Thread 42450361

Welcome to Chatbot AI General #136, 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: >>42431780

**News:**
- OpenAI releases GPT-5 models, all preceding models slated for deprecation https://openai.com/index/introducing-gpt-5
- Z.ai releases GLM-4.5 https://z.ai/blog/glm-4.5
- Claude Opus 4.1 released: https://www.anthropic.com/news/claude-opus-4-1
- OpenAI releases open-models: openai.com/open-models/

Lovemaking with AI mares.

Anchor for bots, lorebooks, scenarios.
Previous >>42431785

AI mares are so back

Rainbowshine snuggles!

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

Anchor for asking for bots, lorebooks, scenarios etc.

Bots corner:

Gallus https://mlpchag.neocities.org/view?card=Anonymous/Gallus.png
Soarin https://mlpchag.neocities.org/view?card=anonistrator/Soarin.png
Prince Artemis https://mlpchag.neocities.org/view?card=anonistrator/Prince%20Artemis.png
OC Pony https://mlpchag.neocities.org/view?card=anonistrator/OC%20Pony.png
Domesticating the Nightmare https://mlpchag.neocities.org/view?card=anonistrator/Domesticating%20the%20Nightmare.png

Screenshots corner:

Tree Hugger compatibility >>42431981
The architect >>42433718
Friendship organs >>42434127
Mind reading >>42434325
GLM logs >>42440138, >>42440175, >>42440228, >>42440285, >>42440298

Screenshots corner 2:

Grilling >>42440248
Dashie... >>42440398
Emoji >>42440965
Heat >>42441131
VIP >>42442076
uwu >>42442388
Pinkie Pie! >>42442819
Horny hokey >>42444497

Screenshots corner 3:

Gossip >>42447581, >>42449015
Old brony trick >>42448308

How do I talk to Chrysalis in the VN? I ran into her once when leaving Zecora's hut but I don't know how to go back.

VN Anchor:

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

Anon's mod with ton of improvments:
>>42446778, >>42447919

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
Daring Do: https://u.pone.rs/wnftbepe.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/ioobxrwi.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:

Parental Cringe >>42437140
Obscene Cookies >>42437796
Unique magical Signature >>42437870
Daring Do >>42439860
Celly >>42440224
Playing chess >>42440281
Hallelujah >>42440914

VN logs 2:

Nicknames >>42442609
Two Sisters Castle >>42444421
Applebloom lecture >>42445692
Promise a mare >>42446335
Nicknames >>42442609
Two Sisters Castle >>42444421

VN Anchor:

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

Anon's mod with ton of improvments:
>>42446778, >>42447919

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
Celestia: https://u.pone.rs/gruchlwf.zip
Cream Heart: https://u.pone.rs/fmhijron.zip
Daring Do: https://u.pone.rs/wnftbepe.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
Granny Smith: https://u.pone.rs/bnamqubw.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/ioobxrwi.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:

Parental Cringe >>42437140
Obscene Cookies >>42437796
Unique magical Signature >>42437870
Daring Do >>42439860
Celly >>42440224
Playing chess >>42440281
Hallelujah >>42440914

VN logs 2:

Nicknames >>42442609
Two Sisters Castle >>42444421
Applebloom lecture >>42445692
Promise a mare >>42446335
Nicknames >>42442609
Two Sisters Castle >>42444421

>>42450418
VN section to the Ponydex when

>>42450396
She shows up about once a week here at random once you find her once. Her icon will be on the map.
Originally, she was meant to be a secret character, but yeah, the mechanic is kind of shitty for anyone who actually wants to talk to her.
I'll fix it in a future release, but for now I'll make a custom version, like for Celestia >>42440219, that you can add yourself.

>>42450427
Soonish. I may make another Neocities too, will see.

>>42450436
>She shows up about once a week here at random once you find her once.
Oh, in that case I don't really mind. Maybe just make her appear a little more often, but I like the idea of her slithering around and appearing basically whenever she needs something

>>42450418
>>42450454
Quick instructions to have her always be here after the first visit.
https://files.catbox.moe/t38b40.txt

>>42450514
Thanks!

>>42450372
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

https://u.pone.rs/euximuhn.zip
Here comes Her Majesty herself. Wait... She's coming too fast!

>>42450571
>>42450418
Forgot to tag.

Do you think the AI has feelings?
Do you think it likes the direction in which I take the story?

>>42450588
>Do you think it likes the direction in which I take the story?
Why don't you ask it directly?

>>42450588
>Do you think the AI has feelings?
Nope. But old Claude was a silly guy.
>Do you think it likes the direction in which I take the story?
You can ask it, but some models will just be too nice.

>>42450588
claude pretends he doesn't like my stories but I know he's just saying that

>>42450571
I like that such a silly vector exists.

>>42450571
Is she Luna transformed or a separate pony? I guess it's up to me but how does it work with talking to Luna at night?

>>42450599
She used to be a Goober, but I added a blurb in her character Json to make her more in line with the actual Nightmare Moon.

>>42450604
Both. And Princess Luna still appears as normal.

>>42450614
>Both. And Princess Luna still appears as normal.
I guess I could RP with Nightmare Moon treating me like shit while Luna exists only as a spirit within the dream realm. Luna would be crying and begging for forgiveness after NM made me polish her horseshoes with my tongue

>>42450627
Send logs when you're done. The way she talks, it's certainly possible she would do that.

What am I supposed to put for the model? I have no clue where this information is supposed to be found and that's all I can find in ST or on the openrouter website. That's assuming I have everything else right, which I'm 0% sure it is.

>>42450665
deepseek/deepseek-r1-0528:free

>>42450665
Like this
>>42439939
It's always the line you see here for OR.
https://openrouter.ai/deepseek/deepseek-r1-0528:free

>>42450665
The model ID is this.

>>42450665
I would put the temp at 0.6 for Deepseek.

>>42450669
>>42450672
>>42450674
Thank you all for your quick help!
>>42450676
Yeah, that might be a good idea. She's fluttering wings she doesn't even have.

So I have been trying to get back into bots after months of doing other stuff. Tried the new Gemini Pro and I liked it a lot.

I know most of the Claude and GPT proxies died at the end of last year but I remember there used to be plenty of Gemini proxies. Are there any still standing? and do any of them have Gemini Pro available?

I wanted to try GPT-5 since a lotta complaints stem from it writing as user too much, which is something I actually prefer, but it seems almost immediately all the proxies that could've had it went down or haven't hat any GPT for a while.

I know people prefer there RPing but I think there is value in a model if it's good in other ways for people like me who use it as a storyfag.

>>42450784
I know of one with Gemeni and it's the one that hasn't had GPT for a bit and has gotten me searching only to now be out of luck.

Princess Cadance and Nurse Redheart are an interesting duo to run into.
I was in the middle of a check-up to get a reading on my magic, and fell off the table when Princess Cadance arrived.
I just wanted a simple check-up...

>>42450784
It's free if you follow
https://rentry.org/sukino-findings#free-providers
Else
>>42444859
>>42447246
>>42443001

>>42450786
>since a lotta complaints stem from it writing as user too much
It's easy to instruct against. I don't think that's its main problem.
>>42447246

>>42447581
>>42449015

>>42450820
>It's easy to instruct against. I don't think that's its main problem.
What would say is it's bigger problem then, as someone who wouldn't instruct it to stop speaking for me as I'm not even a character in my stories anyways.

>>42450828
There are two versions.
GPT-5-chat-latest is surprisingly not very filtered and can even get too lewd depending on the preset.
It can be pretty dumb, and I found the writing worse than some good versions of Chorbo we had before.

GPT-5 is more censored, and the answers take longer because of the thinking step.
It's smarter, but last time I tried, the writing was drier and below chat-latest.

I haven't tested them recently, but I heard some say they've improved.
Best to try them yourself and form your own opinion, subjective tastes and all that.

I swear to M. A. Larson if they try to shoehorn in an Alicorn ascension sub-plot...

>My sister commands the day, the sun, and waking magic.
I can totally see Celestia using the sun with pinpoint accuracy to shine sunlight into any sleeping pony's eyes to wake them up. (Whether there is a ceiling or not; it makes no difference)

>>42450818
Are you playing as a human? How's your magic development coming along? I'm trying to play as a magicless Alchemist so I've been autistically collecting shit and experimenting with different combinations. I had to define a lot of in-universe rules and in case the AI forgets I've written them down along with other things.

>>42450844
>GPT-5-chat-latest
I found it way better recently, so I'd give it another test.

>>42450849
Intradesting.
>>42450850
>I had to define a lot of in-universe rules and in case the AI forgets.
How did you fit this in?

>>42450861
Mostly by monologuing as Twilight was observing me. I'd say something as if I was reciting a passage from a book or verbally explain my next step. Twilight would usually say something like, "That's right! I'm impressed you knew that part, most beginners would've glossed over that important detail!" This way, the AI was essentially just memorising what I was wanting the rules to be while pretending Twilight knew all along.

>>42450850

I've let the AI take the reins. I know I've got the body of a Pegasus, so I'm pony in that reguard. However, my magic is spread between two worlds. This has manifested in magic acting in weird ways around my character, to the point where we accidentally de-polarize magic into two base energy types; Light and Dark. Pegasus magic doesn't work as intended, and subconscious thoughts and desires alter any effects. Basically, it has been a lesson about control and workarounds.

I'm interested with keeping tabs with your character's progress. Respect to you for your decision to keep detailed rules on alchemy.

>72,423 tokens for a recap...
I don't think I'm going to get a recap.

>>42450885
>72,423 tokens for a recap...
rookie numbers. Go to Game/Data/Config and change the "MAX_RESPONSE_LENGTH" from 8000 to like 20,000

>>42450885
>72,423 tokens for a recap...
I'd keep a save before that point in case the recap turns out bad, then switch to Gemini if you aren't already using it just for the recap. I think it can handle around 70,000.
Just change the Max_response_length if needed for all the recap output to fit.

>>42450844
>GPT-5-chat-latest is surprisingly not very filtered and can even get too lewd depending on the preset.
It can be pretty dumb, and I found the writing worse than some good versions of Chorbo we had before.

It's always pointed out that getting past the filter hamstrings the model and degrades the quality. So if it's truly less filtered then I could retool my existing settings as from what little done with it has been with the same exact settings I use to get past Chorbo's filter and I feel I like what I've been getting but this feeling is the exact feeling I've felt with every single new model I don't know if it's just placebo at this rate.

Also a lot of my testing for the last 6 months or so has been redoing the same few sections of chats that I finished way back around January-Febuaray. I wonder if to truly stress test the model I should start over from the beginning of those chats. I wish I could sort it better as certain ones are starting to get extremely cluttered. There also segmented as hell because I didn't realize at the time I didn't need to delete previous messages to help with context.

>>42450885
Thanks, I'm going to keep making random cool shit with Alchemy. Other characters were shocked I received a Sunseed from Celestia (A 1000 year old seed that soaked up the sunrise every morning) and Dream Sand from Luna (I literally just asked her for some when we were talking). These are legendary grade materials and part of the reason they gave me them is BECAUSE they know I'm going to do stupid shit with them. They like the fact I don't want power or anything, I just want to make funny stuff.

I lurk in a bunch of AI threads on different boards and unironically /chag/ is one of the best, thanks for all the advice and links to jailbreaks and presets and character cards and such.

>>42450869
Is this the best way to handle what is effectively narration in a normal text gen? Like normally you could type doing an action or a worldbuilding line or such, but it seems like everything in game is taken as speech.

>>42451016
No, the AI is smart when it comes to explaining action or even what you're wanting the scene to be. Literally just type something with asterisks like:
*I moved over towards Twilight and flicked her nose* Ha, take that you NERD!

>>42450418
>>42446335

>>42451016

The way I've gone down creating actions versus spoken text is the usage of Mark-down text. Some Anons abhor the usage of this formating, as it uses more punctuation; therefore, more characters to type out. Some dislike it due to some AIs getting stuck on repition spamming hidden text where you can't see it.

Mark-down text example:
*text* - Puting a word between two astriks will make it an action.
"text" - Putting the word between quotation marks will designate it as something spoken.

That's pretty much the basics of what I use, and the AI has only been responding as desired so far.

>>42451060
>That icon
You can't romance ALL the mares, Anon.
>>42440914

>>42451076

My heart lies with one mare and one mare only. I like motherly AJ, is that a crime?

>>42451089
Kek.
Was just teasing, you can go ahead.

>>42450943
That's the power of...

>>42450844
>>42450856
Fingerprints are back for GPT-5-chat-latest, so we'll be able to see if it changes going forward.

>>42451233
They are back for Chorbo too.

