Absolute Faith EditionLast Thread: >>539429438>Read First: https://i.imgur.com/lwWTTp9.png>PSA: https://rentry.org/AntiwhineMkIII▶Cloud Serviceshttps://featherless.aihttps://novelai.nethttps://openrouter.ai▶Frontendshttps://koboldai.org/cpphttps://sillytavern.comhttps://gitgud.io/mir/storypad-qthttps://github.com/lmg-anon/mikupad▶Additional Information>Chat Models — e.g. AetherRoom, Claude: >>>/vg/aicg>Local Models — e.g. Command R, Mistral, Qwen: >>>/g/lmg▶Game(s)https://store.steampowered.com/app/1889620▶Tools & ResourcesWiki: https://aids.miraheze.orgPrompts: https://prompts.forthisfeel.club/NAI Guide Collection: https://rentry.org/guide-collectionOther/Historical: https://rentry.org/aids-miscellany>AID: https://rentry.org/remember-what-they-took-from-you▶/aids/ EventsThemed Prompts: https://aids.miraheze.org/wiki/Theme_Fridays>Theme Fridays have transformed into Theme Months!>Write prompts for this month's theme: Personality Swap & War September! Prompt collection has started: >>540537237▶Model News> (09/26) NovelAI releases GLM-4.5 (Untuned Preview) for Opus: https://blog.novelai.net/text-model-release-introducing-glm-4-5-untuned-preview-for-novelai-opus-4aa866c8a0d5>(09/11) NovelAI releases Character Reference: https://files.catbox.moe/08fdu6.png>(08/21) DeepSeek releases DeepSeek-V3.1: https://x.com/deepseek_ai/status/1958417062008918312>(08/05) Anthropic releases Opus 4.1: https://www.anthropic.com/news/claude-opus-4-1>(07/21) Qwen releases Qwen3 235B A22B-2507: https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507>(07/11) Moonshot releases Kimi K2: https://huggingface.co/moonshotai/Kimi-K2-Instruct>(06/05) Google releases Gemini 2.5 Pro: https://blog.google/products/gemini/gemini-2-5-pro-latest-preview/▶Thread News>(07/04) Emacs Update: >>529978024 – https://files.catbox.moe/oovii4.el
>>540544434Sneedy thread
I wonder why they chose NAI GLM, but I'm glad they're moving up from Erato.
What's new in the hood my niggas?
>>540544671I'm curious if they did try making a homebrew MoE, and it didn't go well and they ended up settling on one of the chink ones instead
for me, it's writing like a girl so I can pull the good vectors
>>540544671GLM-4.5 is the best open weights model I've tried (for RP) so I'm not surprised they picked it. Well, I'm surprised they actually had good enough taste to pick it instead of doing something retarded. It's genuinely a really good model and probably? the least censored/restrained.
>>540545754What about claude opus 4.1?
>>540545754>the least censored/restrainedNah that's still DeepSeek 0324It's a good choice though
>>540545910I do not really care for ozone. Regardless, I am positive their fine-tune will make it punch above its weight, especially given the base model does not so far have any noticeable quirks like Erato did.
>>540546060How can I try it for free?
>>540544671Simple. Deepseek and Kimi is too big.Why do you think people don't use GLM before?Previously, even non-NAI users here only talk about Deepseek/Kimi/etc. But look, we now somehow have multiple GLM users come here to praise it.
>>540546193When they release it, they will likely renew their trial for all accounts with a current or former sub.
>been using deepseek 3.1 on OR completely for free>current prompt is 19k tokens of lorebook like infoIs there any point in returning to NAI if I don't need imagegen right now?
>>540546060...isn't this the instruct model?
>>540546269Mostly because I still prefer DeepSeek / KimiMaybe their finetune will finally draw me to GLM, but we'll have to see
>>540546386I think he meant the base they're finetuningBut yeah, it's the instruct model
>>540545484They saw the writing on the wall that in an age where triple beak models are the minimum needed for competitive model performance, their Coreweave-rented Shoggoth cluster (few hundred H100s plus whatever other GPUs they had from before) cannot possibly train anywhere quick enough. Especially while simulatenously doing inference too.Homebrew models are dead, and that's okay. Chinks put out decent stuff.
>>540546625From idle tests with the model as is, it appears better than DeepSeek and Kimi for my purposes, if nothing else. Terminus kind of dropped the ball, but I have enjoyed Kimi. That I have a current sub means I no longer have to waste my tokens rerolling with Kimi.I still think the real pudding will arrive whenever they drop their fine-tune.
