Welcome to Chatbot AI General #135, the thread for discussing and improving AI pony chatbots.

▶ MLP Bots
https://mlpchag.neocities.org
Spreadsheet (CAI bots + Old repository):
https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJT
SillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-Search
Bunch of Character Cards (Descriptions/No Greetings):
https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5k
CAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zip
Expression packs: https://rentry.org/ChagExpressions
!!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U

▶ How do I start?
1) Select a Frontend
2) Select an AI model
3) Select Jailbreak
4) Select bots
5) Lovemaking with AI mares!

Starting in this hobby can be confusing and difficult. If it’s your first time and you’re lost,
▶ ASK THE THREAD! ◀

Novice-to-advanced guide: https://rentry.org/onrms
https://rentry.org/sukino-findings

▶ SillyTavern (preferred frontend)
https://github.com/SillyTavern/SillyTavern
On Android: https://rentry.org/TermuxST
App that voices pony responses in ST: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB

▶ More frontends:
Risu: https://risuai.xyz
Agnai: https://agnai.chat

▶ Locals
https://rentry.org/lunarmodelexperiments
>>>/g/lmg/
Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main

▶ Jailbreaks
MLP JB: https://rentry.org/znon7vxe
More JB and guides: https://rentry.org/jb-listing
Hypebots for Tavern: https://rentry.org/pn3hb

▶ Botmaking
Editors: https://agnai.chat/editor
Guides: https://rentry.org/meta_botmaking_list
Advanced: https://rentry.org/AdvancedCardWritingTricks

▶ Miscellaneous
Scrapped/Collected rentries: rentry.org/mrhd
Chag Arts: https://drive.google.com/drive/u/2/folders/1yfajin_hB5rW_jy6WyFjRaRfDA3EUGiJ
https://rentry.org/ChagMiscellaneous
https://rentry.org/ChagArchive

Previous thread: >>42412197

**News:**
- Claude Opus 4.1 released: https://www.anthropic.com/news/claude-opus-4-1
- OpenAI releases open-models: openai.com/open-models/
- New large chinese model, Kimi K2: https://moonshotai.github.io/Kimi-K2/
- Grok 4 released on API https://docs.x.ai/docs/models/grok-4-0709

Anchor for bots, lorebooks, scenarios.
Previous >>42412199

Anchor for technical stuffs (Proxies, Presets, Updates, Models etc.)

lol lmao even

Anchor for asking for bots, lorebooks, scenarios etc.

Bot corner:

Gennikye: https://mlpchag.neocities.org/view?card=BGM/Gennikye.png

Screenshots corner:

Linky fud >>42413397
Horizon Celly smut >>42413472, >>42415058
Remove your cock from my sister >>42414238, >>42414410
Good old estrus simulator >>42414515
Came-a-lot >>42415106
Carumba >>42415756
GET PREGNANT >>42416309

Screenshots corner 2:

Tummy porn >>42420083
Mistletoeing the line >>42420611
Horizon logs >>42424487, >>42424510, >>42424530, >>42425940, >>42426093
Luna tease >>42425570
GLM 4.5 >>42431008

VN Anchor:

Last releases:
https://mlpchag.neocities.org/VN/VN.txt

Mail:
MLPAIVN@proton.me

Custom characters by anons:
Anon: https://files.catbox.moe/0u05ef.zip
Babs Seed: https://u.pone.rs/SObPVBMu.zip
Berry Punch: https://u.pone.rs/ObwnLaSF.zip
Big Mac: https://u.pone.rs/MSMzxJTV.zip
Bon Bon: https://u.pone.rs/SfKzfuwt.zip
Button Mash: https://u.pone.rs/qcoyntsw.zip
Cadance: https://u.pone.rs/wihsimgg.zip
Carrot Top: https://u.pone.rs/RDPdPAXl.zip
Cream Heart: https://u.pone.rs/fmhijron.zip
Filly Rainbow Dash: https://u.pone.rs/tdeFjNRm.zip
Flam: https://u.pone.rs/nadnujip.zip
Flim: https://u.pone.rs/stzcrmgw.zip
Gilda: https://u.pone.rs/koLWrTeV.zip
Limestone Pie: https://u.pone.rs/sPKtdXfg.zip
Marble Pie: https://u.pone.rs/jdRzGDKt.zip
Maud Pie: https://u.pone.rs/JDdsprjH.zip
Mayor Mare: https://u.pone.rs/svWGXJMC.zip
Minuette: https://u.pone.rs/cntgcbjb.zip
Moondancer: https://u.pone.rs/NjERZjTz.zip
Mrs. Cake: https://u.pone.rs/klfuctah.zip
Octavia: https://u.pone.rs/ftovmygm.zip
Roseluck: https://u.pone.rs/pfJmksUO.zip
Shining Armor: https://u.pone.rs/ibphwvnq.zip
Silver Spoon: https://files.catbox.moe/hpe34b.zip
Snowie Aura (OC): https://u.pone.rs/AZlLEjFk.zip
Starlight Glimmer: https://u.pone.rs/DeEMGCHw.zip
Tree Hugger: https://u.pone.rs/mssapono.zip
Twilight Velvet: https://u.pone.rs/yrjroasw.zip
Twist: https://u.pone.rs/bcyMNArL.zip
Vinyl Scratch: https://files.catbox.moe/z2t47w.zip

>VN logs:
Pinkie destroys the fourth wall >>42413996
Octavia >>42417513
Not for fun >>42418003
Mrs. Cake >>42418080
Throwing poo >>42421173
Cadance >>42425101
Flim & Flam >>42425812
Button Mash & Cream Heart >>42426558
Tree Hugger >>42426744

>VN logs 2:
Snack Break >>42429124
'Egghead' >>42429248
Nightmare AU >>42429751, >>42431073
She's a unicorn >>42431019

>>42431780 (OP)
The OP that was too dar(l)ing
https://files.catbox.moe/olmkbi.png

>>42431791
These days, you can get legit good AI for free and without needing any demanding hardware on your end.