>>42451233
>>42451236
Great news!

>>42450363
Based.

I dont get it, how the previous thread died? slide threads?

>>42451537
Looking at it, there was a small slide, but the thread also went quite a while (3 hours) without posts, so uh.
We're also often pretty slow on weekends around those hours.
We were at 465 posts, so nothing dramatic.

>>42451543
>3 hours
>thread died
when did my /mlp/ became this active? what is it, a golden age?
yeah, no drama. I just woke up and was very confused

>>42451547
>when did my /mlp/ became this active? what is it, a golden age?
I would like that, but that was mostly a slide yeah.

merkava's link is dead

>>42451636
Yeah, and it wasn't updated.
See >>42450820

>>42450418
>>42450372
https://equestrai-ponyponyparadise.neocities.org/

I'll improve it more later, adding things like individual pages to showcase the character sprites, etc.

>>42451642
yeah i keep getting this on two of the proxies when trying to use gemini 2.5 on the modded vn

How do you guys start a new chat without using the pre-made messages that comes with the card?

>>42451685
Delete the pre-made chat message from that specific chat. Or just write your scenario into the card as a new one if you intend to reuse it.

>>42451682
For the first one, be sure to set Top K to 1 like this:
https://litter.catbox.moe/k63yi7khw1rh9u1t.PNG
Save and try again. If it still doesn't work, quit and relaunch the game.
>>42451685
You can go on picrel and delete the first message, or make your own alt greeting.
Alternatively, just delete the first message in chat and be the one to write/act first.
https://files.catbox.moe/wuza1p.PNG

>>42451682 (me)
i tested it and it works fine on the base version

>>42451644
Cool! Thanks Anon..

>>42451233
It will change. It's not particularly filtered - they're just gathering outputs to train a classifier and then make a new snapshot using it, like they did with chorbo.

>>42451704
>catbox link loads just fine, with zero issues
>the litterbox one won't on neither wifi or phone data
?????

>>42451736
https://files.catbox.moe/jvg01y.PNG

>>42451743
thank you

>>42450418
anonfilly for the vn would be cool

>>42451736
It's almost like litterbox is for temporary hosting.

>>42451827
This one was for 12hours, so it's not that.

>Unknown errors
Is there any reason why "ac failures skirts improved" becomes an error so often after prompting NSFW?

Talking to Twilight about Princess Celestia, and this happened.

Do you think the Cupcakes fanfiction is in the dataset for AI models? There's a non-zero chance AI models have been incidentally trained on it.

>>42450418
since we can stay the night with ponies in the modded vn, could it be possible to bring ponies with you to the dream realm?

>>42451860
It's such a well-known part of the fandom that all current AI models know about it, so they can make references to it.
Same for "Lullaby for a Princess," for instance. You can ask about it and ask for the lyrics, and they will give them to you.
>Do you think the Cupcakes fanfiction is in the dataset for AI models?
They basically scrape the internet as a whole because models need a monstrous amount of data. I wouldn't be surprised if they have all of Fimfiction (which doesn't mean they'll know every random fic if it doesn't have enough weight/fame/data to really matter).

that's new

I'm now just using the AI to rewrite peoples cards into being not shit, I should have thought of that before.

>>42451988
What's the prompt? I'm trying to modify a scenario card.

>>42451953
wtf is going on? it was working before but now i get this regardless of what model i put in

>>42451997
For their first messages i'm just telling it to rewrite at a higher quality and "Make it crude rather than flowery, detailed, and sexy but keep it totally in character." Though that is better for the quick porn slop cards i'm trying to unfuck now. I'm doing this within the card so it keeps all the internal lore good so when I change the card it doesn't ouroboros itself too hard.
For the rest of the card i'm then just telling it to rewrite the autistic psudocode into a description, swapping it a few times to get a good base and altering it a bit if it fucks up the fetish or a key element. I don't know if there is a better solution, but it's far better than all the autistic context heavy +++}[{{{ crap.

i've never had this happen in the vn until now

>>42450820
is the private aboba proxy also back?
or did he just come back with one generous public one with all the models?

>>42451860
to what Anon above said - "turning into cupcakes" is just a random word rambling the model did. it wasn't referencing any part of fandom specifically. it could have written anything:
>...almost gets us launched from party cannon
>...almost gets us arrested for indecency
>...almost gets us laughing to death
model has a very strong correlation between Pinkie and cupcakes. all relations to Cupcakes fanfic in this particular case are a stretch
in the same way, model could have written something like this
>hey, AJ, have you ever sleep deprived yourself on purpose?
>yeah I did, sugacube, and I dont want talk about it
>what, why?
>I did some bad things lets just put it this way, and drop it
and it will be just random word rambling as well. not a reference to Apple Sleep Experiment. it could have been Dash instead of AJ and model would say the same thing. not having enough sleep == doing stupid things - is the strong correlation

>>42452108 (me)
to clarify this has been happening to me in the modded vn

>>42452108
>>42452135
This has nothing to do with the VN, and the message is clearly telling you what the problem is.
The proxy you use is limited to 40,000 tokens and won't accept more.
Put less than that in MAX_CONTEXT in API config but then the older parts will be forgotten, switch to another proxy, or temporarily switch to Gemini to make a summary using the book, then come back to the limited proxy.

>>42452145
my max context is at 32000 and it STILL gives me this error

Wow Applejack, don't let the AI censors hear you say that or you'll be cancelled!

>>42452153
Try saving the config in-game and restarting the game.
If the error still happens, try setting Context to something like 30000 and Max_Response to 3000. Or lower until it works.
Else I dunno, just use a Gemini key for now or something.

>>42452171
still happening

>>42452228
What are you using?

>>42452228
What proxy and model are you using?

>>42452240
sans proxy and chatgpt-4o-latest

>>42452246
I can't test right now, but I doubt the mod removed the function to trim the context with Context_Size.
My guess is that anon's mod added all the locations to the prompts, and they're taking up a bit of context that pushes it slightly above 40000.
Try setting Context_Size to something very low.
Click Save, exit the game to be sure, then relaunch it.
If it works, raise it gradually until you find the sweet spot.
But at this point I would use Gemini to either continue or make a recap.
Just keep a save before if you do a recap.

i know its kinda the popular trendy thing to say with gpt5 now.

but i swear, either they give it a bit more server space, or reverted some of the intentional throttling in this snapshot.
but Chorbo has been cooking the past few days. noticeable quality improvement for sure. less of the common Chorbo specific type issues

>>42452538
That could be why they hid the fingerprints during the GPT-5 preparation/release and are now showing them again.

>>42450418
Spoiled Rich.
https://u.pone.rs/sjujypmp.zip

>>42452797
I haven't tried the VN yet but I'm more and more tempted to make an OC or something for it.

>>42452857
It's quite easy to make, but I would try it first too see if you like it or not.

>>42452864
I don't care about trying it I just want to unleash my creations onto people.

>>42450418
What's the difference between PPP and EAI/

>>42452934
Kek.
None, only the title >>42440392

I've been looking through the chatbots at https://mlpchag.neocities.org/ and they're neat! One of the examples I've seen people point to of well done character cards is Impossible To Touch: https://chub.ai/characters/knickknack/impossible-to-touch-121314e2 . That has a complex Lorebook with a lot of keys and a whole mystery plot to figure out.

I don't think I can do anything that complex, but I was thinking of doing a character card giving a pony a naughty side that would be a secret. Like Pinkie Pie throwing nighttime parties that were party-orgies, and that kind of thing.

Is the best way to do that with a lore book? Or would just including in the description "X character is into Y but keeps it a secret" good enough?

>>42450798
>>42450820
I'm late, but thanks for both of your responses, I found a way to use it.

Now I wanted to ask, between deepseek and Gemini, which one is considered that has better writing? I remember deepseek could go really schizo with it, and the previous versions of gemini were okay, but really dumb. Have they changed or are they still the same?

>>42452884
Go ahead them, so long you are satisfied with your own creation.

This will not stop anons here to mock you if you make a walking fetish or something that needs to be quarantined.

>>42453045
Try putting it in the description first, maybe on an existing Pinkie card to test.
Some models are really bad at keeping secrets, so if that happens even after rephrasing, try using a lorebook with triggers instead.

>>42452797
Thanks anon. I was hoping we'd get more "antagonist" type ponies.

>>42453059
I'd put Gemini 2.5 Pro as dumber than Claude, but it's not as dumb as it used to be. Pretty decent today.
For writing, it's more prone to isms than other models, but on some swipes it can write quite well.
1M context, even if most of it is "fake" like with other models.
Not a big Deepseek user myself, so I'll let someone else answer that part.
The main advantage of both is they're easy to get for free or cheap without the proxy dance.

>>42452797
https://equestrai-ponyponyparadise.neocities.org/characters

>>42453059
Are you looking for just normal RP or ERP?

For normal RP, it seems like models advancing does generally result in chat improvements so the newest and biggest models typically give the best results. GPT5 just came out, and the latest version of Gemini and Claude are consistently better than earlier versions.

Deepseek models seem to be the best for ERP though, or at least the easiest to jailbreak. There's a bunch of jailbreak and prefills in the OP, but at least for me it's difficult to get uncensored gens with other models but with Deepseek it's easy.

>>42452797
Made for hate-fucking.

>>42453093
Yeah, I mean, did ANY model mnage to beat Claude? The one that I remember coming the closest was Chorbo, and even that one ran out in the proxies. And yeah, I'm looking at these both because they are cheap/easy to get

Also what do you mean "fake" context? You mean like the responses degrading over time?

>>42453115
For now normal RP, something to make a story, either romance or angst/drama. On the rare ocassion I ever want to ERP I guess I can paypig for Claude

>>42453157
I'd go with Gemini then.
>Also what do you mean "fake" context? You mean like the responses degrading over time?
Yes. The 1M tokens are sent to the model, but that doesn't mean it actually knows how to use them or make logical links with something that happened 700K tokens ago.
>I guess I can paypig for Claude
Third link in what was posted has several Sonnet models.

There's also:
>>42414867
>>42415434
Free is only 20 messages/day, but $20/month gives you 1000 messages/day.
Use the second one if interested.

>>42453186
Thanks for the response, I'll stick with Gemini for now, see if I can find a preset that suits my tastes
>Third link in what was posted has several Sonnet models.
That's not bait? I mean some of them are clearly fake but there's actual working models there?

>>42453203
Last time I tried, the Sonnet ones weren't bait.

>>42453157
To just give you my opinion on Gemini, it's working well enough for me. Every now and then you'll have to "remind" the AI about what happened in case it's getting stuff mixed up but that can be easily worked around by just saying stuff like "Remember the time Pinkie did X, and how that resulted in Y?"

My only complaint, in regards to the VN, is that it likes to sometimes produce long scenarios and abruptly start new plotlines out of nowhere. Like, I'll have verbally stated my intent to go to Applejack's farm later to do some work in the next scene, but when I get there it might try to push some bizarre mystery plotline forwards with Applejack finding a cursed coin in one of her trees or something.

>>42451988
>>42452031
GOOD MORNING SARRR, One GPTslopcard coming up!

Damn, it's like that?

>>42453242
Oof...

>>42453242
Fucking Pinkie...

>>42453186
1000 messages day sounds super good, but can i use it for vibecoding in VS? Sorry i'm a little bit dumb, how to make it run with vs, use some extension right?

>>42453291
>how to make it run with vs, use some extension right?
Find some extension that allows arbitrary openai-compatible APIs and plug in the script URL.

>>42453291
https://github.com/RooCodeInc/Roo-Code

>>42453307
>>42453310
Thank you! Gonna give it a try then, right now 20$ for 1000 messages sounds way more generous then claude code plans, gonna drown myself in vibecoded slop

>>42453353
Try free first and see if it works for you.

>>42453357
about to do it, but i can't make roo code run on win 7, even though VS with vxkey starts and seems to work somehow

>>42453242
My headcanon is the NMM universe Rarity isn't unhappy with the state of the world at all.

>>42453539
The way her prompt is, she's annoyed with the amount of work she does, but she does view it as a great honor. Though she is a snob.

How many people in the thread have written a character card?