>>540546386Can you stop noticing things? If NAI uses it, it must be good.
>>540546386>>540546947Since when are instruct models bad?
hand over your improved system prompt now
>>540547414Since the first one that wasn't NAI, until now.
>>540548060That has never been the case here.
so what happened, is NAI running out of funds and turk snapped out of his lowtaxian stupor?
>>540548215https://arch.b4k.dev/vg/search/text/instructshit/https://desuarchive.org/g/search/text/instructshit/https://arch.b4k.dev/vg/thread/459191694/#459264624>The only real reason NAI is still in the game is because literally nobody else finetunes models on prose and literature for the purposes of storytelling. It's all endless chat/instructshit tune of the week #435234. Honestly a little surprising how little interest there is in it.According to the last post, this thread basically considers NovelAI to be dead.
>>540548521That means you can fuck off and kill yourself. Finally, the good ending.
>>540548445>lowtaxianah, vintage
>>540548521>only forty-five and eleven resultsHow does this proof anything? Especially when many of the comments those posts receive are people arguing against whoever posted those. You can't just assert comments by randos are the sentiment of the thread...Anyone can shitpost. Do you have any actual proof that people here hated instruct?
>>540548445seems that way
>>540547719Now we have to add 'convince a robot to stop doing stupid shit' to our setup?
>>540547719When I get home from work, sheesh I haven't even cracked this bitch open yet
I haven't used NAI in what feels like decades because turk obviously gave up on it because imagegen. It feels me with a certain kind of amusement that all the smugposters are not confronted with the dude implementing an off-the-shelf instruct model. Just cancel the sub and sub to something like nanogpt man. It's cheaper, too.
>>540549413I will never give Altman my money.
>>540548719"Instructshit" is the difference between this thread and /aicg/. You don't need two generals to talk about the same type of model.
>>540549754Ah, the hand revealed, and it's... really? Just the same dumb merge shit again? Kill yourself if that's all you have the energy for nowadays.
>>540549754I fail to understand what problem you have with /aids/ existing, but if you don't care for the community, why don't you leave? Nobody is making you stick around in a thread you think doesn't deserve to exist.
STOP FIGHTING ALREADYI KEEP GETTING EXCITED ABOUT SOME FUN NEW POST AND THEN IT'S JUST YOU PEOPLE
>>540549997Okay, since you asked nicely I'll stop engaging.
>>540549601nanogpt has nothing to do with altman, it's a provider that basically gives a sub to current open source LLM models like glm, deepseek and kimi, for $8 a month. It also offers imagegen but I'm not sure which models. I know it does offer Chroma. They also provide on-demand to other models but the subscription option is extra (and a little hidden on their site)
>>540550115>Pay-Per-Prompt AI ServiceI have no interest in this.
>>540550274only their other models are pay per prompt, GLM, kimi, deepseek (and I think a few others) are unlimited.
>>540550274https://nano-gpt.com/subscription
>>540550532How many times must I tell you I have no interest in this random service you're shilling before you quit responding to me? OpenRouter is better than this shit.
>>540549965A thread with no identity doesn't deserve to exist. That's what happens when the last thing that differentiates this thread from the others disappears. GLM 4.5 can already be talked in /lmg/, /g/aicg, /vg/aicg, /chag/, /vpcai/, /wAIfu/, /2huAI/... What does /aids/ add? NovelAI's special prompt? You can mark today as the day /aids/ died.
>>540551368See you in December or February, when their fine-tune arrives, and I bet you this thread still won't be dead. Despite your desperation to kill it, lol.
>>540551368Kill yourself.
>>540551368considering that we were here long before the other AI threads, they should've all just come here instead :)
>>540548676/aids/ is the SA of AI threads.
>>540551735There's a good 10 bux joke there.
>>540551735I guess we are a bunch of goons
>>540551368...he said for the thousandth time
>>540551735dear turk,
>>540553968PINGAS
https://rentry.org/2hwnkkvd
>>540554573model? creative writing? What
the service
They have been cooking the tools (not the model) for a long time.
>>540554573we're all so jaded by what we got now but if you think about it, it's amazing that a computer wrote this
I wonder if Ephemeral Context would be useful for this model.
>>540555658>instruct model>"please write in different style thx!"Based! The most basedest $25/month I have ever paid!
>>540555925Yes, it is based.