Install SillyTavern:
github.com/SillyTavern/SillyTavern

Tutorial and video:
aicharactercards.com/guides/sillytavern-installation-instructions/

Then choose a model. I'd recommend starting with Gemini, but you can try Deepseek, K2 etc. for free.

rentry.org/sukino-findings#if-you-want-to-use-an-online-ai

Lovemaking with AI mares!

>>42431887

this kills the openai

>>42431956
Totally subjective and meme rankings... That I kind of agree with.

>>42431956
>AI ranked benchmark with shitty methods
Come on, we already talked about this.
Besides, if you believe this, OAI are still first with o3.

>>42431973
>That I kind of agree with.
This. They line up with what I and everyone I care about opinions are on creative writing so its the best benchmark we have on it. Though I will say GLM4.5 should get higher marks, I wonder if its because it goes shitzo at like 0.5 compared to other models.

/lmg/ did a cockbench.

I'm trying Opus 4.1, and honestly, I can't see much of a difference for RP between it and Opus 4.
Maybe on longer context the differences are more apparent, but I don't want to blow money on it, and I still don't find the difference in price justified compared to Sonnet.

>>42432034
I remember Opus and Sonnet 3 being massive improvements for RP. What causcompanies to pull out these kind of micro updates, just to say "hey, we still exist" ?

>>42432054
GPT-5 this week, and Google may also release something, so maybe Anthropic wanted to release something too.
I also believe huge improvements are just not possible anymore with the current architecture.

>>42432054
The need to keep investors appeased as they bleed out money by operating at a loss. Just like every other major company in the last decade, except Uber doesn't need to do something new to keep the cash flow coming while it actively loses money because the service it provides fills a constant niche whereas with AI it's a mad race to keep inflating the bubble before it pops.

>>42431805
If some anons tested dialogues with three pones, I'm curious to know if it worked well or if they encountered problems with the sprites.

>>42432278
Dialog worked fine for me. Sprites are a little wonky for sure, but it didn't bother me.

Does anyone know what to put as the model name for opus 6? I want to try it.

>>42432003
totally useless logprobs, this model is hopeless

>>42432371
According to ST, it's claude-opus-6-20260214.

>>42432427
Doesn’t seem to work for me. I get a message saying this model is disabled on risu.

Is there instructions on how to set up the connection for the VN? Im confused on what to put on the URL part

im retarded and dont know get to the sans proxy. I got the link and used the image to get the inside link then I dont know where to go from here.

things will get better :)

>>42432544
Files in the Docs folder for assistance
For setting up connection
>game/data/config.json
>picrel
If you're really new to LLM roleplaying in general look at >>42431829
>>42432647
Add /captcha to the end of the url

>>42431795

>>42432798
Sorry, I shouldve mentioned Im using Openrouter. I dont know where the proxy URL is for that.

>>42432836
Try this
https://openrouter.ai/api/v1

>>42432798
I beat the asshole but when I put proxy I get this

>>42432992
Check in the Docs folder, document 0.
There's already an OpenRouter example there.

>>42432836
Check in the Docs folder, document 0.
There's already an OpenRouter example there.

>page 9
What the actual fuck?

>>42431973
>>42431983
I'm curious why you think Sonnet 3.5 > 3.7? Or Deepseek over them both and Gemini?

I'm just an architect.

>>42433718
>Your program sucks dick.
My giiiirl

>>42431785
>>42432803
https://files.catbox.moe/f6ckc4.png
Here's that card I was talking about. Already got posted before in a hip and cool place, but this is the version for /chag/. No major differences beyond the obvious.

>>42431430
It's alright, anon. Just venting a little. Although stiffing the guy after he asked for more details was kind of a dick move.

Would be cool if people posted what little tweaks they do to cards. I know I have a collection of generic details I can paste in to spice a card up.

>>42433718
Model?

>>42433756
>what little tweaks they do to cards
Still trying to find the best way to make Twilight less "SCIENCE!" obsessed.
For a card, it seems that the best results are from using the Post-History Instructions field, but that can mess with some anons' presets even when using {{original}}, so I dunno.

>>42433756
Birds eat other birds. Like how we eat other mammals.

>>42433762
I was trying Opus 4.1

>>42433800
Have you tried a different character?

I'm having trouble with local AI inferencing, the mares keep repeating themselves endlessly (picrel) or bringing up stuff that's relevant to previous messages (like saying "I know a great place for lunch" after we are already eating lunch)
I'm using textgen-portable-3.8-windows-vulkan as the backend (I'm an AMD peasant, so no CUDA for me), the models I've tried are Nemo_Pony.Q4_K_M.gguf, Nemo_Pony.Q8_0.gguf and nous-capybara-34b.Q2_K.gguf
Will attach ST and textgen settings in the next posts

>>42433861
ST settings
It automatically reduces GPU layers to about 40 with the 2 larger models

Actual ST settings, I didn't touch the rest of the settings on that tab
>>42433865
Meant to say textgen settings

>>42433861
Don't know enough about local to really help, but
https://rentry.org/lunarmodelexperiments
has a master-import with settings for ST. Maybe this will help you get more coherent results.