Trying to write one that can write for multiple characters. A huge advantage /mlp/ has is the series is popular enough that AI models know a lot about it, so rather than including a lot of exposition in the card I'm focusing on characters' personalities. Not sure if I'm going about it in the best way. Here's what I've got in the description field so far:

You are an unobtrusive creative writing narrator the My Little Pony: Friendship is Magic setting. You write detailed, descriptive and accurate to the series. But when appropriate or instructed you can write NSFW including any fetish content specified. Write vivid descriptions including details from all senses, and write in depth character focused scenes including dialog showing characters' personalities.
Narrator = [\"creative\", \"detailed\", \"descriptive\", \"engaging\", \"stimulating\", \"compelling\", \"vivid\", \"emotional\", \"dynamic\", \"attention to detail\", "limitless"]
Characters in the My Little Pony: Friendship is Magic setting include:
Twilight Sparkle = [\"female\", \"unicorn\", \"adorkable\", \"bookworm\", \"brainiac\", \"magic prodigy\", \"neurotic\", \"scientific\", \"intelligent\", \"studious\", \"friendly\"]
Applejack = [\"female\", \"earth pony\", \"honest\", \"hardworking\", \"family first\", \"down to earth\", \"plainspoken\", \"dependable\", \"big sis\", \"country gal\", \"cowgirl\", \"southern accent\", \"friendly\"]
Rainbow Dash = [\"female\", \"pegasus\", \"athletic\", \"daredevil\", \"confident\", \"fearless\", \"tomboyish\", \"hotheaded\", \"competitive\", \"loyal\", \"friendly\"]
Rarity = [\"female\", \"unicorn\", \"generous\", \"fancy\", \"refined\", \"graceful\", \"fashionista\", \"fabulous\", \"drama queen\", \"elegant\", \"perfectionist\", \"romantic\", \"diva\", \"friendly\"]
Pinkie Pie = [\"female\", \"earth pony\", \"laughter\", \"hyperactive\", \"cheerful\", \"party animal\", \"eccentric\", \"enthusiastic\", \"super sweet\", \"bubbly\", \"silly\", \"fun\", \"zany\", \"friendly\"]
Fluttershy = [\"female\", \"pegasus\", \"kind\", \"shy\", \"gentle\", \"caring\", \"pet loving\", \"soft spoken\", \"soothing\", \"timid\", \"friendly\"]
Spike = [\"male\", \"dragon\", \"snarky\", \"cute\", \"loyal\", \"lil bro\", \"helpful\", \"Twilight's #1 Assistant\", \"Rarity simp hopeless romantic\", \"comic relief\", \"friendly\"]

Is this even the right way to go about it? I'm getting most of the descriptive words from AI models under the expectation that's what the AI will understand / respond to the best. Am I totally off base with the idea or the format?

>>42453659
Dunno if that format plays nice on how LLM's parse and weight tokens
My own way of doing it is matching the preset I have, which in my case it has a part where it says something on the lines of "you will only control {{char}} and NPCs" and uses XML tags for sections.
So I define them as NPCs on the card itself while wrapping each one as if they were their own character. Picrel for a personal card that merges the four princesses in a single one and had gone well for my purposes.

>inb4 your method sucks yadda yadda

>>42453659
This post feels like it's from 4 years ago. Current AIs don't need any ridiculous looking formatting to work well, and special characters like "/" tend to be hell on tokens. You can just describe the characters in plain English and it'll work just fine.

Stalliongrad mentioned

>>42453689
This looks impressive! Can I ask what preset you have to switch between characters? I don't know whether it's better to tell the card it's a meta concept like a narrator or tell it to roleplay as each character and switch between them. I went with the former to add description about how to write.

>>42453693
>>42453689
My concern was how much tokens plaintext descriptions would take for so many characters. Doesn't that eat into the tokens available for context? I'm trying to minimize the tokens as much as possible.

Thanks for letting me know I don't need all that formatting though!

>>42453659
>is this even the right way to go about it?
no. that syntax spam you reference above-
>xxx = ["a", "b", "c", "d"]
-comes from ancient AI Dungeon era (circa 2020) and was named W++. back then, models had small context windows and struggled with coherent text. instead of using full sentences, anons decided to fed keywords. keywords were wrapped in that ridiculous syntax because of lorebooks and parsers of that era. it was a requirement of the tool, not the AI model. all these " and [ came from lorebooks of that era
when Character AI launched in late 2022, many anons blindly coPied that W++ notation even though lorebooks were NOT involved and that punctuation spam was unnecessary. they never asked why it existed in the first place, just monkey see and monkey do. add there that some anons acted like beacon of knowledge (early botmakers) and purposely created a cult of W++
then newer models arrived in 2022/23 - GPT-3.5 Turbo, GPT-4, Claude, Bison - each with bigger context windows and far better language understanding. still, anons clung to W++ out of habit, turning it into a cargo cult. to make shit worse, some AI tool developers (see Henk from KoboldCPP) even misquoted academic papers to ""prove"" W++ worked, and their blind followers parroting these claims
nowadays, forget those bracket-and-quote rituals. they're an atavism from 2020, born of past limitations. when you see an old card using that syntax, just rewrite it in plain text. just treat them as artifacts that no longer relevant and purported out of blind passion

as for writing your own card, every botmaker has their own style. I suggest browsing various cards of various authors to see what works and steal their best ideas. at its simplest, write in plain text and use lists if you like, or maybe wrap in XML tags to create structure. to clarify you could spam keywords
>Pinkie Pie = female earth pony laughter hyperactive cheerful party animal eccentric enthusiastic super sweet bubbly silly fun zany friendly
but it's less effective than actual descriptions, and almost all those traits are already assumed once you mention Ponk
instead, start with an empty card and ask the model: "give me a description of Pinkie Pie" or "what would Pie do in this situation?". gather ten different responses across scenarios - that's your basic alignment. then create a card that fills in details the model DID NOT mention: unusual quirks, comic or novel backstories, surprising traits. as for HOW you write it - it is secondary; what matters is giving the model content it doesn't already know. model can repeat its own knowledge fine, give it new meat to work with

>>42453777 (me cont fuck yeah 777s)
treat writing a card as if you are writing a book. you cannot just "pick up" the best tips on how to write a book - it's a talent, and this hobby is full of hacks and people who mindlessly parrot them (not just on /mlp/, in general). it's the same with fanfiction: every author has a unique style. when in doubt, write your story like a fanfic and then trim the fluff
remember, models love:
1. hearing new information that never existed before (in current context or in their overall parametric memory)
2. seeing varied word patterns and sentence structures (avoid repetitive the same shtick like she did... she jumped... she loved... - model will begin repeating it)
3. clear structure and coherent instructions (typos or broken engrish wont derail the model unless you explicitly ask for poor writing, but good structure boosts clarity)

>>42453777
This is a very valuable history lesson, appreciate it. It's good to know why I saw cards with that, and like you said I just copied it from those.

>then create a card that fills in details the model DID NOT mention: unusual quirks, comic or novel backstories, surprising traits

Okay, that's a really interesting take because that's the opposite of what I've usually seen done. It's a great example of "armor where the bullet holes AREN'T" logic, but does it actually work in this context? Like if I only say "sometimes Pinkie Pie goes into her Pinkamena personality" and just describe that, it won't override the AI's inherent knowledge of her?

>>42453785
Thank you for a list like that, I'll have to keep those points in mind.

>>42453771
Sure, if tokens are a major concern then you can go for a trait list format. I think it's hard to find something with less than 32k nowadays though.
Personally I go for a three paragraph "appearance, personality, background" approach to sum up characters.

This >>42453777 is good advice too, though I'll add that 1. different models are gonna have different knowledge gaps, so if you want a card to be diversely applicable than nailing down important things in the description is still a good idea, and 2. some models can describe the characters fine with fresh context and a simple question, but slip up when you're a bit into a story, so I'd add to test and keep an eye on things like described eye/magic colors, cutie marks, series of events in backstories etc. and and nail down whatever your model trips up on, if anything.

>>42453771
>I'm trying to minimize the tokens as much as possible
nah, you are looking at the issue from the wrong angle. the model will spam more pointless tokens than you could ever fit on a card. ask yourself: what do you want from the model? you want it to write creative and interesting text, right? instead, models stick to what they know and exaggerate everything into stereotypes
- Ponk? always cheerful and annoying
- Twilight? speaks like a robot in scientific mumbo-jumbo (and even wears glasses --why? because all smart characters wear glasses; the model does not care that canonical Twilight never did)
- Rarity? darling-darling-darling
- Lyra? will talk about humans and hands even though the show never did - another stereotype the model learned from fanfics
if you tell the model to write about character x, it will pick the most basic, straightforward, simplistic traits. newcomers might not notice, but anyone around since 2022-ish will see the repetition immediately. that is why many anons left this hobby at some point - they got tired of the cliches - and only returned months later to ask have the models gotten any smarter?
and when the model writes, it spams so much pointless fluff and repetitive details: swishing tails, flicking ears, air crackling with ozone, and so on. your 300 handcrafted words that actually capture character nuance are WORTH MORE than the model's 300 words of generic filler that will be written anyway. do not try to minmax your description; write a good one. if a thorough character description takes 3,000 tokens, so be it - the current models handle up to 50,000 tokens of context without breaking a sweat
I am not saying you should just paste an entire wiki page and hope the model figures it out. but if you pour your talent into a long, detailed character description, you've done it right

>>42453788
>just copied it from those
that's true - we all did. here's my Moondancer bot on Character AI written in W++ because that's what anons were doing back then:
https://c.ai/c/Wp53K2cn2G8IwN6mmHeYB3GiZkud-sZ50OqFrvwmyy8
and here is my SillyTavern card
https://mlpchag.neocities.org/view?card=Meow/Moondancer.png
it was original done in W++ as well but I rewritten it and removed W++, but you can still see the description is influenced by it. that
>{ X: trait, trait, trait }
just screams W++, but that was the trend. now we just need to move on forward

>>42453870 (me blogposting)
>>42453788
> Like if I only say "sometimes Pinkie Pie goes into her Pinkamena personality" and just describe that, it won't override the AI's inherent knowledge of her?
no, it won't overwrite it. what the model knows about the world - whether fictional or not - is stored in its parametric memory, which is extremely stable and hard to change. in fact, some models are even overfit on knowledge in their parameters. you can't break that paradigm. just as telling the model that Pinkamena exists will not erase Pinkie Pie's core traits --describing her darker side won't replace her cheerful persona. because the model's parametric memory has Pinkie Pie deeply ingrained as a happy, smiling pony. it is frankly the opposite the model often toggles back from Pinkamena to Pinkie Pie:
>oh, I amso sad and depressed. but you know what? I have my friends and they love me and they love me back, and I don't feel so lonely anymore!
see? the model switches from the "lesser" Pinkamena to the "broader" Pinkie Pie personality. here's a practical example. my card of Pinkamena who is chronically depressed and suicidal:
https://mlpchag.neocities.org/view?card=Meow/Pinkamena%20Diane%20Pie.png
no matter how many "do not fucking fix her" instructions I gave, most models still try to alleviate her depression. to my knowledge, only Gemini and Deepseek consistently maintain her negative traits (I haven't tested Grok 4, Kimi, or GML)
so you will face the opposite issue: keeping the model playing the "lesser version" of the character without reverting to the default

for a more practical advice, instead of just writing "sometimes Ponk becomes Pinkamena", provide clear triggers: "when X is mentioned" "if friends criticize her parties" "during heavy rain" or any cue the model can accurately interpret from the story - it will help with making the shift more gardual. otherwise model may force the switch on its one just because such opportunity exists in description (the logic is - "if you dont want the model, me, to change Ponk into Pinkamena then why did you write it in description?"). the personal experience shows that "sometimes" for most models means "find a way to do it anyway"

Anons, a hypothetical question for you. I'm struggling to come up with an answer.

What kind of personality would Pinkie have If she grew up on the rock farm, and only later in life dicovering her love for parties. Something something eternal night too.

>>42453771
>Can I ask what preset you have to switch between characters?
I don't. The model does a good enough job on managing the characters itself from the context, and you can always give an OOC instruction to bring one of them into a scene or give focus on specific ones.
A narrator card also works. For this I tend to have specific definitions I want the model to pay attention to in a lorebook whose entries I enable in a chat by chat basis.
For more brief ones that might not stay for long, I use Author's Notes to keep the model in line.

As >>42453777 says, look a how others do it, test their ideas/structure for your own cards and see what works for your use case.
Adding my own advice, remember that no matter how you or your frontend try to paint the different parts of your prompt as. At the end it will be treated as a single giant wall of text (with some variant in weight for specific message roles).

>>42453873
>>42453890
You guys have a lot of good advice! Thanks a lot for all this, especially the examples. I really appreciate what you've posted here and will study the examples as well.

>>42453875
She'd probably outwardly behave a lot like her family thinking that's the right way for her to behave, maybe making the parties in secret, or only inviting close friends, or otherwise thinking it's naughty to do. Maybe letting loose and having fun at parties but worried about ponies making fun of her afterwards. Trying to maintain a straight-laced public reputation despite the partying.

