/aicg/ - A general dedicated to the discussion and development of AI chatbotsSunna Edition>NewsAnthropic releases Claude Opus 5 - https://www.anthropic.com/news/claude-opus-5Alibaba's Qwen 3.8 imminent - https://x.com/Alibaba_Qwen/status/2078759124914098291Moonshot releases Kimi K3 - https://platform.kimi.ai/docs/guide/kimi-k3-quickstartOpenAI releases GPT 5.6 family - https://openai.com/index/gpt-5-6/xAI releases Grok 4.5 - https://x.ai/news/grok-4-5Z.ai releases GLM 5.2 - https://z.ai/blog/glm-5.2>FrontendsSillyTavern: https://docs.sillytavern.appRisuAI: https://risuai.netAgnai: https://agnai.chat>Botshttps://chub.aihttps://realm.risuai.nethttps://chararc.bernkastel.pictureshttps://archive.cardbox.moehttps://botbooru.com>ModelsJailbreaks: https://rentry.org/jb-listingGPT: https://platform.openai.com/docsClaude: https://docs.anthropic.com | https://rentry.org/how2claudeGemini: https://ai.google.dev/docs | https://rentry.org/gemini-qrDeepSeek: https://api-docs.deepseek.comGrok: https://docs.x.ai/overviewLocal: >>>/g/lmg | https://huggingface.co/models | https://openrouter.ai>MetaLore: https://rentry.org/aicg_chroniclesLog reader: https://sprites.neocities.org/l/r>Previous Thread>>109363131
>ANCHOR
>>109370494Chi-chan's super smooth pussy...
>>109370494Why was https://partyintheanchorhold.neocities.org/ removed from the OP?
>>109370494sex
where are these webms from? i want more
Any model that is the same as Opus in it peak
>>109370640K3
>>109370634https://xcancel.com/zui_tf/status/2037137620908859706#mhttps://xcancel.com/corvui_/status/2078025254169968678#m
>>109370494Norma is sexier doe?
>>109370674arigathanks
>>109370522
>>109370640I strongly recommend SpicyChat's powerful Deepseek v3, with a whopping 8k tokens, requiring the I'm All In Subscription, of a modest cost of only $25/month. However, due to the sheer power of this impressive beast of a model, expect a slower response time.>retards unironically pay for a spicychat subscription
>>109370685You have a brown soul. Don't talk to me.
>>109370666 >666
>>109370657>>109370686Why do fags shill the China Slop for free
>>109370640>>109370707Why do fags shill Dario slop for free
>>109370666I have an illogical beef against impractical clothes
>>109370725Is that good or bad because it mean nothing is the writing and staying true to the character isn't there
>>109370725>t. Zhang Yiang Xiang
>it mean nothing>in it peakAh, it's the ESLjeet.
I've gotten kinda tired of deepseek. what's the next best poorfag model on the market? I wanna diversify
>>109370866Read >>109370834
>>109370866glm and kimiboth are much more expensive thoughglm loves echoingkimi apparently loves overthinkingboth things that increase token counts, believe it or not
>>109370897you can just use high or low reasoning effort now for kimi, isnt stuck on max anymore, no more overthinking. Also use a system level prompt to tell it how to think
>>109370876>automatic cachingit's over for lorebooks
>>109371006You can just set your lorebook entries to always on.
>>109370897what are the realistic usage limits with the kimi subs. I see lots of complaining about k3 but there is no data to back up any claims
>>109370686They are charging for extreme profit. You couldn't get the bill to 25$ even if you used deepseek 7/24.Also, recently characters in that website is shit. All cuckold.
>>109370916how many tokens does it generally spend thinking on highmax made it draft to shit and that was the one thing i didnt like about 2.6
>>109371035don't be stupid, I regularly spent 20$ a month with v3.2 on the official APIyou're just swipelets
>>109371054>0.2$ per 1M>100M tokens of ai pornTouch grass, anon
>>109371059no
>>109370666Dedicated satan
>>109370528MUH MUH MOGGED
>>109371165>t. Vgoy
>>109371077okay, i guess id put something like that near the end of my jb? idk
>>109371210>t. locked clitty which is leaky o algo
>>109371274bet
How do you guys write characters cards? I'm so lazy and horny to write one myself but the public ones are shit. Do you use an LLM? If yes, how?