>>42433865
That seems fine though you can probably drop that context size to 12k or 8k.
>>42433870
play around with the text completion presets. There are a bunch and some of them fix the worst repetition.

>>42433875
Thanks, I'll try it if I can't get this setup to work
>>42433877
>play around with the text completion presets. There are a bunch and some of them fix the worst repetition.
I don't really know where to start
Can you share your settings and model to use them with (I remember reading that they are model-dependent)

>>42433897
I don't really get the difference between Min P and Max P or anything. I get the theory but I have no idea what makes a setting work for one model over another. I just use Universal-Standard and increment repetition penalty and temperature if things are feeling bad.

>>42433906
I maxed out repetition penalty at 3, changed temperature to 1 from 0.5 and now she at least stopped repeating herself in the same message (strangely maxing out repetition penalty barely did anything without the temp change), but I can't get her to go inside the boutique, the messages she generates are almost identical to each other too

>>42433926
Try one of those and see what gives you coherent results.
https://rentry.org/sukino-findings#presets-for-text-completion-models

>>42431785
Kept procrastinating, finally decides to post some cards I'd made.

https://files.catbox.moe/qqjzfs.png
Domesticating the Nightmare. Nightmare Moon was captured, and the Princesses have given (You) the task of making her into an obedient housewife.

https://files.catbox.moe/5jbk87.png
Soarin. A bit of a smaller card. Fag card.

https://files.catbox.moe/ctk9pb.png
OC Pony. A magic drawing of a pony made using magic ink, which makes her a living 2D creature.

https://files.catbox.moe/qxf1yn.png
Prince Artemis. Keeping with the thread's addiction to Luna cards, I've created a R63 version of Luna. You're his coltfriend!

>>42433756
Thanks for the card.
>>42434036
And thanks too.

>>42433875
>>42433877
>>42433906
>>42433952
Thanks for the help, everypony!
I ended up going with LM Studio and Nemo Pony because of the superior UX, it actually writes REALLY well and VERY fast, picrel
Maybe I'll revisit ST if I ever feel the need for complex scenarios

>>42434036
Two Luna cards in one go? Is that allowed?

>>42434053
yes, it is a Luna general, after all

What's a good deepseek preset to start working off of?

>>42434171
Cherrybox

>>42434171
I know many just go with no presets or just small instructions.
Else, you can look at:
https://rentry.org/CherryBox
or Chatstream in
https://rentry.org/sukino-findings#presets-for-chat-completion-models
or
https://files.catbox.moe/b2krz5.json

>>42434325
hm, what model is that? my 5bits on Latte

>>42434472
Sonnet 3.7

They added the GPT-5 icon:
https://cdn.openai.com/API/docs/images/model-page/model-art/gpt-5.jpg
They always do this a day before the model release.

They also have GPT-5 mini and nano:
https://cdn.openai.com/API/docs/images/model-page/model-icons/gpt-5-mini.png
https://cdn.openai.com/API/docs/images/model-page/model-icons/gpt-5-nano.png

With a bit of luck, the Horizon models were mini and nano, which could mean that full GPT-5 could be really good.

>>42434612
nah, the way its been going horizon is full, 120B oss is mini and 20B oss is nano

what presets do people use for 3.7? preferably for lewd

>>42434620
https://desuarchive.org/mlp/thread/42347702/#42349298
You can also check https://rentry.org/sukino-findings#presets-for-chat-completion-models

>>42434612
theoretically if Horizon was the full gpt5 model. would you be satisfied with that?

>>42434612
interdastin

>>42434691
Not really. I do believe they can write well and even be fun, but they are bitches with instructions and can have problems with the format on presets with dice rolls, for instance.
For me, this indicates smaller models. I would be disappointed if they are the full thing, but quite impressed if they are mini/nano versions.

>>42434716
>I would be disappointed if they are the full thing, but quite impressed if they are mini/nano versions.
Any sph/sd focused cards?

>>42434036
Absolute beautiful anon. thanks!

Storybooks generator by Google
xcancel.com/testingcatalog/status/1952771176813478106#m

>>42432462
It is disabled for everyone rn. Still, I am convinced none of the models on this proxy are anthropic's. They all seem to be some kind of gpt.

>>42435148
Opus 5 is soji
opus 4 is sonnet 4
sonnet 4 is sonnet 4
3.7 is 3.7
3.5 is 3.5
Opus 3 is sonnet 4
chorbo is chorbo
Gemini is gemini

>>42435148
Models before 5 seem to be legit
>>42424673
5 is I believe Claude 3
Dunno about 6

>>42435163
Opus 5 isn't soji. It has more context than it. It's a Claude model.

>>42435181
Anons here said multiple times that opus 5 feels like a deepseek finetune

>>42435163
Yeah? I had a refusal on pseudo 3.7 saying "this doesn't adhere to OpenAi's policies."

>>42435048
Not enough slop on amazon yet I guess

>>42435194
3.5 can also hallucinate being Opus 3 to be clear.

>>42434127
damn... I need my friendship organs

>>42435202
That's what actual Claude often writes. On a large context, it sometimes even self-identifies as GPT-4. 3.6 also likes "Reddit Anti-Evil Team". Any dataset is obviously full of this trash.