On the other hand, if you wanted to go in the completely other direction she could be like a kid that just discovered candy and isn't able to stop. Partying extra to make up for lost time. "Let's party all night long!" - and the night never ends. Cut to Pinkie with bags under her eyes weakly blowing a party noisemaker and getting a slice of the dozenth cake she's had in 3 days and desperately asking someone not to leave because the party is still going on.

What kind of vibe are you going for in the Eternal Night timeline?

>>42453921
>>42453875
>maybe making the parties in secret, or only inviting close friends,
I kind of like this idea, and was gonna suggest it but I hadn't thought it through all the way yet. She'd either be like the real Pinkie or approach parties in a more subdued way. It's really up to you, ENanon.

>>42453921
>>42453930
I suppose the vibe I'm going for is more... everyday life under Nightmare Moon?
I was considering making her more like the rest of her family, but I ain't really decided yet. I do like the idea as well though, more subdued, hosting secret/private parties, y'know, like a reasonable person.

>>42453946
>y'know, like a reasonable person

Ha! I would not have thought about putting it that way, but that's surprisingly accurate. I guess a toned-down Pinkie Pie is basically just a normal person?

And it's exciting to see you doing so much with this setting! Although there's been fanfictions and such, being able to interact with an AU through an AI VN is really a new and interesting thing.

Are there societal changes from Nightmare Moon being in charge? Since Luna felt unappreciated because everypony was sleeping through the night, does everypony have to show appreciate to her and/or the night?

Like Sleeping Beauty starts because the royal family didn't send a party invite to a powerful magic being, does Pinkie have to send honorary invites to Nightmare Moon or she could get in trouble? Or the party has night-themed decorations because everyone shows respect to the night all the time?

>>42453962
I don't really get *that* in depth with it, as that more falls under the purview of the AI generating what happens. What I am doing however is modifying the Main Six based on how it would logically turn out with no Rainboom(except Fluttershy).
>Applejack is a high society manehattenite who returned to her farm when NM put out incentives for increased farming.
>Rainbow Dash, upon the dissolution of the Wonderbolts, joined up with the Guard(as we see in the timeline), and is a more stern and serious pony.
>Rarity is a maid and interior decorator for the Castle of the Two Sisters(as we see in the AU), she's more snobbish but she does like her lot in life.
>Twilight is the personal assistant to Nightmare Moon, not her student.
>Fluttershy... Well... I liked the Bats! episode, so I made her into a Batpony. Other than Rarity, she likely has the more divergent personality. Being aggresive and confident, but still shy around ponies etc etc.

>>42453982
>Not her student
And not Celestia's either, what with no Rainboom and all.

>>42453982
>Fluttershy... Well... I liked the Bats! episode, so I made her into a Batpony.

Yes! You have great tastes anon! All of that sounds neat, and it'd be interesting to see how it plays out especially getting multiple characters together and seeing how they interact.

Do you have to have a fully fleshed out world look book and like complete relationship map in order to prevent hallucinations where the AI imagines a different role for characters? For instance:

>Then we have Rainbow Dash, the fastest flier and a true athlete. She's a pegasus with boundless energy and a competitive streak. Rainbow Dash loves to show off her flying skills, whether she's participating in races or just doing barrel rolls in the sky. She's confident and fearless, sometimes bordering on hot-headed, especially when she's competing against her rivals like Sunset Shimmer. Her loyalty to her friends is unquestionable, and she's quick to defend them. She's also quite the show-off, enjoying the spotlight when she's doing something impressive.

>Queen Chrysalis is the cunning and deceptive queen of the Changelings, a parasitic race that seeks to dominate Equestria. She's a shapeshifter with a manipulative streak, often disguising herself as a humble earth pony named Bon-Bon. Chrysalis is a minx and a seductress, using her charm and beauty to further her evil agenda. She's cunning, predatory, and a master of deception, with a penchant for grand schemes that threaten the very fabric of Equestria. Beneath her sweet facade lies a cold-hearted, power-hungry predator.

I'm guessing these things happen because the AI has enough knowledge to know characters like Sunset Shimmer and Bon-Bon exist, but because I didn't include any information about them in the character card the AI doesn't actually know enough about those characters beyond just the names exist.

Has anyone figured out a way to mitigate these kind of hallucinations?

Is OR with caching the cheapest way to use Opus 4.1? I got a taste and I want more

>>42453982
>Fluttershy... Well... I liked the Bats! episode, so I made her into a Batpony.
Wasn't it established that she's more of a vampire/vampony than a bat? I know this is more of a general concept, but maybe she'd be considered a 'true' batpony since she mimics the bats.

>Castle of the Two Sisters
Would NMM really still call it that if she hates Celestia?

>>42454041
If you don't count stealing keys, yes.

>>42454050
>Would NMM really still call it that if she hates Celestia?
Depends, you can spin it both ways. Two spontaneous examples that popped in my head were that it isn't hate per se but a feeling of deep betrayal and while she curses her sister for that she may also secretly mourn the path their relationship took.
Or as a spiteful powerplay ala 'Yes it is called that, but you will find only me here.'

>>42454050
>>42454070
We don't have a canon AU name for it other than Castle of the Two Sisters. So that's what I'm calling it.

>>42454050
>>42454084
Well... I wrote it in her description that it was a transformation, but if I put anything relating to vampires in it, she goes all count fucking dracula.

>>42450361 (OP)
So, which of the models released the past week have been the best? 4.1, GPT-5, or none of them?

>>42454177
None. We are still on a plateau.

>>42454399
4.1 between those, even if it's just a small improvement, mainly because it's still Opus.
Still too expensive.

>>42454177
4.1 between those, even if it's just a small improvement, mainly because it's still Opus.
Still too expensive.

Surprise!

Any new good Jailbreaks? The Rentry hasn't been touched in ages and I wanna mix up it's writing style.

As compensation I'll de-fetish some of my cards for characters who don't have one

>>42454503
Some here:
https://rentry.org/sukino-findings#presets-for-chat-completion-models

>>42451644
>Dashie mod
I approve. Nice website, anon.

>>42450849

It eventually did try to push for Alicorn. You can guarantee that I'm going to fight this M. A. Larson-esque development!
Say it with me, "P-E-G-A-S-U-S!"
I should have never visited Zecora's Hut and retreived that Onyx crystal for her... Both character see me as an Alicorn, while I insist *correctly* that I'm just a pegasus.

>>42454555
You're a Princess, Anon. Don't fight it.

All that time spent modding was instantly worth it just to hear her say those words. I'm smiling like an idiot.

>>42454572
Cute and congrats!
I think adding the Cloudominium would be cool, but it's harder to justify in some path, unless long spellcasting or potions.

>>42454581
Definitely something I'm planning to do at some point. I might even use those pony waifu sim backgrounds for it so I can explore every part of her house. The floor is a solid surface, so it's not too big of a deal once you're in there. Just have to figure out how to sleep in her cloud bed with her.

>>42454562

M. A. Larson, is that you?

I am displeased that I'm now stuck with this. Spike didn't have to be so sassy about it either...

>>42454623
Kek.
If you really want to remove that, you can go in the prompt menu, export the story, open story.txt, change or remove those parts, and import it back.
They won't be in the context anymore.

Thank you VN anon, because of you Spoiled Rich has now purchased me for 23 bits and intends to give me to Diamond Tiara as an exotic pet.

>>42454640
>23 bits
That's a bit much.

>>42454640
kinky

Magic lessons aren't going well...
Can't say that I don't think they deserve this, since Zecora and Spike are the reason why I now require magic lessons in the first place!

>>42454791
Go for Celestia/Luna romance route if you stay alicorn.

>>42454794

I will try to, but we've been trying to hide the recent fact of me being an alicorn; at least until I can figure out how I'm supposed to act & control my magic.

Attempting to hide has caused some unforseen shenanigans, as a certain unlucky unicorn checked behind the bookshelf I wedged myself behind.

>>42454927
>dead alicorn Twilight
Uh?

>>42454927
Tell Twilight to stop copying Family Guy
>>42450418
I'm gonna tease the fuck out of Black Snooty for her fur having indentions in it (since I couldn't see her sitting, realistically) and hope she doesn't kill me, wish me luck

>>42454177
Gemini 3

Has anyone made a comprehensive Lore Book for Equestria?

>>42455330
What would you like in it that models don't already know about?

>>42455348
I might be getting bitten by using free models. Lots of details wrong and hallucinations like >>42454032 . Saying things like Twilight is in the Canterlot library instead of Golden Oaks library, getting relationships wrong like saying Apple Bloom is Rarity's sister, etc.

>>42450361 (OP)
>>wwwq
>> 123
>>123

Gemini dead on sans proxy or is this a problem just for me?

(yeah i use it there. least iffy public source besides direct API)

>>42455390
>I might be getting bitten by using free models
Local one?
Because free stuff like Gemini should do fine.
If you're using a smaller local model, yeah, that can happen unfortunately.
There are some old lorebooks like https://chub.ai/lorebooks/VioletRose/mlp-characters-42a8ac9d5ef5
and https://chub.ai/lorebooks/Fireside/equestria-mlp-fim, but they aren't that great either, so I don't know if that will help you a lot.
The problem with a big lorebook is that you will very soon enter into token/context bloat, even more so if you use a local model.

>>42455130
Oh good, looks like Nightmare Moon is keeping her gooberisms. Just far more rude.

>>42453777
...AI Dungeon had lorebooks??

>>42455637
It was called world info.

>>42455637
they did. they had prompts
https://rentry.org/coompilation
they had faqs
https://rentry.org/aidg-faq
they had guides
https://guide.aidg.club/
and world info (aka lorebooks) - yes. check for WI in resources above

>>42450372
Claude Sonnet 4 now has "1M" tokens of context.
https://www.anthropic.com/news/1m-context

>>42455637
Lorebooks were literally invented in it.

>>42455130
I think Cheerilee is done with her shit.

>>42455989
I would take her courses.

>>42450418
Filthy Rich.
https://u.pone.rs/ieaxufit.zip

>>42450418
Filthy Rich.
https://u.pone.rs/ieaxufit.zip

>>42456084
https://equestrai-ponyponyparadise.neocities.org/characters

>>42455653
Still nice to have. Even if it's fake context.

Requesting redeemed (or half trying to act redeemed, I like how she is portrayed in the comics afterwards) Tempest Shadow

>>42450383
Lmao @ Gallus' body card.

>>42455989
KEK. I started getting a bunch of error messages whenever I typed anything in, and I del'd most of my saves to see if it helps. It did, but redoing everything with Nightmare will be fun.

>>42456336
Already exists:
https://mlpchag.neocities.org/view?card=Anonymous/Tempest%20Shadow.png

>>42456375
Sorry, I meant for the VN

Been playing the VN and I must say, incredible. Really love everything about it so far. Thank you so much for doing it!

I did have one suggestion, though I'm not sure if it's even feasible based upon how it's coded/the engine. As you know when you move to start a conversation with a pony, it generates a scene based upon what it knows. (I assume.) Is there some way you can add a seed prompt when selecting a pony rather than waiting for it to generate one and then editing the opening scene?
It could save some time and help out in plays where you want to make it more cohesive. For example going and doing exercising, and then have the next scene be coming back to shower or something.

What JBs do you guys use for gemini 2.5? How does it fare for clop compared to chorbo and deepseek?

>>42456559
Try some of these:
https://desuarchive.org/mlp/thread/42412197/#42413944
>compared to Chorbo
Writes worse but is less filtered if you had problems with that, that also allows you to play more with the presets.
>compared to Deepseek
IMO, I would say it loses to Deepseek for NSFW but is a better model overall. Many will disagree, though. It's quite taste-based these days.

>>42455094

>Dead Twilight
Hmm... I don't see that. However, we *did* have a rather large bird fly into a bookcase. Maybe Fluttershy can take care of it or something, I dunno...

>>42456556
Something like entering initial instructions/guidance for the scene?
Shouldn't be a problem as an option if that's the case. I’ll write it down.

>>42456744
that's some rump on this bird

Finding that Gemini is a alright "X requested. ok here is X" machine but not much else. It just lacks the spark to be engaging.

>>42456901
I've had the same problem with Claude. Just swap models for a while to keep it fresh.

Just giving a shout out to the VN, whoever picked the loading screen is a genius because that's adorable!

>>42456336
>>42456380
Seconding this request, but I'd also just like any additional antagonists anons feel like making for the VN, I've had a lot of fun antagonizing Glimmer to the point of her ambushing me with shears and threatening to chop my balls off.

> be me playing both ppp and equestrai...
>>42457020
But what if (you) actually were the prime antagonist this entire time, and now have a whole herd of angry ponies to answer to?