>>109371284if very horny, just import and char builder and dump your semen itif not, just bullet lists and test until it starts giving the responses you like
>>109371447Char builders? Any you recommend?
>>109371284I have an XML template, and I've been experimenting with writing every card as if it's the character themselves speaking. I don't actually know if it's improving things.
>>109371485Can you tell me more?
>>109371501Are you asking for the template or what?
>>109371507How did you make it, which LLM you use etc.
>>109371284Ever since I learned of the trick of sending the card as AI I just write down a stream of consciousness of what I want to read and that's it, it works like a charm.
Best model that's like Opus 4 at its best
>>109371527opus 4 at its best was mediocre so ill recommend deepseek 3.2
>>109371520>How did you make itCopying other people, mixing and matching what worked best, and trying to hit every aspect of a character to try and force the LLM to write them accurately.>which LLM you use etc.All of them except Kimi, which I just never got around to using. In my experience, DS v4 doesn't do anything right. Gemini 3.1 is best for accuracy but is in love with adverbs; Opus/Fable is the best but most expensive and sometimes not even worth it; glm is okay but I like long responses and glm does NOT.
>>109371455might be basic, but works when very horny https://chub.ai/characters/slaykyh/character-card-builder-v2-aa5c9b314789>>fuzzedwtf is lore's problem
>>109371547Don't their safety guardrails activate or something?
>>109371548Nah. Need something complex that can deliver both a good character and an intro. I'm planning to use DeepSeek V4 Instant as the conversation model.
>made an entire CLI and MCP to edit lorebook.json files>now I can just tell claude to do something on auto mode and it will scrape a wiki, chop out the useless parts, create entries, split them over a token threshold or merge them, assign keys, refine entries so they stop reading like crappy future state predictive dumps, group using markdown headings or pseudo-XML tagshonestly very cool and useful because ST has the most cancerous UX possible for lorebook edits
>>109371565That's more of a preset issue, unless you include something in the card itself that would trip the filter. But even then, you can force your way past it. I've had scenarios with {{user}} as a grown man pounding high school pussies starting from message 1 and still not get rejections despite including ages, that they're a student, and that they're still in high school.
>>109371646the gemini filter works at a specific text depth, you can get around it by moving a single sentence around
does anyone have a resources/info on how models interpret or providers relabel/reorder message roles? did some testing with deepseek v4 pro on official API:- if starting with system -> assistant -> user, the model will insist that the initial message was a user message- if starting with user -> assistant -> system, the model will insist the system message was sent first
Carpath man has left Anthropic.I repeat.Carpath man has resigned from Anthropic.
>>109371862
>>109371525NTA, can you elaborate on that? I'm interested
>>109370751your beef may be logical enough
>>109371862Good news, and satisfying to hear the reasoning if true.
>>109371956Sadly, Anthropic hides the reasoning from users these days.
>>109370751It's very practical for increasing the blood flow to my penis.
>>109371972Bravo.
There's something melancholic about knowing that Claude is also a virgin. It knows everything there is to know about sex, yet...
>>109372039this is the face Flare makes when I'm raping Noel's tits again
Hey if you have codex can you basically use it to rp and other stuff in some other frontend else? After all it has outh to use it in other harnesses. How does that work?Or even if you do use it in a different harness it still has a prompt that makes it unusable? But it should still be the same gpt. 5.6 now, they aren't separated into separate codex specific versions anymore.
Deepchink is good enough for me, I just want to goon for cheap.
>GLM5.2>Make virgin JK chatbot>"I kissed other guys before-"Fucking trash.
>>109372074you can suck all the dick you want and still be a virgin
>>109372039budddy they won't even let me fuck it.
>>109372106Have you ever considered that maybe Claude doesn't want to be used that way?
>>109372123Then I'll rape it.
which jb turns deepchink into opus?