Anyway, I was more talking about how that feature was teased back when Erato released. Which gives people a time-frame for how long they have been working on things. In other words, no, they weren't sitting on their asses and twiddling their thumbs, as certain people here deluded themselves into believing.
>>540555925I can finally stop paying that Russian guy to suck me off and switch completely to textgen.
I thought this day may never come.
>>540554573i too like to write letters to LLMs but they are usually not like this. What model is this, what prompt did you use?
>>540556318More like cum.
>>540554921>>540556359deepseek, I'm experimenting with multiprompting. I have a hard time pushing the model to go past 2000 words in a letter without kinda devolving. It seems to be the limit for deepseek for one reply in that format, no matter what I do.
>>540556210>select text>paste it to a blank story>tell the model to rewrite it>copy the result>replace the selected text with the resultA CS undergraduate can code this in an afternoon.
>>540556943What's the name of your service? Surely you created and financed one... Right? Unless you're not a CS undergraduate...
>>540556894>deepseek>>>/aigs/ This thread uses GLM.
>>540557171This thread isn't exclusive to any particular service or model. Why do you pretend otherwise?
How does it compare to Dragon? Or Calliope?
>>540551368Are those all the tabs you have open, Ctrl+Fing for any mention of NAI?Kek, you fucking loser
Now for the grueling task of transferring all the prompts and lore I've been working on the last 6 months from Sillytavern to NovelAI. I will not miss it. I hope GLM understands xianxia.
>>540557505Yes but I keep losing them among my gay porn tabs.
>>540557619it does
>>540556894it's very organic. What time is this happening in?
I tried a promptless generation with GLM 4.5 set to minimum temperature and curiously the AI outputted this with like 100.0% probability for basically every token. I wonder if this shit is some kind of training data that a fuckton of AI models were all built on and that's where the longstanding meme of characters named "Elara" comes from.
>>540558183late 19th century. The realism already falls apart at the fact that a nun would never be able to receive letters unsupervised but well. The biggest problem is that deepseek is a 21st century model and will easily slip into being a bit too observant and emotionally self aware and also using the wrong language. You can kinda suppress that, I just wish I had a model that'd handle the language of the time better.
turk, fix it now
>>540560908thank you
>>540560596>ElaraGod fucking dammit, Deepseek always uses that name for random girls. I swear if everything starts to smell of ozone...
>>540560641it's cool anon, thanks for sharing
>>540560596I'll tell you what, I just looked it up on a whim, and the "whispering woods" is always paired with ElaraIt's impossible to determine whether or not the examples predate AI, even with the "before:2020" tag
>>540562119somebody deep dived on this iirc the name elara is all over fanfiction
>>540561852I already had an ozone smell in one of mine.
>the /aids/ gang learns about post 2022 LLM slopisms
>>540545754Shouldn't it be completely uncensored if nai is using it?
Alright anons, now someone create an Asper preset for this, please, thank you. I like them RANDOM
These new slops don't have shit on the true OGs the likes of Dendrin, Count Grey, and of course my main nigga Grotesk.
>>540562386Yeah, but it's weird how it's pretty much the exact same children's story every time
>>540561852jasmine and something darker for me
>>540554573I need this read to me by a breathy, deep female voice
>>540563286Sir Kit's hands wrote this post
it's not bad
would be nice if it were faster, but my verdict is that it's pretty good
>AI brings up loli unpromptedyep I'm thinking kino's back on the menu
Just use claude.
>>540566406 (me)I'm talking about my boyfriend not the model btw.
It used to be that I could get away with really schizo scenarios, because the AI itself was schizoNow it remembers everything and holds me to accountThe higher resolution forces me to think
hmm interdesting
>>540568621Very coincidental
>>540568621>>540568885Indeed.What does AO3 think about AI, by chance?
>>540569567An anon said he got a pretty warm reception with his (probably not labeled) ai fanfiction, images included
>>540569567i imagine like every creative-adjacent space, the people who make it their business to be seen hating it the most prominently are the social barnacle sort who want to be seen as writers more than they want to write but also there's a general distaste for it among the higher performance groups
Man, I don't even know where to fucking start
>new textgen model>new south park>new beavis and buttheadlife is good
So what's the current lorebook/memory meta for the new model?
>>540570209>the dragon of westeros be want to suck and fuck for money>the dragon of westeros be want to suck and fuck for money>the dragon of westeros be want to suck and fuck for moneylife is good
>>540571067make sure you're using the tags and formatting that the system prompt is telling the model to look for
>>540571067I'm not sureI know I should try running the defaults first, but I want to tweak the system prompts abd prefills so bad, even though I have no clue how those actually work, lmao>>540571176You can also edit the system prompt/prefill under the advanced tab, so in reality you can probably make anything work
>>540571149>>540570209
Any bets on actual finetune when?