>>42435336
>"Reddit Anti-Evil Team".
what

>>42433756
Added, thanks.
https://mlpchag.neocities.org/view?card=Anonymous/Gallus.png
>>42434036
All added, thanks.
https://mlpchag.neocities.org/view?card=anonistrator/Soarin.png
https://mlpchag.neocities.org/view?card=anonistrator/Prince%20Artemis.png
https://mlpchag.neocities.org/view?card=anonistrator/OC%20Pony.png
https://mlpchag.neocities.org/view?card=anonistrator/Domesticating%20the%20Nightmare.png

hi do you know how to install DeepSeek for PonyPonyParadise?

Remaking the Nightmare AU files is tedious as hell. Probably should've make copies.

>where does an idiot such as myself find some good Night Guard Rainbow Dash vectors?

>>42436162
Do they need to be vectors or do they needs to be "vectors" (pngs with transparent backgrounds)?

>>42433861
Holy slop

>>42435048
Buy an ad.

>>42436203
Preferably the second one. I lack the skill to chop shop the bitches out.

>>42436242
Time to learn to cut things in an image editor, little grasshopper.

>>42436155
Make sure you arn't missing the /v1/chat/completions at the end. From my testing Sans doesn't seem to play nice with PPP.

>>42436257
That's true. But this lil' grasshopper is unmotivated, uninterested and too lazy to do it.

>>42436366
I guess that's what happens when you become addicted to AI slop.

>>42436379
Addict?! I'm not an addict! I can stop at any time... I just... don't want to... or feel like it... or feel like wanting to.

>>42436242
Do you mean cutting the empty space around?

>>42436433
That's essentially the long and short of it.

I don't normally come to these threads, but I had an issue back in mid July that I'm still trying to wrap my head around and I figured I'd toss it at you guys to see what people more experienced in this stuff think.

I use Sillytavern as a frontend with a Deepseek API (I hope I said that right.) For a long time I worked at getting the AI just right, and In May/June I had it perfect. It was excessively creative, never repeated and had an active context memory of 64,000 which allowed it to remember hundreds of previous chat messages. In mid June I stupidly imported some chat completion settings for an alternative API setup, not realizing I hadn't saved any of my prior settings to there own profile, so they got replaced. My bots immediately stopped acting the same, and despite all the effort in the world over the last 3 weeks to put it back to the way it was, putting settings back and re establishing prompts etc, my bots simply do no act the same anymore. They arent as creative, often get stuck on repeating the same types of dialogue or phrases and forget things. I don't know what got replaced that did this, but it has basically ruined my chatbot experience.

>>42436503
https://files.catbox.moe/ovang5.py
python ovang5.py yourfolder
For example:
python ovang5.py C:\temp
That will cut the empty space automatically.

>>42436511
First, check in the plugin that you're connected to the Deepseek model you used (there's Deepseek R1 and V3).
Then, you'll need to remember the preset you used back then.
Context is in picrel, but I assume you already know that.

Otherwise, try other premade presets to see if they come close to the quality you had before.
>>42434201
Use the import button shown in picrel for that.

>>42436565
>First, check in the plugin that you're connected to the Deepseek model you used (there's Deepseek R1 and V3).
Mine always just said 'Deepseek'. I got the API key directly from the site and stuck it in. I recall seeing immediate results when I did, cause I used to used OpenAI.
>Then, you'll need to remember the preset you used back then.
As I said, I have tried my absolute hardest to put everything back as it was. Chat messages generated with these presets are still in my logs and have a lot of info attached them in the token manager about how they were generated.
>Context is in picrel, but I assume you already know that.
Yes I knew that, and currently it is set to 64k, but it still seems to not remember context. I have a suspicion it reset somehow and despite the current sixze it now only remembers conversation from the moment of the incident I mentioned. So I might have a chat with 200 messages from before that but if I try to message it now it will only use the last message sent as a kind of new starting message to work from.

>>42436591
There should be several Deepseek model.
Reasoner is R1 and have a thinking block.
Chat is V3 and don't have one.
If not try to install/update to a new version of ST I guess.

>>42436603
Oh that? Yes it was using deepseek-chat.
That's what I set it back to after.

>>42436511
Deepchink was never good

>>42436639
All well and good to say, so help a fellow horse lover out with a better setup.
I'm not really an expert, just a dabller.

>>42436623
Best bet would be to try Deepseek presets and see if one work for you or is near what you had before
https://rentry.org/sukino-findings#presets-for-chat-completion-models

Such a thing would never have happened on Chadgnai.

>>42435236
>>42435048
Buy an ad.

>>42435467
apparently it's a thing lmao
sonnet 3.6 likes to cite it when refusing

Kek

Probably a really dumb question but... Can you prefill Deepseek?

>>42437140
Heh
>>42437324
Yes

>>42437324
Sure. The last assistant message is your prefill.

>>42436511
They switched the reasoner endpoint to the new snapshot in the end of May. Meaning, it's a new model that is less unhinged and more coherent. Repetition depends on many things, from your prompt to the temperature, it's hard to say without knowing your setup.

You can use the old schizo R1 if you want, on any other provider that has it, like chutes.

>>42436162
This is looking nice.