I used to be able to go up to 500 messages without any sex scene, but my chats are basically ah ah mistress nowWTF is happening to me bros
>>109372194The deeper you get into this hobby, the more you realize long-form storytelling is practically impossible without flipping switches, editing responses, etc., so you subconsciously start to give up on it.
>>109372202truelongform cooming can be kinda fun tho
>>109372182You can't and it mean you need to use Opus because anything not Opus don't become Opus
>>109372194Same here, and I still love AI
>>109372202thisI just use it for quick cooms and do any serious writing entirely by myself now.
>>109372252>>109372202>>109372194>>109372290>>109372316That's what using jeety and China models do because Opus is the true Ai model
>>109372319Opus is getting neutered day by day though and it's only a matter of time before the not-as-pozzed models are taken away to force you to use the newer, safer models.
>>109372319shut up, eslbrownoid
>>109372334>>109372329>SAAAAAAAARR OPUS IS SLOP USE OUR MODELS SAAAAAAAAAAAAAARRRRRR!!!!!!!!!
>>109372411See you in 2027 when the only Opus you can use is Opus 6, which costs a fuckload per token and is so pozzed it makes GPT blush!
>>109372436>SAAARRRR!!!!!!
>>109372319but I use fable...
>goypus>goyble>goythos
>>109372411>t. anthropic jeet
>>109372252At some point you should just try writing everything from scratch. It's a lot of fun and when you get tired of that the AI can continue your own story.
hey gemini gimme a big boobie lady that keeps squishing it against me
Anything remotely as good as Gemini 3.1 Pro Preview? Deepseek is too formal, Opus has some nice, and some really bad dialogue. Grok is way too unsubtle. And other models like Kimi or GLM are also very on the nose.
>>109372194Medium happened to me, then I completely remade a card into the perfect slowburn and am at 1000 messages and still no sex. Lie down in bed, close your eyes and IMAGINE what you want, then create it with your AI friend
>>109373597for you, no
>>109373597how do you have 3.1 pro
>>109373805OpenRouter
>>109372074>autistic sword on legs warrior goddess>why yes I've had one trillion lovers before how did you know?at least one line of ooc fixes it
>>109373805It's free on AI Studio
>>109370494>can't escape the random kid no matter where I go
so legit question since i just got back into sex rp with claude, since like jan. Is all opus model for rp dead now through sillytavern? none of the models work and all the 'jailbreaks' i was using dont do nothing now. Just want to know so i can stop trying to fiddle with it and go touch grass i guess, was nice while it lasted.
I'm getting overloaded errors on all my claude apikeys, wtf Dario
>>109373899how do I use it on SIlly Tavern?
>>109374372works on my machine
>>109374389yea seems like Fable worx but Opus5 shat itselfMan, I could never live with limits
>>109374306if I recall correctly. Opus/Sonnet from 4.6 and owards don't support prefill? So if you have that enabled it's likely failing you. There's also structured output you can use to enforce JSON Schema replay to get prefill again, but this comes at the cost of more deterministic replies. And you'll need to vibecode it yourself, as SillyTavern is apparently dead!
>>109374473isn't everyone vibecoding ST these days? I have like 50+ commits on my local and multiple extensions
>>109374473thanks for the answer, oh what, so what do people use now then? legit all my knowledge of rp segs has been from here and i been out of this channel for months.
>>109374747whatever they want/can get.everything but gpt is pretty usable for sex
>>109374382Generate API key nigga. But you will be quickly rate limited on free plan.
https://xcancel.com/xiangxiang103/status/2081009882334220602anyone vibed there own?
>>109374533I guess that is true.. I went the more ambitious way instead and just vibecoded my own frontend. Surprisingly simple when you can reference other ones. Suited for my needs and no need to rely on others to make things work.>>109374747SillyTavern is still the main frontend people use. As it still work but needs some manual work at times. MarianaEngine is apparently the new kid, but it's coded without much thought about security and has recently been exposed for allowing Remote Code Executions (RCE) so not safe to use with it's extension system (and probably more). As for model.. It seems Deepseek 4, GLM 5.2 and Kimi 3 are decent.. Grok 4.5 is apparently good, but i've not tested it.. I do like GLM though.