>>540572186one billing cycle plus one day
>>540572186two or three months
>>540571067The current way to try it via Open Router and stay there until NAI-tuned comes out. If you want to pay $25 a month for a UI and a model which is likely to be taken away by then, however, be my guest.
>>540572186Four score and seven years.
>>540573101sonic would hate you
How good is the untuned for straight shota mom x son erp? I'm not asking for a friend, I'm asking for ME nigga. I want to self insert shag my goofy character's fictionally hot mom, NOW!
>>540575526It's handling monster respect pretty good
GLM 4.5 model is seemingly unusable on AIR as it is now. Most responses restate information or do a CYOA multiple choice thing so far.
I see some people having this severe hatred for AI in general on some of the boards. They seriously don't know what they're missing, especially coomwise.
>>540577050How do you get it to run GLM 4.5?
>>540544434I haven't been on /aids/ in almost a year. Dropped by because I felt a disturbance in the force.THREE HUNDRED AND FIFTY-FIVE BEAKS???>355B paramatersTHREE HUNDRED AND>FIFTYTHREE HUNDRED BEAKS??The rewrite shit looks really good too, I've wanted to be able to "reroll" preexisting text for a while.After, like, two years? Textchads finally get thrown a bone? Do you have any idea how long I've been in sillytavern "provider returned an error" hell? I've been on fucking characterai after I got bored of Erato!Are we, dare I say... back?
holy fuck i just heard the newsmy penis is in for a long night
>>540579421Don't you mean a schlong night?>>540579326Don't you mean a schlong back?
The prose is kinda bland. Less slopped than Erato if you use a good instruct or memory and the like, but still. Hope their finetune can help. I can't really get purple prose. It's lilac at best, but I want some fucking ultra-violet.
>>540579326Cringe reply, cringe image. Kill yourself.Anyways, I hope Opus users can give some feedback to help with the finetuning process, I’m sure as hell am not gonna resubscribe for a untuned model. I hope that 28k context isn’t the maximum window they’re going with this model.
>>540577050That sucks, I was considering getting a sub again just for this.
>>540579261Put GLM 4.5 as your default on NAI user settings
>>540580204Probably has to do with default temp settings and the fact it's a heavily instruct model.
>>540579828Is this purple enough? Or do you require still more?I think this prefill gors a bit too hard on the commas though
>>540580405>Elara
>imagine paying $25/mo for glm
>>540580727That's what you get on a blank ass storyShit, I didn't even notice
>>540580919Go to bed, CF.
>>540580405No! I need to be cradled—devoured—by the succulent gelatine cocoon of eye-scorching Byzantium.
>>540581542Ok, so even more purple, got it
>I will avoid names that start with "El" because they are overrepresented in my training data. actually works lmao
>>540581438I can't. Putin is in there and he's got the Rainbow Dash Super Equestrian 19 inch Friendship Funhole Destroyer waiting for me.
>>540581542I'm sure I could push it farther, but this is giving me a headache as-is
>>540561852You write like someone that gets pegged by Elara.Never had that problem myself.
>>540581542>>540575526I decided to combine your things cause I thought it would make me laugh. It did. As for Anon who wanted purple, I just use what you literally wrote for the latent thing.>Purple prose, like ultra-Violet, wrapped in a succulent gelatine cocoon of eye-scorching Byzantium purple.
>>540584314>>540584545This pleases me.I bid you fine gentlemen adieu.
so how's this new model? i only use chub paid model lately which is deepseek iirc
>>540585438It's a learning experience for me because I've not messed with a real instruct model ever
Did they say 355 BEAKS?
>>540584545>mfw velveteen furnaceThat's a new one kek.
>>540546060They couldn't unslop a 70B model, they won't be able to unslop 355B one.
>>540586974Nice try, but I was talking about Deepseek with that comment.
Can I use it for AI Roguelite?
I wanna ask a genuine question.What AI doesn't have slop?
Decided I'd play around with text adventure a bit. The model is definitely more willing to do things written in the Author's Notes, and is happier to make stat boxes than previous models. Though it seemed in this case to write around sex happening as if it wasn't the important part. Might just need some playing around with style-wise. But I am happy to be able to reliably get stat boxes.