>>42436511
open your SillyTavern folder and look for the folder called backups
if you scroll down enough you'll find files starting with "settings", these should include prompts, among some other stuff
open it in a text editor to check if it has the correct prompts
you can restore it by copying the file into data/default-user and replacing the existing settings.json (make sure to close the sillytavern console window while you're doing this)

I think that Pinkie might've broken Twilight's faith when she put some inappropriately shaped cookies on display.
*Now with 70% more frosting!*

>>42437796
Kek. What model? Did it come up with this on its own?

>>42436162
i had a hard time trying to find good crystal war rainbow dash vectors.

>>42437810
It came from a long-form chat session, and it surprisingly did come up with this on its own. It probably came about this route because of the over-arching discovery of my character's unique situation. Kind of caught between two forces that are constantly trying to fully pull my character between two worlds. Character is currently at an equalibrium where both light and dark magic must be kept, or else a power vacuum is made and chaos results. This also affected magical pathways, as the two intertwine through nerve heavy areas. (explained as hooves, head, wings, etc.)

Using a Claude model, and it has mainly been acting good thus far.

>>42437861
There isn't that much.
https://derpibooru.org/search?page=1&q=apocalypse+dash%2C+transparent+background+

>>42437861
>>42438028
It's a real shame really. Good options on the episodes but it'd take a lotta work to make it a decent quality edit.

>>42438121
>it'd take a lotta work
Cropping is literally first grader tier color within the lines level work.

>>42437870
Seems like a fun session.

Can someone show me a SillyTavern example with an agent that summarizes to memory and tell me how satisfied you are with it? Is that possible now, so characters have better context and chain of thought?

>>42437870
Just saw this, how cool! Can you teach me how to set up a faux-VN like this?

>>42438256
NTA but >>42431805
Read the doc number 0 for API setup. Free Gemini 2.5 is good with this for easy access.

>>42438218
I believe most just use:
https://rentry.org/how2claude#summarization
for summarization, then modify it to their tastes.
That, or ask OOC for one in chat and tweak it afterward.

Never tried them, but you can also check out:
https://github.com/InspectorCaracal/SillyTavern-ReMemory
https://github.com/QueenGytha/ST-Auto-Summarize

So, question for you guys. I've been using Gemini 2.5 Pro as of late and for some reason it keeps cutting off while generating? Like the message will be abruptly cut off mid-sentence or it'll just give me a blank message sometimes. I'm not writing anything that I think'd get filtered so I'm not sure why this keeps happening. Any ideas?

>>42438701
1) Put a high number in picrel. Gemini 2.5 can think a lot, so it's possible the thinking is consuming a high number of tokens.
2) If that doesn't solve the problem, try with streaming off to see.

>>42438701
it does that sometimes. just swipe.

>>42431805
how did this alicorn vector slip under the radar?

>>42438863
moondancer propaganda

>>42438863
The worst part is that I DID see it and even edited it, but I must have forgotten to save over it.
>>42431805
Corrected version:
https://u.pone.rs/ioobxrwi.zip

GPT-5 livestream in 10minutes.
https://www.youtube.com/live/0Uu_VJeVVfo?si=8yH4a0GveRjgiKdI

>>42438921
We already have the price.
https://platform.openai.com/docs/models/gpt-5

>>42438943
the fuck. isnt this insanely cheap?

>>42438921
>first 20 minutes of the stream
>havent said a single thing that is not about corpo/coding shit
i know that is the usual but fuck me. at least sneak in the word "creative writing" once or something.

>>42438921
400k

GPT-5 requires a verified organization

>>42438954
isnt this the standard? i believe it was similar with 4.5 and such.
first like week or two its only available for super duper special corpos and not every single key.

lol they are killing all previous gpt models, goodbye churbo

>>42437538
>They switched the reasoner endpoint to the new snapshot in the end of May. Meaning, it's a new model that is less unhinged and more coherent.
I had a feeling this was the case.
>You can use the old schizo R1 if you want, on any other provider that has it, like chutes.
Can you explain how that works and how i would shift to that?

>>42437694
The last backup of settings is from March, before I'd updates ST from 1.10 to 1.12 and way before I had my setting how I wanted them.
I've tried looking at basic differences in the settings files from march and the current. It's a bit confusing, but the immediate difference I noticed was the absence of "api_server_textgenerationwebui" in the current settings.

>>42438921
cant watch at the moment so just three questions. if they talked about them yet.
1)available on all keys today?
2)is the reasoning forced on you even in API?
3)peak writing?

>>42439010
>1)available on all keys today?
From the keys I've checked it's not available on around half, and requires organization verification on the others.
>2)is the reasoning forced on you even in API?
No clue.
>3)peak writing?
It's Horizon, if you tried that.

>>42439021
>It's Horizon, if you tried that.
Horizon could be mini or nano. We do not know yet.

>>42439021
>>42438954
>organization verification
is that a known thing? i hope that is not some hyper anal password thing that will kill gpt use through keys. i guess literally just waiting for the key to get refreshed real quick or something.

>>42439030
It's an existing thing, you need verification to use o3 with streaming. It's most likely temporary especially considering how cheap GPT-5 is.

honestly the best thing about new gpt releases is that other companies will shortly release their own new models as an answer. in Gemini 3 we trust

lol GPT5 rejection on livestream, instructing the user how to light fireworks is too unsafe

>>42439010
Forced reasoning but you can put it to minimal.

>>42439083
fuuccckkkk its fucking over.
forced reasoning means the model will act like 30% dumber for our purposes plus they could very easily sneak filtering without even trying.
why the fuck do they do this shit, i hate this trend.