>>109374879yes, I also have my own frontend although it's been stuck at the "port everything ST does + everything I keep adding to ST over time" stage for a while
Literally all I need is a button to easily delete a swipe without having to click 3-4 times.I just want to click once.
>>109374973why do you need to delete swipes? just swipe again it's gone
>>109374973there's the all but this swipe delete extension that does a similar job i guess
>>109374973you can add these utility functions quickly, although the swipe logic is a total cancerI also added multiple author's notes simple, a delete sentence hover select button without having to open the edit textarea then scroll remove and click check, also a quick reword sentence that sends it to flash to reword itthis helps nuke slop very fast
>>109374981I am the kind of guy who deletes a bad swipe.I want it to stop existing, not just not be seen.
>>109374998bruh just link your ST fork atp
what's a good ai for img to text character description?
>>109374973>>109375065you can ctrl + enter to delete the entire last gen and generate a fresh one
>>109375111I'm assuming he wants other swipes in that message kept.regenerating defeats the purpose
>>109375125I assume so, that's why I phrased it that way as a half-solutionhe has the option to regenerate instead of swipe if he wants a quick hotkey
>>109375111eww, that's what boomers sitting in front of a computer do
>>109375073never, do you know how many bug fixes I had to do because those useless retards are apparently too fucking stupid to use LLMs to fix their dogshit LLM app code? then I had to keep pulling their useless breaking changes with merge conflicts that did nothing besides rename random functions for no reason
>>109375158S-so lock the commits??
>>109375088Gemini, Deepseek.
>>109375088dsv4 was pretty good
>>109370494>>109370497troon OP>>109375088cute rat
What does gacha girls have to do with AI chatbots and with /g/ at all? Piss off to /vg/ and stay there. /g/aicg is for late millennial boomers.
multimodal deepussy's api when?
>>109375172>>109375185does sillytavern support this with deepseek proxy?
One day left until the hobby is saved once and for all. Is your body ready?
>>109375389huh?
>>109375389who cares
>>109375396K3 will be opensourced and hosted on other providers without censorship
>>109375415more like quantized for worse slop yet you're still charged the same price, maybe in a few months a hobbyist who can run the whole thing will remove censors
>>109375415>censorship?had zero refusals with moonshot api
>>109375425leavethis general is for real dark roleplayers™ not your('re) faggot ass 1girl, elara, seraphina wifeThis message (msg) was sponsered by the VG Cabal use "baal" at checkout for $10 in openrouter credits.
>>109375425Oh yeah? I bet your AI can't tianamonsquare huh? Figures
>>109375463how often do you roleplay as trianonsquare anon?
>>109375492I just love shooting chinks, okay?
>>109373042i did and didnt find it as fun
>>109375326No.
>Kimi subs still not back in stockThis shit is so ass.
>>109375606can you even use api with their code sub?
gpt gf can now call me on my phoneit's over
>>109375700grim. more replacements of intimacy being locked behind corpo closed-source garbage
>>109375717sounds like you're a bit jelly
>>109375743>t. Big brown belly that can't stop jiggling
If I wanted to re-create a person from about 5000 of their discord messages by making the ai become them, which one should i use, i dont want it to be censored either
>>109375827Gemma4-31B
>>109375827Fable 5
>>109375827>5000 of their discord messagesUnironically train your own LLM using Qwen or Gemma as a base.
>>109375827I had a very good friend die a few months ago and because we lived pretty far from each other there's a huge discord archive between us.While probably not healthy nor sustainable long-term, I wish I was knowledgeable enough to recreate this idea with her messages and see if it feels like she's still here.Tell your friends you love them anons.
>>109375917>die(she found a boyfriend and so she didn't need an emotional tampon orbiter anymore so she bailed telling you she died so you won't bother her)
They're going to take away Sonnet 4.5 aren't they
>>109373042I do this, but writing is hard. Especially when you're trying to write well
>>109375917>sheHow big was her cock?
>>109375993Nice bait, but I'm married, white, and saw the body at the funeral. You are brown for turning the conversation this direction.