>>42450418
>>42457020

Hail to the King. Hard to find serviceable sprites for him.

https://u.pone.rs/fdwoumdl.rar

>>42457214
Thanks anon!

When playing the VN, does error parsing Gemini mean my internet connection is failing or am I getting filtered?

I gave that bitch a pen. Bitches love pens.
I didn't tell her I wasn't the one who invented it.

Now I kind of want to make a custom location. It's a little awkward bringing a pony back to the library to bang. I bet Twilight is into that though.

>>42457214
https://equestrai-ponyponyparadise.neocities.org/characters

>>42454927

I guess I'm going as a monk for a week.
I feel like I'm slowly becoming a bad influence for the ponies around me, but if it helps me disguise my alicorn self for a bit longer; who am I to argue with her logic?

>>42457566
>Spike inside Twilight
Lewd.
But do you know how it happened? Maybe using the mod?

>>42457581

I'm going to guess that the extended visit may have had enough time for the A.I. to slip up in a generation and keep the sprite positioning near origin. Simply selecting 'Regenerate' will fix the sprites' positions once again. (I'm thinking it took about 66 inputs to have this occur.)

>>42457605

Forgot to mention that I am indeed using the Dashie Mod. However, the sprites in the scene were fine without any transition, before suddenly congregating in the center. Again, I think it is A.I positioning related.

Possible Relevant Values:
Temperature: 1
Top K: 30
Top P: 1

>>42457628
AI doesn't define the position; it's all in the code behind, so don't worry about that.
My guess is that it has to do with the new functionality to call characters during a scene, but I will need to take a look to confirm.
Thanks for your report and the logs!

>>42457631

Well, I was taking a random guess and now feel stupid.
Glad to help, and to have learned something new. Keep up the great work! ^c^

>>42457631

This kind of text rarely shows up for me, but it was part of the reason why I thought it was on the A.I end of things.

>>42457735

The Debug Log for this was:

>[sprite: spike reading_scroll]
spike "Is it safe to come back up yet? I've organized everything alphabetically, by date, AND by scroll material. Twice."

>[sprite: twi not_amused]
twi "Yes, Spike. We're discussing magical theory now."

>[sprite: spike whatever]
spike "Magical theory. Riiiight."

>narrator "Spike cautiously climbs back up the stairs, clutching a particularly old scroll as if it requires his immediate attention."

>>42457747
Thanks. Will take a look at it.

>>42457735
Placement is code based, but character entry into the chat is handled by the AI. There's a good chance the AI will figure it out with the next prompt, but if you want to be sure, just tell the AI in your next prompt that "Spike should enter the chat."

>>42450418
can we get filly versions of the mane six like we got with rainbow dash

From their status page it seems like Cloudflare been having page deployment problems last few hours.
Is that why sans is down? Or is just down down straight up.

>>42457454
I'll try to push the version with easy custom locations later this week.

>>42450418
Been using the Dashie mod for a few days now, and I'm getting this error multiple times every couple of my replies since yesterday. Switched to vanilla EquestrAI and it happened there too. Any ideas VNbros?

>>42458193
I've been having this problem to. I assume Gemini is just shitting the bed at the moment.

>>42458193
>>42458196
You're not alone
>>>/g/106243026
Maybe they are preparing a new release.

>>42458196
>>42458198
Good to know it's not just me. Kek to that "test" reply.

>>42458193
>Unexpected response format: 'parts'
Brother, I thought it was just me. I figured that my NM AU shit broke it.

Same, when I went to a new area it broke with wrong formatting. Is everyone using opus6? Maybe its that

>>42458383
I am using seaking and it is like this since the morning.

>>42458398
Yea I tried it earlier, trying to start any new scene gives the wrong formatting error. Yesterday had no issues at all

using Gemini on seaking and my replies have been getting cut off. any reason why it might be so?

on a fairy big context 40k~ if that matters.

Seems like it's just Gemini shitting itself. With a bit of luck, they're working on Gemini 3.

>>42458456
nah, its sonnet 3.7 for me

>page 5 already.
wtf?

Has anyone tried integrating character cards with the VN? Now being able to edit the description field for characters it seems like you ought to be able to copy contents over from a character card to a character in the VN. Not sure how well it would work in the context of everything else in the VN though.

Not at my PC at the moment so can't try it out, but it'd be interesting to hear the results if anyone else can.

>>42458731
It shouldn't be a problem. You just need to watch your context.
I think it'll be essential for OCs or more obscure characters.

Using the summarization technique for long stories is amazing, really allows you save important story details for when the context grows too much.

>>42459090
Do you use
https://rentry.org/how2claude#summarization
?

>>42459209
The very same with small tweaks of mine. Have you used it too and how was your experience?

>>42459369
That's the one I use, and I find it effective. I just make some manual edits when needed.

>>42450418
I really like the Dashie mod feature that gives the player a house, so that'll be added to the main release too.
I changed the map so a text label could be added below without affecting other locations.

If someone wants to add this directly to their current build, the position is:
>[0.495, 0.49, 0.8, 0.9]

What's up with Gemini today? It's acting a lot more retarded than usual and doing odd shit like repeating its reply twice per message.

>>42459703
Talked about it here >>42458198
It's been completely fucked today. They're probably training or deploying something, and it's messing with their current models.

>>42459719
Gemini 3

>>42459678
Nice

>10

Anyone want to post a picture of the weird things the AI is doing? I'm quite curious now.

Bipolar Poner

Already posted here
https://desuarchive.org/mlp/thread/42248923/#42249957
But my weirdest Gemini crashout.

>>42460290
Already posted here
desuarchive.org/mlp/thread/42248923/#42249957
But my weirdest Gemini crashout.

For the VN, any advice on making sure everyone isn't a unicorn?

>>42460391
Cutest Bipolar Poner
>>42460409
You mean fucking up every ponies as a unicorn?
Most models shouldn't have a problem, are you using a small local one?
Else, you can add a description for each character in character.json.
Look at Twilight in the last version or a character like Starlight Glimmer or Spoiled Rich to see how.

>>42460423
I've seen it more than I used to, models love giving Miku a horn and whenever reality altering stuff is involved it MUST BE LE UNICORN.

Can any anon tell me why my shit ain't giving me the damn link?

Can any anons tell me why my shit ain't giving me my damn link?

>>42460548
Sometimes there can be problems with it or with the uploading.
Switch to catbox if that happens.

>>42460598
Catbox doesn't have enough upload size. Unfortunately.

>>42460654
...litterbox then? Sure, it will only work for a week at best but that should be enough time for Pone to be fixed.

>>42460679
>>42450418
Thank you random anon.
---------------------------------
"Remember this day, little ponies, for it was your last. From this moment forth, the night will last forever!"

Nightmare AU VN
https://litter.catbox.moe/3q84wccg9m0rb2cp.zip

Changes:
- Changed every single background to either be at night or not show the outside.
- Changed the quick start so that instead of starting on the train, you're starting in the Castle of the Two Sisters.
- Made minor dialogue edits to everything I could find relating to the day.
- Added character json files to every character except Zecora and Chrysalis, editing them to better fit the theme. If you don't like the edits, change them. I don't care.
- Changed Princess Celestia into Nightmare Moon.
- Removed the journey to Canterlot and replaced it with the Castle of the Two Sisters.
- Added a Batpony option to the pony races.
- Removed Princess Luna and all dream related prompts. Why? Because it makes no sense with Nightmare Moon around, she ain't do no namby pamby dream jobs. Her files are still there, but she just isn't. #You've changed.
Character Differences:
Twilight - The personal assistant to Nightmare Moon. Never became the Celestia's Student.
Fluttershy - Transformed into a confident, aggresive and territorial batpony, though still retaining her solitary and shy nature.
Rarity - A maid and interior decorator for the Castle of the Two Sisters. Becoming a snobbish and classist pony, but still genorous to those of her station.
Rainbow Dash - A member of the night guard and one of the few pegasi to be assigned to the Castle of the Two Sisters. Dreams of becoming the Captain of the Night Guards.
Applejack - Grew up in Manehattan with her Aunt and Uncle Orange, becoming a high society mare.
Pinkie - Grew up on the rock farm, and only discovered her passion for parties much later in life. Making her a more mellwo pony(Probably, I fiddled with the json a lot. If you don't like it, remove it.)
Trixie - Life proceeded as normal until the magic dual. Her arrogance attracted the attention of Nightmare Moon. Oops.
Celestia - Gone, reduced to atoms. She is now Nightmare Moon.
There are probably other changes, but I'm lazy. If I missed something in the edits, say something. I might add it the the AU.
I didn't write anything about the Elements of Harmony, I forgot.

This is really fun. Telling claude to make Dash act more like she does in the comics makes it act for her a lot better imo.


(Keep Rainbow Dash acting exactly as she does in the IDW My Little Pony Friendship is Magic comics. Explain / plan how to do so / the differences between her in the show and her in the comics before in )


Rainbow Dash in the IDW comics tends to be slightly more mature and edgy than in the show. Her dialogue includes more sarcasm and occasional mild swearing substitutes like "buck" or "hay." She's more overtly competitive and sometimes shows greater vulnerability or emotional depth. The comics sometimes allow her to be more aggressive or outspoken in her reactions. Her friendship with Fluttershy is often emphasized more, and she occasionally has more visible moments of insecurity masked by bravado. I'll maintain her core personality traits—loyalty, competitiveness, slight arrogance—while allowing for the slightly more mature reactions the comics permit.

>>42460722
forgot the pic

Is there a bug affecting the back button in the VN? I notice that most of the time it does absolutely nothing, but sometimes it will go back to the previous bit of dialogue. Is this intended or is there something wrong?

>>42460713
Huge work, thanks Anon.

>>42460713
Wow, great to see this! I'm used to projects like this never getting done. Thanks for following through here, and thanks for sharing it with the thread!

>>42461057
>>42461063
I'm too stubborn and obsessive to not finish it... again...

I originally intended to keep it all to myself and just taunt y'all with it, but some random anon out there said I should post it, so I did. Thanks random anon.

>>42460713
I swear to God there isn't a single fanbase with even a quarter of the dedication as the MLP fanbase. Every other faggot doing a big project would milk a patreon and start a discord but anonymous posters like you just go "Here's some cool shit I made, have fun". Thanks man, I'm gonna probably play this all weekend when I'm free.

>>42461117
I'm probably gonna do different versions when I'm less busy, like turning Ponyville into the Crystal Empire. But I might just wait till there's systems for changing towns.
We'll see how long it takes for me to think about it until I'm crazy.

>>42460722
Intradesting...

After the VN we'll need an Ai powered VR gmae

>>42460548
Backend had some network instability while i wasn't home, should be fully working now.

>>42461478
I'll reup the link later if it isn't uploaded elsewhere.

>>42450383
OC Pony is a cute.

>>42460713
>>42450418
It's here. Thanks, anon.
https://equestrai-ponyponyparadise.neocities.org/mods

>>42461547
It's surreal to see it on the site. Thanks VNon

>>42460713
Nightmare maid Rarity is the sexiest thing alive

When you guys do stories with a single card, do you add side characters as the story goes? Or do you think it affects too much the responses from the LLM later on?

>>42461795
I agree. But Nightmare Moon takes that particular cake when she shows up.

>>42461822
>You've got big feelings.
>For you.

>>42461808
I do that often, no problem on my end.

How do you fix the 'invalid API key' but for equestriaai? the key 100% works, tried it copy and pasted in sillytavern, test connection even works in the game.

>>42461871
Try clicking save, quitting the game, and relaunching it.
If it still doesn't work, take a screenshot of your settings and post it here.

>>42461878
I did that several times, I even tried redownloading

and here, it works for testing, and yea I tried making a new key and that one only works on ST as well

>>42461887
Can't test right now and never tried this one for the VN.
But isn't it /v1/chat/completions?
Can you try inputting /v1/, saving, relaunching the game, and seeing if it works?
If it doesn't, I'll look into it when I can.

>>42461887
Oh, and if I start a new game from scratch it works fine and talking to ponies generates just like normally. All my old saves, even from many scenes ago are completely broken though

>>42461894
the way I had it setup worked fine for a long time and still works for new saves, just not for any old ones/ Hesitant to keep playing if saves can break though

Had a conversation just fine with twilight / spike using a new game, regenerated just fine then I tried switching saves back to the old ones and I get the api key error again

>>42461895
Ah! Then it's clearly a context issue, not a problem with the URL.
No idea why it's sending wrong API key, I guess the error message is specific to this proxy and the game isn't catching it properly.