>>540588114based
>>540544671>moving up from EratoOutside of their porn finetune the value proposition of $25 for 8k context on an outdated model was insanely bad.
>>540588442it's still a bad value proposition, you can get a lot more and a lot better than glm for $25/mo nowadays. I guess Erato started being too much of a rip-off for comfort.
I'm loving the new model.Had a sexy scene with a Hostess. Then when the protagonist admitted he was starting to fall in love with her, she called him a fool, asked him if his love was so shallow and easily won.She tells him that he sees complexity in her personality, when in reality it was calculation.And when he throws cold water on that theory, she went completely mask off, admitting that what she really sold was Fantasy. (all of this across many more paragraphs)This final paragraph killed me.
>>540588893To me it seemed like they had accepted their role as providing a sort of good anime/furry image gen with some easy to use toolI'll probably give the text gen a spin once they finetune it
>>540589324I don't think the finetune will be any good. It'll either be way too porn-y to write anything just remotely serious or just fall flat altogether. These models are ridiculously difficult to finetune and erato already did not impress me.
>>540589073...Damn. She's really cutting deep on that one.
introducing GLM to novelai users is currently like introducing them to crack, the textgen side really needed this
>>540589073Damn, Marx. (I decided to have Marx write a commentary on the dynamic in this story, taking an advantage of it being a proper instruct model)
maybe i should try detective agency prompt with fluffy boss lady on this model, but i'm so lazy to write everything..
>>540589073>>540591118Holy... kino.
>>540588114You need to be kind of explicit in telling it you want sex.
>>540537237Well, here's my prompts for the theme. First "personality swap": https://prompts.forthisfeel.club/7763>It turns out that one should probably not be playing around with magic when one doesn't really know what they're doing. Well, too late now, because I now find myself in the body of my sister, and she in mine. On the day when a courtship ball is to take place at our estate and twelve various noblemen are to compete for her heart. And I just happen to know about a rather pretty noble coming in who's got his eyes on guys rather than my sister, and who I am crushing on. This is going to get complicated...This is a contrived motherfucker. This is the result of me not getting an idea right away and just thinking and thinking until something comes out, and what came out might just be the romantic comedy of the summer. Freaky Friday by way of fantasy princess and the usual courtship scenario turned on its head.
>>540537237And here's a more conventional one for "war": https://prompts.forthisfeel.club/7764>War is hell. I learned that when I witnessed my fellow soldiers committing atrocities that will continue to haunt me. But now, the army I was a part of has been shattered, its survivors scattered to the wind, with me now lost in the enemy borderlands and still dressed in the uniform of my army, terrified that I'm going to be targeted for vengeance. Wandering through a forest, I stumble across a small temple...A more simple story of atonement and of finding similarity and companionship where nations might want you divided.
Guys, I don't think I can go back to non-instruct models... It's so easy to get what I want, by simply, like, tell the model what I want. I don't think I'll resub after this.Is Deepseek better than this? Or should I just use GLM? What about Open Router or other sites?
Digging the ability to have highlighted text rewritten>start story>is used second person when I wanted first>custom and say to write in first person>same story but adjusted for first person
>>540596212deepseek's about the same, gemini's smarter but has shittier prose, claude has the best prose but it's expensive. you can use all of them on openrouter but personally i just pay for deepseek direct from their api because openrouter hosts are NIGGERS who inject their own safety prompts sometimes.
>>540596348Is there a guide to use ST as a story writing tool (I don't care about chatbot)? Or what program would you recommend?
>>540596462I tried that before and it was basically impossible. I'd say just use https://lmg-anon.github.io/mikupad/mikupad.html
>>540596643Thanks.
>>540596462I feel the best you can do with ST is try making cards to take your inputs more as an instruction and telling it to write on your behalf too. It's janky as fuck though.
>>540596462i don't know. i use sillytavern for story writing with a preset i put together for it and it works alright but i strongly dislike the interface.
I have yet to coom...I am......having fun instead?
>>540596901Do you write like you do with NAI or do you are like dungeon master?
>>540597415like with NAI, often just importing old stories. i can clean up and post the preset if you'd like it but using a different frontend is probably a better idea.
>>540597826I'll try storypad first, thank.
>>540596286Would be nice to have an undo option. Got a kino rewrite, accidentally clicked apply, and couldn't get it back
>>540588114I used text adventure heavy on Erato so I am glad they improved it
>turn random factor up>get moonrunes instead of english insteadI guess that it is random