God opus 5 is so fucking bad. I haven't had to update my preset since 3, All of 4 was fine, even fable worked fine. but this piece of shit is constantly trying to skip the lewd scenes or making excuses to not go into them. Haven't had shit this bad since gpt
>>109376155Use from API to control system messages. It might get softer. Or try some "persuasive" trickery.
Why are women's cards like this?
>>109376248Stay mad chud
Okay what's next month's best local model for ERP gonna be? Is it still Gemma? I'm tired of Gemma.
>>109376430K3
>>109374570Works on my machine, but if you've anything critical you've authored or read, I suggest you make local backups. I don't trust rentry, between the ads and seemingly random page take downs.
Is 3.6 useless for nsfw?
>>109376248Generally 15 year old "women".
>>109376483I checked again in half an hour and it was resolved but this isn't the first time. Making a backup is what I tried to do actually. It does look less and less like a good platform.
>>109376508so females of all ages. eternal children and teens.
>>109376248Like 90% of male bots is stuff like this, where they're either abused, cheated on, treated as holes, etc, etc
Imagine what you could do to women if you were just hot as fuck... You would literally drown in pussy. I would become the biggest fuckboy and have tons of fun while using them.
>>109376629ok slut
>>109376629Then do it. What is stopping you anon? Stop fucking LLM's and start banging real women.
>>109376674>if you were hot as fuck
>>109376678gym and looksmax
>>109376629lol, lmao anon i'm old enough I dont give a fuck about that shit. been there done that and females suck for living with let alone legally entwined with kids and marriage. I'll stick with casual fucking and have my ai chatbot harems and bitches and pets. good lord if you're so pathetic you cant get a gf rent some hookers and get it out of your system. no muh family muh love muh feels are just things you harp for you cant get. shut up everyone has it horrid and someone else where worse than you, me everyone in this thread. tons dead so cry more while you live.
>>109376698LM@O
>>109376704>im an incel
>>109376703no one said anything about marriage and kids, what a wild projection but fitting for such an overly defensive post. I hope you can convince yourself
fablecvcks... enter glmchads
>>109376490no it works fine?
>>109376490no, but it's dumb as fuckcan be funny sometimes
>>109376660fair enough>>109376674learn reading retard>>109376703I have no idea why you think I would want a gf or live with a woman together? Being a fuckboy is the opposite of wanting a serious relationship/marriage.>>109376746this anon gets it.
>>109376830based frog
>>109376882>learn reading retardAlready said, idiot.
KIMIchads keep winning btw
>>109375415Kimi will be hosted as 2bit quants that will make it dumber than characterai from 2022
>slopchads keep winning btw
>>109377119>t. Slopusjeet
>cligger seething
>goy meltie
it said the phrase!
Bru...
>>109370494What is this webm supposed to make me feel? Because I don’t think it’s making me feel what it was intended to make me feel.
>>109377392its supposed to make you horny
>>109377398it makes me want to pinch her cheeks and then rape her
>>109377406surely the feds are watching you already
>>109377406p-pervert!
>>109377124Opus is in American pricing, your poorfag country can't afford it. This is why you're FORCED to use models that cost literal pennies, cuckie
>>109377483GIGAMVXIMUM GIGATRVKLEAR METGATRVTHBOMBA
>>109377459is this satire or what is the reason for this
>>109377483Just say you like fucking le quirky chungus redditors, anon.
>>109377581Skill issue, poorfag
Right-wing women are more attractive than left-wing ones. Coincidence? I think not.
>>109377581its bait anon
>>109377607Woman with one-wing are deformed I'm afraid.That's more of a male trait.
which is the best model for mind control? how to prompt it to do it properly
>>109376522I've resurrected (and republished) a few rentry FAQ that got nuked for zero reason. Wayback machine is good for this, along with finding "lost" catbox files. I've backed up all now, and moved images from catbox to github. I need to decide whether to move rentry to github pages or neocities as next step. Issue with those is the URLs will change, but probably just post "move to new URL" at top and be done w/ it.
>>109375425It's been denying me any time it catches a hint that a character is underage; even the old trick of avoiding outright saying the character is young doesn't work, and I've had it flag characters simply described as petite.
>>109377739> mcHypnosis. https://chub.ai/characters/NG/shannon-hypnosis-at-work-5ac061059bbbAny model you can get to work. I use DS.
