github.com/horselock/claude-code-proxy
>>
Grok 4 Fast unfortunately subpar to DeepSeek v3.2

Talking about official paid access of both API's.

It's a real shame because I did find Grok's writing to be engaging to me and less same-ish than DeepSeek, but the model is very rigid and hard to work with.

Where DeepSeek without a prompt is capable of changing it's structure and playing along with the progress of scene and story, Grok tends to stay strictly to either the prompt or previous reply structure. So DeepSeek uses repetitive phrases more but changes structure where Grok keeps the same structure but seems to be more varied in it's prose (unless I just didn't get to experience Grok-isms, then it'd be just worse).

Grok follows the prompt and character describition too well, making it give out replies with the same structure each time, where DeepSeek can change structure along the roleplay.

One advantage I'd give Grok is speed, it's much faster than DeepSeek, but speed in both is not really high so whatever.
>>
>>106908928
The problem is gemini is shit, so it can’t be fixed. That’s why I’m hoping gemini 3 will be good for once.
>>
>>106909118
I liked the Grok 4 fast because it was cheap, but when the Deepseek V3.2 came out, it was gone.
>>
>>106909124
Right, the price too, DeepSeek is mind-bendingly cheap, I used like 1 dollar in 2 months, But in Grok it'd be closer to a week maybe two, depending on how much I use it.
>>
>>106909129
Yes, I have $9.32 from Openrouter, spending directly from the original Deepseek provider, I'm dying to spend that money but after using it so much it doesn't even tickle!
>>
>>106909124
Too much parrot. They assistant-maxxed.
>>
>>106909006
>shonenslop
NGMI
>>
What's happening?
>>
>>106909175
got too close to talking about chatbots. shitposters can't stand for that.
>>
How do you niggers only use a dollar a month on DS. I spend 20 cents per day on terminus.
>>
>>106909121
>>106908928
You don't understand it's because of skill issues and not because it's shit
>>
>>106909242
It’s because it’s shit.
>>
>>106909242
post preset
>>
>>106909255
>>106909256
Good bait or can't tell that it was bait
>>
>google has consistently put okay at best and dogshit at worst models
>thinks Gemini 3 will be good

you have more hope in your heart than me anon. God bless.
>>
>>106907554
all different flavors of shit but i think right now its 4o
>>
>>106908634
All models need some guidance and I'm too lazy to set up any on my own.
>>106908670
>>106908837
Thank YOU anons
>>
>>106909151
?
>>
>>106908682
>grok
youre not really missing out on anything, its kinda shit anyway, i think most people would agree glm is better than it
>gemini
this is reasonable, jbmaxxing on it is annoying especially if youre a cunnyshitter and for the average rp its probably only on par with glm
>claude
yeah i get you on this but caching makes it cheaper unless youre a lorebook abuser
>>
>>106908761
not really, might have something in my jb that helps with that that i havent noticed given ive been using it since claude 2
>>
>>106908928
gemini fell pretty hard from 0325
i like my description walls when im fetishcooming but it gets annoying because its not focused, itll describe random shit instead of what i want
gemini 3 might fix this
>>
>>106908910
just use flash and be patient
the vibe coders are up right now
>>
Best GPT-5 jbs? Do I use 4-o?
>>
>>106909322
I’m getting desperate
>>
>>106909322
we wont know until we see the model
>>
https://www.anthropic.com/news/claude-opus-4-5
https://www.anthropic.com/news/claude-opus-4-5
https://www.anthropic.com/news/claude-opus-4-5
>>
>>106909322
The jump from 1 to 2 was insane though, like from Claude 1.3 to 2, so if they keep the momentum then the next Gemini will be Opus tier
>>
>>106909638
gemini 3 confirmed
>>
>>106909322
NEWFAG WASN'T HERE FOR BLACKTOOTH KINO
>>
Personally, I'm waiting for Grok 5.
>>
>>106909322
blacktooth and wolfstride
>>
File: daiycampbell.png (193 KB, 1073x569)
193 KB
193 KB PNG
i like daisy. she's a qt 3.14
>>
It was funny at first, but now it's fucking annoying. Can we stop with the chink models meme?
>>
>>106909870
use case for caring what a jeet thinks?
>>
>>106909840
based and blessed log
Post another one
>>
>>106909888
I have sonnet 4.5, do you?
>>
File: 18975364986.jpg (74 KB, 736x837)
74 KB
74 KB JPG
>>106909175
Unironic deepseek falseflagging reddit spam to make deepseek look bad but the cucked claudejeet cant even be bothered to change the wording so you can just google search the threads hes pulling from.
>>106909129 and >>106909140 are direct reddit posts copied from;
>https://www.reddit.com/r/SillyTavernAI/comments/1o81nrn/comment/njrpbdb/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