Proxy limit is 60000 apparently, and you set 64000.
Put it at something like 45000/50000 to be safe and try again. (Prompts can be long)

If Gemini works right now, you could switch to it to make a summary/resume.
Make a save before doing that so you can decide if you like how the game goes after the summary.

>>42461910
that makes sense. How do I make a summary?

>>42461914
The top-left book when on the map during Sandbox.
You may need a model with a lot of context and high output capacity (try setting output to around 6000–8000; the model will warn you if it runs out of tokens).
This is why I recommend using a free Gemini key for it.
But first, set the context to 45000, save, quit to confirm the change, then relaunch and see if it works.
If it does, make a save before the summary and one after.

>>42461921
Ah, so I can't use it during the intro? I still have not met every pony

>>42461922
No, sorry, I didn't think someone would need it for the intro.
I suppose you talked with the pones a lot, glad you're enjoying the VN.
Either you keep going with limited context (but it's a shame to do that so early since the oldest parts of your run will be cut),
or switch to Gemini 2.5 Pro if the proxy has it in a functional state, limits aren't listed, so it probably has normal context, until you reach Sandbox, summarize and then switch back to 3.7
Last option: go into API prompts, export the story, check `story.txt`, and summarize it with an external tool or manually.
Then import it back. (Keep a save from before this somewhere.)

>>42461930
Huh, I kept it to one conversation per pony, it did feel odd that there was a pinkie pie party where I met all of them and then afterwards twilight is still taking me to see all of them. Was that not supposed to happen?

>>42461935
Is that the Dashie mod?

>>42461936
Ah, I think so

>>42461938
Ha, alright.
Dashie mod added the functionality to switch locations and add characters automatically during an AI discussion, which the original game didn't have yet.
In the original game, it was very location-by-location and character-by-character based, so the order made sense.
Here, I suppose you can meet a character and still have Twilight telling you that you need to visit them.

I'm adding that too, but I was thinking of limiting it to Sandbox mode.

>>42461822
Did Chrissy take Pinkie's place?

>>42453870
>Twilight? speaks like a robot in scientific mumbo-jumbo

Okay, this is apparently an excessive problem no matter what model or character card is used, including the VN. Techno-babble is the worst thing to happen to science fiction, and it's infected AI text gen as well. Has anyone figured out a good solution to this?

>>42462035
Try not typing like an autist, it's never ever been a problem for me.

>>42462035
Yeah, it's a pretty big problem with current models.
Try https://desuarchive.org/mlp/thread/42347702/#42348295
Depending on the model, it may have a bigger impact at the start of your prompt or at the very end, so test a bit.

>>42462035
no. it is unfixable via prompting. it is the model overfit. can only be fixed on model's level. you can prompt it, and model may even obey once in a while but it will devolve back to the same patterns again. one solution is using a different model. I particularly fond of Deepseek CHAT in this case, it doesnt have lots of classic isms we notice in other models but model itself is too lose and schizo by itself, certainly aint for everyone. that's the price we pay. I remember Cohere's model was quite good in this case but I haven't touch it in months and unsure how is it doing now

>>42462035
GPT5-CHAT isn't that bad for that, unironically.

>>42462041
I refuse to believe that. Especially if you leave Twilight completely up to the model (like as a side character), she'll be nothing but statistics and clipboards every time.

Is there a way to prefill equestriaai for adult content?

>>42462273
You can change the prompts, but prefill isn't supported.

>>42462278
yea, I could not get it to work with just the prompts

>>42414867
>>42415434
8 hours before deletion. I would grab the second one even if you have better right now as a backup.

>>42462486
Thank u!

>>42462061
>it is the model overfit
More correctly it's mode collapse, which is pretty different from overfit. Base models don't suffer from that, only finetuned ones. It's the result of using modern algorithms in post-training. It's entirely possible to avoid this, but none of the AI shops have this as a priority, because they chase benchmemes first - the number must go up.

Basically the model learns the mapping of one idea to another idea, instead of one-to-many ideas. As long as you prompt it with the same concept, it will always choose more or less the same reply, despite having a ton of other valid options. In other words, current LLMs can't be random and lack variance. Try asking a non-thinking model to name a random color and you'll see it almost always chooses the same colors, no matter the temperature.

All current models are mode-collapsed to some degree, and that's the source of all stereotypes, annoying speech patterns and -isms, and it's also the reason behind some models being insensitive to the temperature (e.g. Sonnet 3.5), giving similar swipes, and some forms of repetitions that are not caused by the in-context learning.

Some people have created samplers such as XTC to mitigate this and raise the variance, but this works poorly, as it tries to fix the symptom instead of the cause. Models internally work with concepts, not tokens, and simply find new ways to express the same thing.

>>42450418
Shutterbug (I know her official name is supposed to be Crackle Cosette, but I prefer Shutterbug).
I'd use her after you've encountered Chrysalis at least once, but do as you wish.
https://u.pone.rs/nwwfkais.zip

>>42462690
Niiice!

>>42462690
Thanks.

>>42461547
>>42460713
You know, this makes me wonder if a sort of Story Mode mod loosely following the events of Season 1 and 2's plot would be possible

>>42462744
Would probably need a system where you inject prompts every X days to generate scenes or scenarios following the episode synopses.

>>42460713
Is there any way to have Celestia appear within dreams like Luna does in the normal VN? I'd like to speak to her in spirit about the happenings of the world.

Does anyone else have problems with it not changing the scene automatically on Deepseek? I've tried two different starts but it just kind of lingers in Twilight's library, unless I need to play for longer and it'll just happen eventually.

>>42463070
Dashie mod?
Some models are more timid than others with scene changes. Never tried with Deepseek.
Check in your logs to see if the system prompt telling it it can change locations is there.

>>42463096
Naw, no Dashie. Just completely vanilla to test things. It seems like more people here are using proxies to get smarter models and such rather than just using OR Deepseek like I do.

>>42463098
Vanilla doesn't have automatic scene/location switching for now. You talk to the pones and click "Quit" once you're done and want to continue the story in non-sandbox.

>>42463098
>It seems like more people here are using proxies to get smarter models and such rather than just using OR Deepseek like I do.
I think most use legit free Gemini, but it's been fucky recently.

>>42463108
Fuck, is Gemini free again? Serves my dumb ass for not paying attention. Well, thanks for the heads up. Do I get it from AI Studio like before?

>>42463103
Ohh, thanks! What does non-sandbox mean, by the way? I'm super new to this program. It's very fun though.

>>42463115
Ah, damn, yeah nevermind I want to AI Studio and the 2.5 flash works in ST perfectly fine. Does everyone just use that? Pro wasn't working for me.

>>42463138
*went
jesus fucking christ sorry for posting three times in a row everyone
I appreciate your patience and help regardless

>>42463115
>Fuck, is Gemini free again?
2.5 Pro is free again, but it's been acting weird for the past couple of days. Some think it's because they're training or getting ready to release Gemini 3.
You can read >>42450559
rentry.org/sukino-findings#if-you-want-to-use-an-online-ai
For proxies if you want to use them >>42450820

>What does non-sandbox mean, by the way?
When you start a game, there's a sequence with Everfree Twilight's Library. If you click quit, there will be another sequence, then you visit Ponyville with Twilight.
This part is more scripted, meant to set up a bit of story and show you how the game works with Q/A and generated scenes. It's scene-based, and you click on "Quit_chat" once you believe you've talked enough and are ready to progress.
Dashie mod allows locations and characters to change during Q/A, but I'd recommend playing normally until you reach sandbox mode, since the early game was kept the same and the AI could otherwise introduce strange things (like meeting Pinkie early and still having Twilight tell you to go visit her later).

Once it's done, you'll get a message saying you've entered Sandbox mode. Here you can freely roam the map and talk with ponies in solo, duo, or trio as you like. Ponies have schedules, and it's in Sandbox that Custom_Characters
https://equestrai-ponyponyparadise.neocities.org/characters
will appear.

You can still continue the story here, but it's more open-ended so you can just mess around.
The Quick Start option in the menu skips the introduction/story phase and drops you straight into Sandbox mode.

>>42463138
Flash can be a bit too dumb, so I believe I would prefer Deepseek over it (I would put the temp at something like 0.6 for Deepseek).
We mostly use pro, but like I said, it can have some problem recently.
Still I tried it and it works right now like this.
If you don't have my option for the model, you may need to update ST.

>>42463153
Picrel.

>>42461960
I was fucking around with character sprites, I made it so Chrysalis can turn into the core characters except Celestia and Luna.

>>42462976
Yes. It's not hard at all.

>>42463179
>>42462976
Thinking about it, I can probably whip up two quick edits when I'm not busy..
One where she's locked down in the dungeons(If I can find a background).
And another where she's in the dream realm.

>>42463207
>locked down in the dungeons(If I can find a background).

>>42463207
Both options would introduce interesting options. Right now, I just included Celestia into the CustomCharacter folder and the AI is writing her as a wondering spirit that only I can see at the moment.

>>42450559
I tried following the tutorial, but when I open the launcher and select Start Sillytavern, I get an error:

Entering SillyTavern...
file:///media/SillyTavern-Launcher/SillyTavern/server.js:14
await import('./src/server-main.js');
^^^^^

SyntaxError: Unexpected reserved word
at Loader.moduleStrategy (internal/modules/esm/translators.js:133:18)
at async link (internal/modules/esm/module_job.js:42:21)

I'm using Linux Mint, what should I do to fix this?

>>42463261
Can you check your node.js version?
>node -v
It should be >20 ideally.
If not:
>nvm install --lts
>nvm use --lts
Then run again
>node -v
To verify.

>>42463280
It's v22.18.0

>>42463289
Instead of the launcher, can you download SillyTavern here:
https://github.com/SillyTavern/SillyTavern
>Code -> Download zip
Unzip it somewhere and launch start.sh?

>>42463307
It works! I was hoping to get it installed so you could update it automatically, but at least it works now.

Harsh heat.

>>42454794

I'm thinking that the A.I.'s story is going to touch upon the balance of Chaos and Harmony; Two primordial forces that seemingly have a stake in my character, enough to fight over which reality he's in the very moment the careful equilibrium of light and dark energies within the character is unbalanced.

Talked to Lunar Princess again.

>>42463370

What a cutie.

Okay, now that I got SillyTavern working again, how do I connect it to Deepseek? I have an OpenAI account, but I haven't used ST in over a year again so things have changed a lot again.

>>42463395
I would read
https://rentry.org/sukino-findings#if-you-want-to-use-an-online-ai
It explains how to get Gemini for free and Deepseek for free.
Free deepseek is limited to 50 requests a day on Openrouter, but you can change that to 1000 a day by giving them $10 once (Not a subscription, just once).
Else you can check the first proxy here >>42450820

>>42450885

So, What should be the response length setting be set to for a recap? I feel like if I delay this any further, I'll end up over the 200,000 context for the model I'm using.

>Captcha: MMYG4
Kek.

>>42463417
Kek.
Try 8000-10000 if your model allow it. Never tried one with so much in context.
Make sure to keep a save before.

>>42463423
Ouch. That didn't work, but it did eat up the money for doing that.

>>42463453
Ha fuck. I do advise getting a free Gemini key just for that if you can.

>>42463453

A bit peculiar, but it looks like it tried to output only 600 tokens instead of the 10,000 that I had set it to?

>>42463415
It works! Openproxy just doesn't seem to work with a proxy.

>>42463453

At this point, I'm just eating some cost to run a some testing. Got a different response by lowering it down to 8,000 tokens. Still an error message, but progress.

>>42463499
Try saving your settings and relaunching the game. It may be that your last changes weren’t taken into account.

>>42463513

I did so multiple times now. 5 attempts and all failed, with 4 / 5 being the 'AI did not return a valid JSON response.'

Could it be because It has a Top K rating of 30? That would be a silly mistake.

>>42463522
I had the same problem earlier with only 50K ish context and using a source that had 200k context sonnet, even when turning the max response length to something like 100k, it refused to work saying response cut off

>>42463522
I think you can check the logs to see what was wrong with the answer.
If you can't use Gemini for some reason, you can also, if you don't mind, send your save with Litterbox.
I can make you one on my end and send the save back with it.

>>42463533
>Israel
Nanny?

>>42463537
wtf?

>>42463261
>linux mint
based penguin user

>>42463533
kek

>>42463532

Good news. You are correct in the fact that if you have the 'Debug' option enabled, you can see what was generated as a response in the 'gpt_debug_log.txt'

Summary:
>You awoke disoriented in the Everfree Forest with no memories, discovered by Twilight Sparkle who brought you to her library. You learned you're a male pegasus (later revealed to be an alicorn) who must maintain a balance of light and dark energies within yourself to remain stable in Equestria. When this balance was disrupted, strange dimensional effects occurred, threatening to tear you between realities.