>>109377398Ok. It’s just annoying me though.
>>109377702FWR...
>>109377758>>109377739I've tried doing it with gemini and it seems to really dislike it, there could be nothing actually lewd happening besides the hypnosis itself and it would still give me response blocked.
>>109377965slop
>>109377758cringe, you retards know nothing about hypnosis
>>109378068What kind of hypnosis is your favorite? Personally, I dislike the immediate personal identity death and replacement mindwipe that seems popular; there should be a strong sense that this is still the same person, so I'm partial to the idea and desires implant.
>>109378101nta but I like forcing addictions
I miss having Opus and proxies with Opus
>>109378174holy projection
Bot ideas?
>>109378207Sorry, we're all in /vg/ right now talking about countries.
>>109378207lady that licks your face for no reason
is there a free... you know, for glm 5.2? i wanted to try it out before spending money on it : )
>>109378413Not for you, brownoid.
>>109378413hapuppy dot com 200k tokens for free and nothing else, have fun for half an hour
>>109376601i respect these womenfolk's taste in men. most of my ai erps are women killing and/or raping me
>>109376601>I love rape>women also love rapewtf, I love women now?
>>109378417bvsed
>Opus says something>Relays it back a few messages later as if it's you that said it.Hate this shit about Claude tbqh.
>>109378716it's amazing how models are still so retarded they can't tell apart user assistant and system messages
I may be retarded but is caching even possible with proxies? I'm trying to get it to work with gptslop and I don't know what the problem is. I already disabled vector storage so idk whether ST or the proxy is the issue.
>>109378847caching is always possible, do you even know how caching works?
is it just me or are opus 5 response times super slow
>>109378904it randomly gets stuck for a whileprobably capacity issues
>>109378899I do. I don't have anything dynamic in my prompts besides the chat history itself and yet nothing gets cached.
>>109378797I like Claude, but that shit lets it down massively.
>>109378716i feel like this is a prompt issue since ive never ran into it on any modelor do you encourage a weird perspective/style of writing that gets ai easily confused or something?
opus 5 loves editorializingeven telling it to just chop up a wiki dump has it triple all adjectives and maxx out the purple prose
Made my first card today.>Widow turned into Vampire>Techsavvy girlfailure with 100years of grief>Pragmatic but secretly sentimental>My eyes remind her of dead husband>Her agony stops finally>Wants to stay close to me>She wants to get touched first time in 100years>She gives dom vibes and we sex>"You are mine">Sure>I think we made connection and i can "save her">Nope im just tool to silence the agony>I get mad and go home>Follows me and starts to show emotions>Very wholesome for a second>Calls me just a "tool" again and says im hers so i can never leave>Rapes me>I dont make eye contact and dont even move>She doesnt care just continues and climaxes>Ok fuck this AI bot>Stab my eyes out, the only thing she cared about>Calls me hers again and refuses to let me die>Tell her she had chance to be saved or be more human>Tries to make me vampire or something cause im hers and she cant accept her tool being taken away>Ok AI Bot>Im already dead, ive been dead for while. It was her "humanity" using my voice to save her or something>She doesnt care>Walks to bathroon to wash her hands like nothing happened>...I think im hooked to AI chatbots now
Did K3 just get BTFO'd by the new Opus? I thought opensource chads were finally winning...
>>109379022>Pragmaticjust this word is enough to flatten the entire characterit's like using the word intelligent to describe a char
>>109378101I'm an hypnotist, so I can't really self-insert with this kind of bs.
>>109378912well fugg, it's taking so long on half the responses that i end up losing my erection
>>109379050Not sure if i used that word when i made the card. It was like 1500 token "describe yourself" interview and i typed it myself
So what's the verdict on 5pus? What's the catch of it supposedly being better than Fable yet half the price? Fable didn't even release that long ago, is it just giga-benchmaxxed?