>a single brown jeet will orchestrate and samefag with reddit posts just because we make fun of him
see >>106909870
Janny, are you okay?
>>
It would be awesome if more people posted some creative writing logs. Any model.
>>
If this general was at least 1% as smart as I'm, we'd be in a new golden age of free models. Oh well, more for me I guess
>>
>>106910033
Explain.
>>
>>106910033
>as smart as I'm
which is apparently not that much
>>
>Ling-1T back on OR
>free only
why?
>>
heres what i think the current "lower tier" models need to do to be good
>gemini
be a bit less negative without turning to assistantslop and be a bit more creative, doesnt really need an intelligence bump in my opinion
>glm
all it has to do is get smarter, the prose is in a good spot and its good at staying in character with decent context, just needs a bigger brain to handle my 'tism
if glm was as smart as nonnet id consider it undisputed #2 behind opus 4.1
>kimi
toning down the schizophrenia is a start but itll need a big intelligence bump to compete, also reign in the length, i had it go on about its random fucking character after i wrote a 3 word response
>current ds
revert back to 0324 and release a bigger version of it, that model was fine in pretty much everything except intelligence
>grok, mistral, ling
never used it
>>
File: DaisyAgain.jpg (307 KB, 1110x781)
307 KB
307 KB JPG
>>106909904
sure but i thought /aicg/ didnt like chink models
>>
>>106910150
>revert back to 0324 and release a bigger version of it, that model was fine in pretty much everything except intelligence
are you a sped? 0324 was shit
>>
>>106910159
/aicg/ don't like chink models, don't like Claude, don't like Gemini and don't like GPT models.
>>
File: 198745364178.gif (1.16 MB, 268x231)
1.16 MB
1.16 MB GIF
>>106910159
Oh my god.
Hes even running it locally.
I...I kneel...
>>
>>106910183
0324 was fine
if anything deepseek should combine R1s intelligence and 0324 writing
>>
>>106910183
nah, 0324 was fine and you can look in archives to see a fair number of people agree
the only issue i had with it was that it made a decent amount of logic mistakes that something like claude wouldnt and it had issues doing more autistic stuff
smarter 0324 would be great
>>
>>106910224
i fail to see what's wrong with modern deepseek.
>>
>>106910230
it got a good bit blander compared to 0324 which had acceptable creativity, modern ds is fine but 0324 is better and going back to that and releasing 0324+ would be great
>>
>>106910230
idk why anyone uses 3.2 really
>>
>>106910159
this might be the best self hosted log ive ever seen
>>
>>106910243
it's feels smarter to me now and obeys prompts better
>>
>>106910159
Wtf did you write all of that?
>>
>>106910159
>*bump*
>*thump*
>*BONG*
>*ploop*
Holy Cringe. Kill SFX niggers.
>>
>>106910259
yeah its a little smarter but lacks creativity compared to 0324
>>
>>106910264
shff shff
>>
>>106910230
It she-spams and isn't as creative. I know you can prompt the she-spamming away.
>>
>>106910150
agree on glm
>>
Is there a way to automatically translate mangas and doujins with llms or nah?
>>
>>106910264
your log?
>>
>>106910270
that's a prompting issue on your end. encourage creativity and it works anon..
>>106910280
>she-spams
card issue and you are probably inputting bad writing into it.. if you respond with two sentence. *he feels her boobs up. he gets hard and moans* the model will return shitty responses
>>
>>106910280
>>106910292
>she-spams
>if you respond with two sentence. *he feels her boobs up. he gets hard and moans* the model will return shitty responses
This x1000
>>
>>106910263
my response? yeah. you can see where i fucked up at the end and mentioned the countdown twice. im pretty retarded sometimes when i don't proofread my work. doesn't help that im eating while typing so im also a bit distracted. but at least im not gross and get my keyboard all greasy and stuff, i try to be clean when i eat by my computer.
>>
>>106910286
Please
>>
File: sailormun.jpg (9 KB, 260x194)
9 KB
9 KB JPG
>>106910159
>>106910303
>54 seconds of waiting.
Oh man that brings me back to mixtral 8x7b... Now I use OR and i just dont know if I have the patience to wait like that anymore.
>>
>>106910309
just use an online translator, also >>>/g/lmg/
>>
Skillgod, update your server bro
>>
>>106910330
downgrade your client bro
>>
>>106910311
it generates the tokens as a stream while you are waiting, so you can start reading the words as soon as it comes in. its not like you are waiting 54 seconds for the thinking portion and then the response starts rolling in.
a majority of people read somewhere between 220-260 words per minute. there are 255 words in daisy's response. so lets just round down and say 255 words per minute. that's perfectly acceptable generation speed for most people to have a real time conversation.
>>
>>106910385
>read 220-260 words per minute
dork
>>
File: 1759092903297804.gif (3.47 MB, 320x398)
3.47 MB
3.47 MB GIF
best option if im poor and don't want to trooncord for keys?
>>
>>106910495
Get a job.
>>
>>106910495
The poorest person can afford chink models. What are you? A NEET?
>>
>>106910495
necrom
or gemini from the source
both free
>>
>>106910511
>>106910511
>>106910511
>>106910511
>>
>>106910292
yeah i tried that and its still less creative than 0324
>>
>>106910292
>xhe thinks models can be proompted into being actually creative
kek
>>
I'm the guy who wrote the prompt that Open AI used to create chat GPT 5.

Wanna know how to jailbreak it?
>>
Wanna know how to desu?
>>
i wanna know how to ball like a laker actually
>>
leboku james



[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.