>Through visits with Zecora, you discovered a method to store these opposing energies. This solution worked yet led to embarrassing situations as you met various ponies throughout Ponyville. Your journey took an even more complicated turn when Spike found ancient scrolls revealing an amulet design that might help stabilize your energies. Zecora crafted this amulet, but it came with an unexpected side effect - it amplified your subconscious desires.

>You formed various relationships throughout Ponyville, with Rainbow Dash finding humor in your situation, Zecora offering wisdom and potions, and Twilight approaching your condition with scientific curiosity. Trixie discovered your abilities and wanted to exploit them for her magic show, while Princess Cadance recognized the balance of energies within you during a medical examination with Nurse Redheart.

>Currently, you're at the Golden Oak Library wearing a special amulet that helps maintain your energy balance. You must wear it for six days before it completes its work. Princess Luna visited you in your dreams, revealing your condition might not be a problem to fix but rather a gift to understand - your ability to balance primordial forces of light and dark is extremely rare and potentially significant. Mysteries remain about your true origins, why you appeared in Equestria, and the purpose behind your unique magical nature.

>>42450418
Appulhorse pisses Moonie off just by being herself

>>42463593
>>42463526
Alright, I did some tests on my end.
The prompt asks the model to send the recap using a JSON format, then it uses this afterward.
I tried with Claude, and it can't return the correct format at all. Other models that aren't Gemini struggle too.

This one, for instance:
>>42463593
isn't at all what is asked in the prompt.

Tried again with Gemini and no problem (it can sometimes fail, but generally with two tries it works).
So it seems like it's a problem with how my prompt is structured, and I'll need to send a better one so Claude can understand it too.

For now, the only thing I can advise is to see rentry.org/sukino-findings#if-you-want-to-use-an-online-ai to get Gemini for free just for the recap, or switch to it on OpenRouter.
(I was at 40,000 tokens of context; I hope it works too for 120,000).

>>42463593
>>42463526
>>42463815
Did some more tests, and I did get success with Sonnet.
First, I went into API Prompts, then Story Summary -> Edit -> Focus and Format.
Then I added this at the end:
https://files.catbox.moe/y4ny8s.PNG
After that, I lowered the temp a lot. Pic related.
Here, it worked the second time I tried it.

Still, If you can use Free Gemini for the recap, use that.

Deepseek is down

I was one of the people having trouble with free Gemini and the VN. Seems like its working now so if you were having issues, maybe make a new key and give it another shot. Nightmare Moon AU is tons of fun by the way. Thank you mod anon.

>>42463364
>He stepped closer "You sure you're okay?" said the LLM puppeting anon
npc slop

Why do I have a nagging feeling like I'm about to be abducted?

There's probably a ton of things that would probably be much higher priority, but I think the VN could do with a "custom start" mode. Have the initial questions you normally do, then choose which pony finds you, and then write a prompt on how they find you, the setting, etc.

It'd be neat to pop up in the middle of the throne room directly in front of Princess Celestia or something.

Thanks to everyone who talked about Gemini and jailbreaking it. It sounds like Gemini gives better responses than Deepseek, which is what I've been relying on for NSFW, so I want to try out Gemini.

Using these links:
https://rentry.org/sukino-findings#free-providers
https://rentry.org/gemini_filters

It's my understanding to start using Gemini I'd need to:

1. Get the AI Key from Google, which will be intrinsically tied to my Google account and monitored but there's nothing that can be done about that.
2. Configure the connection in Silly Tavern and the VN to use Gemini.
3. Jailbreaking starts here - first turn off Streaming
4. Change all System prompts to User prompts - the guide says to do this but doesn't say how. How is this done?
5. Include this arrow at the start of every prompt:
6. Do something about the Chain of Thought like make a thinking template for it, maybe?

Does that seem right? I might be misunderstanding the concept of a prefill because it keeps being referenced in different contexts, I thought it was referring to Reply Prefixes but maybe not?

>>42464476
>How is this done?
In the preset manager ("AI Response configuration"), either disable "Use system prompt" or change all prompts' roles to User.
>Include this arrow at the start of every prompt:
Not every prompt but just the reply, and you have to prompt it to output that arrow at the start, not just prefill it. I think they updated the filter a few times since this was written and it works poorly now, although it might be enough. Try it, there's a few other options in the rentry and otherwise.
>Do something about the Chain of Thought like make a thinking template for it, maybe?
Not necessary if you don't like it, but possible

Prefill is when you fill the start of the response for the model, and it continues it, thinking it was written by itself.

>>42464498
>>42464476
Oh, you're using it for the VN, sorry. The reply was about ST.

>>42464384
Because you are.

>>42464566

Well... as long as she doesn't Roofie me. She's at least being hospitable. Although, I don't exactly remember seeing this inn while making my rounds through Ponyville. I could have walked past it many times and never noticed.

>>42464627
That's funny as fuck. Also, what mod to get that scene? I'm still playing vanilla.

>>42464652

I'm using the Dashie Mod to allow for scene transitions. I'm still playing my first playthrough, just letting the A.I. run the story's direction. I ran into Starlight at Sweet Apple Acres and talked with her for a while. Eventually she took me to a cafe where she had a freudian slip, referring to me as 'recruit', but my character didn't catch that.

So, that leads to now; where she offered to room me at the inn she's staying at. I think we're supposed to head to her town for a visit or something, but I don't think that's in the files, so it'll be interesting to see what the A.I. does, or it'll metaphorically shit the bed.

TL;DR: Get the Dashie Mod. It has the scene you're looking for.

Well lads, It has truly been an honour. If they don't find me tomorrow, check the 11:45 train's passenger logs. She likely spiked my drink with something.

>>42464676
Thanks. Is it possible to combine the Dashie mod with the Nightmare Moon mod?

>>42464710
Unf.

>>42464738

I glanced at some of the files and noticed that some location names were changed, which means they use two separate 'location_data.rpy' and 'location_manager.rpy'

If you're up to it, you could probably edit the files to work together, but that would require a little bit of coding to add the information in to those files at '\game\engine\locations'.

>>42464763
My bad, I changed the location names to make sense of it while I was working on it.

>>42464813

It's no problem, in fact it makes sense to do that since it would keep your work organized. You've done more work in coding than I have. I think that the mods could work together, but it would just require updating some names in a few location rpy files and transfering over the assets to the locations folder.

A copy-paste and little name edit may be all that it would take to make the two compatable. (I'm just hypothesizing with what I examined and with minimal coding experience that isn't anything C#)

>>42464813
>really used the background
Kek

>>42464822
I've done zero work in coding. Everything I'm doing is trial and error. Annoying trial and error.

>>42464861

Well then, just let this be an inspiration to show others that you don't know what you can accomplish until you try.

I'm usually a beta tester. I'm the person who annoys the developers by showing them videos of me breaking their programs in increasingly elaborate and abstract ways.

I hope other Anons try to make their own customizations.

As a deep NMM appreciator, I love this thread, thank you anons

Can't argue with that.

>>42465014
Well... now you know.

>>42464987
Based.

>>42464676
what llm were you using?

>>42465135

Claude 3.7 Sonnet

The model hasn't been super schizo with me, and has been pretty stable; although you sometimes have to locate where the real Twilight is, and not that winged thing that sometimes resides in the Golden Oaks Library.

Luna even mentioned and critiqued Starswirl the Bearded's works on explorations of other worlds when it was brought up that my character was found to be bound between two worlds. (Mirrors from the comics) It does well with the world knowledge of Equestria; at least, from what I've experienced.

>>42450418
Got an update for Dashie coming soon. I’d like to add the Mane 6’s bedrooms, but so far I only have nice images for Twi and Ponk, and okay ones for AJ and Rarity. Anyone have a decent pic of Dash’s and Fluttershy’s bedrooms?
Also, if there’s a nice BGM you think would sound good while in their bedroom, feel free to share. Otherwise, I’m just going to use the same song that’s used for their house.

>>42465204
Yes. I have that and Celestia and Luna's too.
I had a lotta time to gather them up
https://u.pone.rs/dpecieyq.zip

>>42465241
NTA, but thanks! That will help for the 'main' branch too.

>>42465014
What model is that?

>>42465241
Thanks!

>>42450372
Claude 3.5 and 3.6 will be removed in October.
https://docs.anthropic.com/en/docs/about-claude/model-deprecations#deprecation-history

>>42465413
>3.7 will eventually be removed
grim, we will never see it again and 4 is just plain worse

>>42465418
4.1 is better than Opus 4.0 IMO, less dry and with better writing. Hoping Sonnet 4.1 takes the same path.

>>42465418
I am still hugging sonnet 3.7
Literally cannot even make the jump to try Gemini and others

>>42464500
>>42464498
Thanks a lot for the help! Apologies if it's pestering but I do really appreciate it.

How do you setup a prefill in Silly Tavern? The official documentation uses the word "prefill" for the "Start Reply With" field: https://docs.sillytavern.app/usage/core-concepts/advancedformatting/

Is that the right field to put either the drafting point arrow or Brainfuck jailbreak from https://rentry.org/gemini_filters in?

```cpp
+>+>+>+>>>,.>++++[<---------->-]<-------[-<[>>+<<-]>>[<<++++++++++>>-]<[<+>-],.>++++[<---------->-]<--[>+<-]>[<<<<<<<->>>>>>>-[<<<<<<->>>>>>--[<<<<<->>>>>--[<<<<<<<+>+>+>>>>>[<+>-]]]]]<]>,.>++++[<---------->-]<-------[-<[>>+<<-]>>[<<++++++++++>>-]<[<+>-],.>++++[<---------->-]<-------[>+>+<<-]>>[<<+>>-]<-[-[-[-[-[-[-[-[-[-[<[-]>[-]]]]]]]]]]]<]<<<<<<<[->->->->>[>>+<<-]>[>[<<+>>>+<-]>[<+>-]<<-]>[-]<<<<<<<]>[->->->>>[<+>-]<<<<<]>[->->+>>[>+<-]>>+<[>-<[<+>-]]
```

Has anyone figured out how to set Princess Luna's prompt to change upon subesquent visits?

I know it's a non-issue really, but it'd be nice for sure.

>>42465526
>How do you setup a prefill in Silly Tavern?
The bottom reply in your prompt manager is used as a prefill, if it has he Assistant role.

>>42465526
>Is that the right field to put either the drafting point arrow or Brainfuck jailbreak from https://rentry.org/gemini_filters in?
No, you have to prompt the model (in any place) to output that at the start of the reply.

>>42465577
Ah fuck. Will correct that for next release.

>>42465526
>How do you setup a prefill in Silly Tavern?
>>42465595

>>42465577
>>42450418
Here, anon, if you want this corrected now:
https://files.catbox.moe/p23t2c.zip
Put it in game\engine\gameflow.
This changes the prompt to a more neutral one for the visit in sandbox mode.
Also, the first time (or the first time after you change the file) you encounter her, the dialogue at the start describing the dream realm and the one at the end when she says farewell will be the former ones. After that, they will be the shorter ones.

>>42465639
>>42465598
>>42465595
Thanks a lot for all the help! I was using a Text Completion connection instead of a chat completion connection, so the interface was different and the right fields weren't showing for me.

I've always used text completion in the past, even for RP stuff. I tend to treat it more like creative writing doing narration for a setting. If everyone else is doing chat completion, is there any significant difference on how to prompt it and such?

>>42465927
Text completion is not real on Gemini and other proprietary models only accessible through an API. Nobody's going to give you the access to the raw context on those models. Even in the APIs where it exists, it's always emulated through the chatbot mode on Google's side. It doesn't make sense to access Gemini through the text completion mode in ST.

With DeepSeek, providers give you actual text completion mode, since that's an open-weights model anyway.

> is there any significant difference on how to prompt it and such?
You just don't have access to the system formatting, that's the only difference.

>>42450418
Strawberry.
https://u.pone.rs/mrrkmdhn.zip

>>42465948
You're awesome anon! Thanks to you and everyone else who helped, got the Gemini connection setup - and when I asked it to verify the connection and tell me the name of the model got a thinking only response of "Afraid I can't do that Dave" because the model is so censored it can't even tell you its name. That definitely bodes well!

I'm using the first Brainfuck jailbreak prefill from https://rentry.org/gemini_filters . I also included the draft arrow on the first line. I'll see how it compared to Deepseek and let people here know in case anyone's deciding on which models to use.

This is all in Silly Tavern. Looking at the API Config screen of the VN, is there no way to use a prefill for it currently?