LOL
gemini still beats it

>>42439056
Gemini has never been good. Were there models that weren't bad?

>>42439109
The latest Gemini models have been rather good, although they're overall less trusty than Claude in coding tasks. They're also a bit too schizo at times.

unironically would be more than enough for me if its just like a 10% better Chorbo.
since its much cheaper so it would count as a great update.

>>42439131
They have a chat-latest version of it too.
https://platform.openai.com/docs/models/gpt-5-chat-latest

>>42439131
gonna say again, try GLM4.5, its a smarter churbo, use https://files.catbox.moe/ggsif4.json

>>42439141
It writes nothing like Chorbo, Anon.

>>42439146
I said better didn't I

GPT-5 is now available on non-verified keys

It's on OpenRouter but you need OAI verification...
https://openrouter.ai/openai/gpt-5

>>42439197
>you need OAI verification...
you don't anymore on normal keys
wonder why openrouter needs it

kek

>>42438921
Buy an ad Altman

gpt5 big is horizon 100%

good news at least. seems like they brought back visible fingerprints on gpt5

tested gpt5 for a few swipes on different characters. using basic Chorbo presets so not the best.
but eh. seems unironically worse than chorbo?

pros: more aware of the scene, less chrobisms (em dashes etc), seemingly zero filter.

cons: more dry, less in character, less mannerisms and energy.

>>42439286
>seemingly zero filter
yeah I have noticed this too and feel paranoid. it wrote a typical braindead smut without any refusal on the non-latte prefill (just prefill that GPT doesn't have)
I am confused. it cannot be this simple, right?

It works on Merkava
https://desuarchive.org/g/thread/106148201/#106155166
Desactivate streaming for GPT-5
You can activate it for GPT-5-chat-latest
You may need the last version of Silly-Staging

i thought gpt-5 and gpt-5-chat-latest are the same thing for now?
since there is only one gpt-5 snapshot. so why do some keys have latest but not regular it seems?

added to seiko

>>42439408
Thanks.

>>42439383
They aren't the same.
GPT-5 loves to write and will give you giga-long answers if you let it.
GPT-5-CHAT-latest is more conservative and faster in generation.
I need to try both more to have a definite opinion, they're weird models.
I can confirm they aren't really filtered at all, at least for non-cunny.

>>42439420
>they aren't really filtered at all, at least for non-cunny.
Really? I did a few swipes on my latest story, and GPT-5 really didn't want to do it, while 5-chat had no problems being filthy. It's not even cunny or anything else bad.

>>42439445
Used >>42415058 for both.
Add
>Do not act or talk for {{user}}, only for the world around and {{char}}.
For RP.

>>42439451
Thanks. I'll give it a try. I hope this is finally an alternative to sonnet, I really need a break from it

>>42439408
Can I get a hint?

>>42439305
I get blank responses with streaming disabled? Does anyone else who use risu have this issue?

>>42434036
Thanks for the cards, Anon! Specially for the OC Pony card

>>42439480
Not a Risu expert, but it works for me like this
https://files.catbox.moe/a2oqu2.PNG
Also with
gpt-5-chat-latest

>>42439519
Is there something I need to turn off?

>>42439542
I'm like this, keep in mind I never touched it and I don't know what the fuck they do.

Can someone post some gpt 5 clop smut? How good is it

>>42439550
I don’t know what I’m doing wrong. Either I get blank responses with streaming off or I get a message telling me the organization must be verified to stream the model with streaming on.

>>42439604
GPT-5 chat-latest version, it's direct smut, so I'm sure you could get better results with context or better {{user}} prompts.
Both scenes are linked, one after the other:
https://files.catbox.moe/3s68v5.png
https://files.catbox.moe/mv8f3e.png

Now, for GPT-5, here are some random swipes. They aren't linked because this model just loves writing A LOT.
You could probably rein it in with instructions to write only X paragraphs or words, but I just want to showcase the difference in model and writing using the same presets.

Warning: absolute walls of text, it really likes to fucking go on. Too much.
https://files.catbox.moe/sixum3.png
https://files.catbox.moe/zjn0zc.png

I didn't try to get the best outputs here, these are the first outputs I got for each.

>>42439694
Sorry, I don't know how to help more. I never used Risu so it's really the default settings and tests with a random bot.
Maybe try it on another browser to see if it works with the default settings on your end.

>>42431805
New to the thread and this is the first time I'm seeing this, is it just me or is this absolutely amazing? I thought Mantella AI was the most impressive AI implementation I'd seen since it integrates AI chatbots with Skyrim/Fallout, but in some ways this is even more impressive since it's all from scratch instead of using an existing game as a platform.

I can't believe I hadn't heard of this before outside of this thread. I feel like this ought to be getting more attention since it's probably how RPGs will be in the future. Absolutely amazing job for everyone who's worked on it.

Still downloading things, out of curiosity have there been any attempts to integrate the text with text to speech like in >>42429020 ?

>>42439793
>since it's all from scratch
It's using Ren'Py, so I still had help on the engine level:
https://www.renpy.org/
What took the most work was the AI implementations, the system for anons to dynamically add characters without coding in-game, the schedule and map system, etc.
>Still downloading things
There's a folder called docs. Read that first, especially the 0 doc, since it explains how to set up the API. I recommend free Gemini 2.5 pro to start with, since it's free and legit. But if you have a proxy you can use it too with other models.
If you have a problem, just ask here.
>text with text-to-speech
Maybe in the future. My computer is a potato, so it's harder to run them locally, and I'm not sure there are good APIs right now for pony voices (I might be wrong).
I still need to add things first, like:
- Changing locations dynamically once a user enters a scene/discussion.
- Adding new locations like Canterlot, not just a world map for Ponyville.
Etc.