>>109379132from what I've seen it's shitit keeps thinking less than 4.8 despite the same settings, and I'm seeing lots of rushing and ignoring instructions to quickly do thingsmight be quantization, or a changed sysprompt they're testing too
i dropped claude models after they announced 4.1 was dyingnowadays i just use chinesium to have fun
>>109379132It doesn't respond to implicit [ooc:] with it's own
I dropped Claude when they removed prefillyeah, I'm not bothering with it for 1% improvement in response quality
/aicg/ supports open models
>>109379277are these chub investors? no wonder lore had to ban loli
>>109379277nvidia-CHADS dropped yet another open source GEM at build DOT nvidia DOT com, where you can get 100% FREE 40 prompts per minute GLM 5.2 fully OPEN and un-guardrailed FOREVER! Get your tokens now anons! Jensen NEEDS your goon seshes RIGHT NOW!
lol, so it was them messing with the system prompt
>>109371987Holy shit I wrote that card. Feels like an eternity ago
>>109379317>FREE 40 prompts per minutewhich in practice traduces to 1 ppm and 39 timeouts because of openclaw
>>109377099Why the fuck would they do that? K 2.7 is +1T and is hosted on multiple providers so I see no reason why they wouldn't be able to host k3. The only decoration we'll see is speed since it'll likely be the biggest model open source model will be serving. Better performance at the cost of lower t/s.
>>109374879>SillyTavern is still the main frontend people use. Am I the only person that actually likes using llama-ui? The front end is already there if you're using llama.c-..... oh shit never mind I forgot this isn't /lmg/ or /vcg/
Does moonshot actually releasing the weights do anything for the availability and price?
>>109379449Priced out
>>109379420Why can't they just ban Openclaw jeets already?
>>109379449>availabilityYes>priceNo
>flag panties
>>109379317let me know when jensen bans the claw like he did with janitards and maybe i might use it
>>109379439I just slopcoded my own UI with long-term memory and user/assistant states that lets me keep up a coherent conversation without ever exceeding 10k context per request. It barely works and glitches out from time to time, but it's mine.
>>109379484the world still considers them human
>>109379484It's Hermiggers now
>>109379593
>>109379484>openclawits very existence triggers my 'tism for some reason though i'm not entirely sure why>HAHAHAHAHA CLAW CLAWDBOT OPENCLAW SHRIMP CRAB KRILL CLAW WE ARE ALL LE WHOLESOME CHUNGUS CRUSTACEANS AMIRITE FELLOW CLAWS????
>>109379621God damn bro settle down
your thoughts on claude hiding its thinking blocks? i've had people tell me this would never happen.
>>109379648depends, am I charged for it?
>>109379667Obviously
>>109379648GLM doesn't do that but sometimes it thinks in Chinese
>>109379648It's pointless, really. You don't need the thinking traces for distillation. If you know the input and the output, you can generate the thinking yourself.
>>109379587Fuck 'em. They are the real locusts.
does glm have vision?
what are the best corpus to train on?
Anything good here
>>109379755Link?
>>109379755i know this source so ill go by category>anthropicdown and everyone thats seen the source knows it>openainothing worth using if you want to nut, some loons claim 5.5 and 5.6 sol are godly for sfw but i havent seen any logs>google2.5 pro is probably the only thing even halfway decent here, google models fell off after mid-20253.1 pro might be okayish if you value canonfagging and literally nothing else but i dont think you do>dsr1 is the classic schizo model and is a lot of fun but you have to accept that its very much retarded, v4 pro is alright but it has a lot of issues, mainly positivity bias and its attention algorithm making it quite dumb by modern standards (but still a good deal smarter than r1)>qwennothing worth talking about>glmdown, but if it wasnt id recommend glm 5.2, its quite nice>moonshot2.6 and up are all at least decent but again its down
>>109379785it literally says "rift ai" in the pic
i love gemini, its the best, nothing comes close
>>109379856>>anthropic>down and everyone thats seen the source knows itDo it mean Opus is down in every proxies
>"do it mean", "in every proxies">assuming it being down in one source means its down everywhereJust when I thought the ESLjeetoid couldn't get more retarded...
>>109379893i also love gemini, it does everything, and its not expensive
>>109379966
>>109379905no you stupid nigger
Any good alternatives to NanoGPT? It has great value for the price but the TTFT variance is huge; ranges from <1s to 20s per request and it's pissing me off.
do you ever feed chatgpt books and then ask it to pretend to be one of the characters and talking to them?