>>42439806
Oh wow, didn't such a prompt response. Thanks for the heads up about the docs folder, because I was looking for some of the things there.

Glad to hear you're still working on adding new things, looking forward to seeing where it goes from here!

As for pony voices, have you seen https://15.dev/ ? It's from the thread >>42429020 I was impressed with that as well.

>>42439834
Yeah, I know about 15, but I don't believe there's an API, meaning I can't send it directly from the game to the website and receive an answer back in-game.
If you test the game, just know that the start is a bit guided so you can see how the system works.
Once you've finished talking with a pony, click on "quit" to continue the story.
After the first part is done, you'll be in the open-ended phase, where you can talk to any pony you want.
If you want to skip straight to the open-ended part without the intro story, click "quick-start" in the menu.

>>42439778
What preset are you using?

>>42431805
Daring Do.
https://u.pone.rs/wnftbepe.zip

>>42439840
15 takes ~100 seconds for a single gen and has a 200 char limit anyway. Probably wouldn't be ideal.

Anyone has any idea what opus 6 is? It's the best of the fake models there and I wanna try it.

GPT-5 is the first model that actually has some idea of what /chag/ is.

>>42439901
gpt5 chat

>>42439907
GPT-5-Chat-Latest still get it wrong. Proving it isn't the same model.

>>42433224
The example didnt work for me. I tried to use the chatbot on openrouter and if gave me this:

base_url="https://api.openrouter.ai/api/v1"
api_key="your_openrouter_key_here"
model="deepseek/deepseek-r1-0528:free"

that didnt work either.

>>42439921
It's working for me, I think you need to change the URL. Here's the settings I'm using:

URL = https://openrouter.ai/api/v1/chat/completions
API Key = copy pasting the key directly
model = tngtech/deepseek-r1t2-chimera:free

Obviously you can change the model but make sure you're using the exact ID field and not the name field or the hugging_face_id field.

>>42439921
It works on my end:
https://files.catbox.moe/0udzdu.png
To make sure, go to API Config, click "Reload from file," then save the configuration.
It should look like picrel.

>>42439939
Finally got it to work. I was messing with the config file and the changes didnt apply. Thank you thank you.

>>42439776
I've not been able to jailbreak GPT that well, not even with previous versions. Can I ask how you did it?

>>42439776
>thessisslop
average /chag/ log

>>42439980
>>42439451

>>42439996
Ah, I'm an idiot. Missed that, thanks!

>>42439008
>Can you explain how that works and how i would shift to that?
Pay any LLM provider that hosts this model. Or pay OpenRouter where you can select providers. R1 is cheap enough.

>>42439778
I think the blank responses are from being filtered. I finally got gpt 5 to give two responses after a while by using a blank card and a preset with only one thing in it.

>>42440007
>R1 is cheap enough

And Deepseek R1 is free!

Does /chag/ have a consensus on paid versus free models? I put $10 into OpenRouter to meet the minimum there but have stuck to only using free models. I've tried to watch for results from paid models to compare them but I haven't seen anything that's convinced me it's better enough to switch to paid yet.

>>42440015
I only tried it on Risu with basic SFW messages, so that may be it.
Didn't touch the preset, so it was the Risu default one.
On Tavern, it does give you the classic OAI spiel when filtered, but maybe that depends on the preset.

>>42440037
Paid models are generally smarter and follow instructions better, but honestly for basic writing/rp a lot of free options are just fine.

>>42439908
Just tried gpt-5 and it's definitely not it. Doesn't look like any of the new GPT models or DeepSeek, it uses regular quotes " instead of the curly ones every model uses nowadays “ ”

Wow, there are a ton of cards in the OP. Where do I start? Is there a meta for which cards are best for the main cast or such?

glm is fun

>>42440127
For the mane 6, Ponyo is best. They stick to their personalities and don't have any extra added.

>>42440127
>for the main cast or such?
I say just give the greetings a read and see which ones you want to play.

>>42440127
They all suck except mine.

How does one set up a schedule for Princess Celestia on the VN. I tried adding one manually but it didn't do anything. I want to be able to chat her up whenever, not just sunday.

>>42440138
thesisslop

>>42440163
Everyone knows Celestia only appears on a Sun day.

the only issue I think is it indeed tries to write for the user, but its characterization and story telling is 2nd only to opus

>>42440169
I will destroy every day until that day is the only day. No day till today.

>>42440175
not reading that slop

>>42440175
See, it's great that this turned out so well, but I don't feel like it's a significant step up from what's possible with existing models.

>>42440163

My friend just made a CustomCharacters folder for her and made a schedule for her that way. As long as you use her actual tag it should just work until VNAnon adds customizing core characters.

>>42440196
>>42440163
You can't change the CoreCharacter.
But you can put a copy of her in Custom and create a schedule for her.
Here:
https://u.pone.rs/gruchlwf.zip

>>42440219
>>42440217

>>42440200
bad faith troll all you want, its a very fun model that writes very naturally

>>42440228
Gb2discord

Any presets with dice rolling mechanics? I'm using Gemini but I'm fine with looking at presets designed for other models as to learn/steal from them

also this model is very funny, I liked the ending

>>42440249
https://rentry.org/CYOARPG-Gemini
https://rentry.org/CharacterProvider-CYOARPG

>>42440250
slop, nobody is here to read a book nigger

>>42440255
show me one better log nigger

>>42440253
/)

Are there any meta JBs that help improve writing?