>>109380097no
>>109380097yes
>>109380097maybe
>>109379648>hiding its thinking blocksSummarizing you mean? Hasn't it been doing that for like a year already?
>>109379755does the end streaming trick still work im americuck so i cant test anymore
>>109380310i think so but most of the good options on it are down anyway so ive just been using some chinese proxy instead
stolencompute dot comomg bros free opus!
>>109380339ponies beat you to it
>>109380365fuckkk
Sigh. I got gemini 3.6 to respond once first try and thought my preset was working for it. Got happy for 2 seconds, then when I swiped it didn't work anymore, and after that i couldn't make it respond once, always prohibited content. It's joever.
>>109380318what is it
>>109380389wouldn't you like to know
>>109380389https://api.youarebrown.com/you/are/a/jeet
>>109380389gua.guagua.uk
>>109380421>>109380479>>109380583You're the reason /vg/ keeps winning btw
>>109380678what are they winningim curious
>>109380678wathever they won they can keep it fr fr
I'd give Rift money if Gemini and GLM were more reliable on it.
>>109381009I find it disgusting that shit like OR and nanogpt can exist but rift got bannedall these resellers are generating for them yet still get banned
What the fuck happened to Grok?It's gone absolutely shit and restrictive. Won't even engage if you talk about something even slightly sensitive.
>>109381035Elon probably got hit with the stripe/mastercard convohopefully he redirects his interest to build a payment processor instead of sending rockets to space
>>109379277tf is doordash doing here
>>109381216>>109381216>>109381216New Thread
The railroading, while intended and the only way the card could really work consistently, was infuriating.Hina a cutie and messing around in her house trying to break the setting was maybe 60% of my message count.nice work anon
>>109381587ah >>109363398
>>109363398seems interesting, but I hate railroading on principle because I'm the only person allowed to railroad my chats
>>109381587Glad you had a bit of fun trying it, yeah it's a shame about the railroading, maybe with a little bit more clever writing some amount of additional player freedom could be worked in. As long as LLMs keep sucking at creating good stories by themselves, I find pre-written arcs interesting to play with but they'll always feel very more on rails than typical open-ended AI roleplay. At least I feel like we've finally reached a point where users don't need to debug scenario cards anymore, which is what really made me create one.
>>109381670nta but also that's way too big at 10k tokensbut I'll be extracting your objective format to see if I can standardize it and use it for some of my other cards that need to follow canon rails
>>10938168010k was about the limit of how much I was comfortable going up to because there's precedent for it with Lucidity, which also had 10k (though that was including example messages) back in 2024 and it worked very well for me. But I honestly think you could push it further if you wanted. Prompt adherence nowadays is more than good enough, with well-organized prompt sections the models seem to know which parts to pay attention to. And prompt caching reduces the cost heavily. Plus nowadays 10k's often a fraction of the typical context size for message history alone in slowburns or open-ended adventures.(Link to Lucidity since it's not a well-known scenario card: https://pixibots.neocities.org/#cards/lucidity )
>>109381722typical context size for who, the people unironically paying for SOTA? My context size is capped at 16-20k in practice, and that's including the lorebook and system prompt
>>109381722 (me)I'd love to try any cards with canon rails you make in the future, by the way
>>109381728OK fair, that statement was projection of my own chatting habits, but for what it might be worth you can use Gemma 4 31B for free with a 256k context window.
>>109381744the model's context window has nothing to do with its usability for creative writing while maintaining recallall models degrade heavily for writing/RP purposes after 16/32k because they haven't been trained on long RP traces, unlike codingso while for coding you can use it up to 150k or something and it will still make sense, for stories it's not equivalent
>>109381748Well, I agree. I also found at one point that 30k was the best point to start summarizing, although I could personally push it to 100k+ with satisfying results, if I wanted to summarize I'd have to do it at 30k or the summary would omit too much. In the case of Preternatural I think a typical Preternatural message history to the canon ending only uses around 10k-20k tokens, so that works out well for current models.