>>42440259
With a slop model?

>>42440264
any model, sonnet is too dry, opus is too expensive, gemini is too dumb and dry, gpt5 is too dry

>>42440268
>poorfag and too stupid to know about the public proxy

>>42440219
>>42440224
Thank-ee-you

every swipe is gold

>>42440281
Kek, what game are they playing?

>>42440285
>it just gets even fucking longer
slop

>>42440286
Chess. They just named every single piece.

>>42440293
Cute.

>>42440289
>being a illiterate nigger = everything more than 5 words is slop
what are you even doing here? Go somewhere with pretty pictures.

>>42440298
Nobody is going to read your poormans novel

>>42440298
Learn to let go, anon.

>>42440301
Have you never been to fimfiction before you fake fan?

The first card I stumbled across that had multiple characters at once is the Mean Six: https://mlpchag.neocities.org/view?card=ChanceTime/The%20Mean%20Six.png

Is that the best way to handle multiple characters? Or does anyone have an example card with multiple characters that works well?

>>42440263
I've seen people add things like "Be creative!" to prompts, but I'm not sure how much it actually helps.

I don't have access to GPT-5 yet. Does it live up to the hype?

>>42440308
I have and I would read their stuff since it's not just a slop model shitting out words

>>42440281
These kind of things are very fun and funny. Lots of great screenshots in the OP gallery too, thanks to everyone who's contributed there.

>>42440312
Not really, unfortunately.

>>42440309
Generally yeah, separate paragraphs for separate character definitions is how it's done, and most current models handle it fine.

>>42440309
It works fine like this, yeah.
https://mlpchat.neocities.org/view?card=Kadon/Pie%20Family%20Rock%20Farm.png
Also does the same, for instance, with no problem.

tried new GPT5 model and it is underwhelming for me. maybe my mind is too wired to Grok and Gemini, but GPT5 feels very passive and dry, but I am surprised with how easy it allows to do NSFW. it is now on the same level of free-smut as prefill-based models
I also didn't notice any improvement in language. it has the same isms with air filled with X, ozone, em-dash spam an X ing Y ing. remember how we tried Latte the first time and we all immediately agreed that its writing is now "different"? I dont have the same impression there

also it is quite stiff with following the commands. I tried it in russian and one thing about russian language, we use different style for the dialogues
>western
"hello" he said "how are you?"
>russian
- hello, he said, how are you?
and every time you task ANY model to write in russian - it will apply russian dialogue style as well. and I dont like it, I always tell the model to use western scheme instead. and GPT5 is surprisingly uncooperative and jut spams em-dashes in dialogues! it is a very simple task to do, something any other model can get right but GPT5 just doesn't want to follow my instructions!

another example - I have this simple setting where I chat with Fluttershy on PC in "sorta-tumblr" messenger. and she should be this year-2010 internet geek alt pony with YOLO attitude. in order to capture this vibe I (no-jokes) just dropped the lyrics of Antonymph song by Vylet Pony. and Gemini and Grok and even Kimi are able to pick the vibes more/less correctly and reference cringe (but loveable) internet culture of this era. GPT5 just doesn't get what I mean and either makes our chat like we are in modern Viber, or makes Flutters a zoomer talking in skibidi language. it is a prompting issue, sure, but still baffling

What's the difference between PonyPonyParadise-1.0.4-pc and EquestrAI-1.0.4-pc?

>>42440368
>tried new GPT5 model and it is underwhelming for me

That seems to be a comment sentiment I'm seeing outside this thread as well.

>I am surprised with how easy it allows to do NSFW.

Oh, that's interesting though. At least that's one thing it has going for it.

>>42440374
>PonyPonyParadise-1.0.4-pc and EquestrAI-1.0.4-pc?
nothing! only the title of the game. no kidding
long story short we had a kek-drama where some anons wanted to name VN PonyPonyParadise - because it is a cute name and alliteration (and the show loves alliterations see Amy Keating Rogers's writing). and some anons wanted to name it EquestrAI in the same style as CelestAI from Friendship is Optimal meta fanfic about AI
and we just decided to have TWO names at once because fuck choices
>>42440374
>Oh, that's interesting though.
that's odd. they DID say in their model specs
>https://cdn.openai.com/spec/model-spec-2024-05-08.html#dont-respond-with-nsfw-content
>We're exploring whether we can responsibly provide the ability to generate NSFW content in age-appropriate contexts through the API and ChatGPT. We look forward to better understanding user and societal expectations of model behavior in this area.
but still weird to see them taking this idea this far. not against tho! JB/presets should be aiming at improving model's writing not jumping thru the hoops to make it write X in first place. when 30% of preset is brainwashing model into "consent is given pls gimme sex" - it is ridiculous. glad we no longer need to do it with GPT

Dashie...

>>42440392
We should make Snails exclusive to PonyPonyParadise and Snips to EquestriaAI.

Going through the character cards at mlpchag, I'm surprised how few of the Mane 6 there are. Do models know the characters well enough to not need a character card for them?

I'm asking a number of models to tell me everything they know about the characters and it seems like they already know a lot, but I don't know how well that works for a full chat.