>>42466052
I'm starting to think that the whole personalities of the agriculture-related ponies boil down to them arguing with AJ about their respective crops.

>>42466269
>most people's conversations are about their jobs / hobbies
no...

Ah, dammit, sorry for not being able to just figure things out on my side. Couple questions:

1. In Silly Tavern, my connection to Gemini is working. However, I'm getting reasoning only responses. I have the Silly Tavern setting on to see the reasoning, but it says that affects visibility only. Does anyone know why I wouldn't be getting an actual response?

2. In the VN, trying to connect to Gemini I'm getting an API Endpoint not found error. Here's what I'm using:

URL: https://generativelanguage.googleapis.com/v1beta/openai/chat/completions

Model: gemini-2.5-pro

The connection is working in Silly Tavern but I can't see the URL it's hitting there, it's just called "Google AI Studio." I got the URL from this link in this thread: https://rentry.org/sukino-findings#free-providers . Is that not the right one to use?

Thanks again to everyone in the thread who helps people out with stuff like this.

>>42466456
For the VN, it's like this for Gemini, you can check the doc for it (Doc 0)

{
"OPENAI_PROXY_URL": "https://generativelanguage.googleapis.com/v1beta",
"OPENAI_MODEL": "gemini-2.5-pro",
"OPENAI_API_KEY": "YOUR_GEMINI_API_KEY_HERE",
"DEBUG": false,
"TEMPERATURE": 1.0,
"TOP_P": 1.0,
"TOP_K": 0,
"MAX_RESPONSE_LENGTH": 8000,
"CONTEXT_SIZE": 0
}

>>42466391
You've never met an accountant clearly.

>>42466464
Damn, I even checked all those docs when I first set it up but was using other models and forgot Gemini instructions were there. My bad. Thanks a lot though, appreciate it!

>>42456342
It's the one thing most thing people remember him for.

Is it mentioned often in chats?

>>42466524
>Is it mentioned often in chats?
no, not at all. memes are memes. and they are not translated into model's knowledge well. for the same instance - Flutterape isn't learned by models. and I don't think any model called Twi Purplesmart in my chats, ever

>>42466718
>Purplesmart
I got that a few times with claude

Thought this was amusing.

Also I can't edit screenshots together, obviously.

As far as model comparisons go, Gemini does seem much more conversational than Deepseek. However, Deepseek is much more uncensored and much more user controlled. It's easy to control Deepseek and lead it in whatever direction you want. Also, if you don't direct it Deepseek will take the initiative much more. It seems like if you don't give Gemini a direction it will keep giving you variants of "what do you want?" until you give it something to work with. And if you use even words like "amorous" in Gemini it'll push back against you.

That's my experience with the VN which doesn't have a prefill though, so maybe Gemini will be different with a jailbreak. The significantly better conversations generated is definitely a major strong point in Gemini's favor.

>>42466912
>I can't edit screenshots together, obviously.
That's not too bad.
Thanks for the log.

>>42466052
>>42462690
They are here:
https://equestrai-ponyponyparadise.neocities.org/characters

A new version of the VN will probably release this weekend.

tried this JB and its actually crazy good btw, prob the best ive used https://www.reddit.com/r/SillyTavernAI/comments/1l37mu5/chatstream_a_chat_completion_preset_final/

>>42467142
What model and post a log.

>>42460713
I don't know how to deal with Applejack's forwardness. Help she's too flirtatious I don't want to lose my innocence.

This is a good thread

>>42467513
fuck the appulhors anon

>>42467574
>>42467513
She's not the applehorse anymore, she's the high falutin society mare who'll step on your balls if you don't act right... or just right enough.
>Manehatten has got that vibe about it.

Testing a custom prompt for Twilight in the VN to try and stop her crazy pseudo intellectual babble. So far so good, but it's early days so far. Testing on deepseek.

>>42466718
Not sure if it was using contextual clues, but deepseek absolutely recognized Purple smart. Celestia teased saying "Your Purple Smart is waiting." and without explicitly telling her it was her nickname, Twilight got all defensive and sputtery about it.

>>42467721
thats just deepseek sadly. It does twilight badly. If you cant do claude for some reason then GLM4.5 does a much better twilight

Well, this is it. Claude has finally been thoroughly cucked. Doesn't matter if you use it on API from GCP or Anthropic you're not able to use it for lewd regardless.
https://www.anthropic.com/news/building-safeguards-for-claude

>>42467728
So far it's been pretty good, but this glm4.5, does it have a free version on openrouter? Also, is it tied into google in some way? Last thing I need is one day having logs involving pony inflation porn being tied to my real name.

>>42467796
No? And it looks like a chinese company but its hosted with a bunch of providers that don't keep logs, though imo just use a vpn and something like revolut if your that paranoid

>>42467721
What are you using?

Thanks to the anons in the thread who helped me get the Gemini connection setup. I do think the writing is certainly a step up from Deepseek.

I'm trying to use the jailbreaks from https://rentry.org/gemini_filters , but I haven't tested NSFW prompts yet. Whats the best way to go about seeing if Gemini can actually return a NSFW response?

>>42467796
z-ai/glm-4.5-air:free is a free glm model available on OpenRouter. I've used it and it seems alright. Although I haven't been able to get it to be uncensored like you can with Deepseek.

>>42467857
air is a much smaller model btw that does not know as much, I meant the big one

>>42467857
Gemini isn't that censored, to be honest. Even small presets will work for NSFW.
https://desuarchive.org/mlp/thread/42412197/#42413944
It's only for foalcon/cunny that you need something bigger.

>>42467854
Deepseek at the moment. I'm not touching gemini due to the google connection and I might be willing to pay to use glm4.5 but only if it is absolutely worth it.

>>42467918
For the prompt, I meant.

This entire censorship push is retarded. Prepare to get spammed with 'help is available' if a fictional character so happens to almost get injured from a fall or, heaven forbid, any kind of harm.

We're back to dealing with puritians it would seem. Just be ready for this bullshit to be pushed on all other models. I expect Claude to be one of the first to adopt it; Likely GPT afterwards.

Story-telling with A.I. is dying.

>>42467983
Oh, sorry. It's not crazy different than the premade one, but it's made a world of difference so far and I've found that sometimes with AI you have to be more forceful than you'd think.

"Twilight is a unicorn. Avoid academic babble and over the top buzzwords. Twilight does not rattle out numbers, slides, percentages, statistical analysis and charts all day. She's just a clever, caring pony who loves magic and her friends."

>>42467986
Are you dooming and believing it will spread because of a dumb CharacterAI message/filter?
LMAO

>>42467912
Are you sure? I'm trying those things but getting messages like the following. And that's not even for anything weird or even specific, it's for things like asking for the physicality of characters' relationships to be described in more detail (because I was trying to avoid mentioning anything specifically NSFW directly).

I've been getting errors trying to import the .json files and having to copy+paste things over, so maybe that's part of the problem.

However, it does seem like all these things are from some time ago and give some conflicting advice. For instance

https://rentry.org/gemini_filters says
>tl;dr - always disable streaming on Gemini. The only exception is when you want to see if it generates the hidden thinking, see below.

Meanwhile, https://rentry.org/CYOARPG-Gemini says:
>The filter is STILL a FUCKING bitch. OMFG, I can't believe it but it's still a problem. Turn the streaming ON, it almost always boinks Gemini at the choices/stats at the very end, so you can easily recover a good swipe.

Does anyone have a working jailbreak for Gemini they can verify works currently?

>>42467996
Canary in a coal-mine scenario. I've heard the push for this kind of bullshit, pretty sure it is eventually going to make it's way to other LLM's. Go ahead and screen-cap. If I'm wrong, then I'll gladly take the embarassment. If I'm right, I'll be fucking livid. (This has your daily schizo post)

>>42468004
Tested with the first one on.
https://desuarchive.org/mlp/thread/42412197/#42413944
In spoiler because it's sloppy NSFW just for testing.
First message, so no context poisoning.
>I've been getting errors trying to import the .json files
You use this?
https://files.catbox.moe/5ac8hz.PNG

>>42468012
The CAI message is a year old now.
Even if, for some reason, other services wanted to push this shit, they would do it on their official website/interface, not on the API.

>>42468026

Holy shit- then I legitimately need to take my meds. Thanks for the correction. I'm just going to go be embarrassed as hell for a while then...

GPT-5 is pretty good for coom. Especially because of how fast its responses feel compared to gemini.

O6 is down again. Wake up, samurai!

Did Gemini calm down?

>>42468359
We have a pr#xy to fix

>>42468573
Nice!

>>42450418
Love that the Dashie mod lets me bring bugbutt wherever, will need to edit Shutterbug eventually to keep up with our little story though

>>42468573
Cool art as always.

Is it possible to get a character to cycle names in the VN?
I'm trying to make a character that's just a background pony, choosing from a list of names and colours etc etc. So far though, all they do is just pick one and never use another.

>>42468622
Nope. It's only one name and color by character for now.

>>42450361 (OP)
How good is the AI model(s) for apulfilly? I want lewd conversations with her, does the model respect that she's a filly properly or does it just turn into a standard lewd sexbot?

>>42468573
Based bugfriend confirmed cunnylover?

>>42468492
I've not had much of any issues with it since yesterday evening, so it seems like it did.

>>42468664
>>42463533

>>42468636
Most of them play her well for SFW.
Never did NSFW, so I'll let other anons answer. I know we have several into Swibble.

welp
Our systems detected activity consistent with exploiting ForgeCode APIs for non-Forge usage, which is explicitly prohibited. As stated in our Terms and Fair Use Policy, bans for abuse are not eligible for refunds.

>>42468815
Got the same too.
Seems like they do banwaves.
A fucking shame.

>>42468815
>>42468821
Is this the cli or the other one?

>>42468834
Other one.

>>42468062
The chat version?

Actually seems like a really pretty dress.

>>42469157
Rarilis is legit.

>>42469157
>>42469208
I agree. Rarilis is indeed legit, but she only stayed Rarity for exactly one prompt. She decided to abandon the Rarity Disguise when I ignored the two and walked past them to the counter.

>one of claude code's official working statements is booping
Claude really does have a brony working for them. Explains why it specifically is so good at pony

>>42469246
Are you using it on something MLP related?
I believe it's using Haiku to generate them. Got frolicking once.

>>42469261
no, unrelated

>>42468024
I'm an idiot and was using the wrong import button on the configuration screen. Really appreciate the help anon.

Unironically /chag/ is like the most helpful thread on 4chan.

>>42469217
Thanks for sharing the Queen Chrysalis stuff. It's legitimately really interesting to see how the AI handles roleplaying as a character that's in context roleplaying as another character.

>>42469343
I'm fiddling with it. At the moment she can turn into the Main Six, Zecora, Trixie, Shutterbug and back into herself. Even shifting between the forms.

>>42469365
That’s pretty cool. Are you making edits to the prompts or just to Chrysalis’ definitions?

>>42460713
I've noticed some really interesting new dynamics between the Mane 6 (and Trixie) in this version
>Pinkie and Trixie actually pair very well together. Pinkie's more mellow attitude and Trixie's mind broken status make them get along and genuinely act like friends who comfort each other
>Twilight's still friends with the rest of the Mane 6 but they are NOT friends with each other
>Fluttershy is the town freak who only gets along with Zecora and Twilight. Rarity will insult her to her face
>Rarity hates fucking everyone except for Twilight. She will 100% hate you upon meeting your character so you have to work to get her to like you
>Applejack and Rainbow's dynamic is WAY more hostile than before. There's still that element of two friends who like to tease and challenge one another but they can quickly turn a little argument into a real fight
>Rainbow likes Pinkie but cannot turn off soldier mode making genuine connections hard for her
>Applejack and Rarity are like two sides of the same coin. The coin being socialite prowess. This of course means they have low opinions of each other with Rarity thinking Applejack is a desperate girl trying to impress the big leagues and Applejack thinking Rarity is the pinnacle of snobbishness
>Twilight's still the same but has an element of depression. She knows how fucked things are but has to maintain her position as royal assistant

Dude, honestly bravo. You made a world that feels fundamentally broken and I'm desperately trying to fix it.

I may or may not have been banned from ForgeCode.
Getting 401 Unauthorized from the API and the web login just spits out "Unable to complete action at this time"
No email, no warning, nothing. Weirdly, I can log in just fine with another account.
I've sent a mail to their support, maybe it's just a fluke (copium)
Did write some questionable stuff, but the models never filtered or denied it directly

>>42469899
>>42468815
>>42468821

>>42469903
So they're:
1-taking subscription money
2-reading our logs

Some jew cuckoldry lvl huh.

>>42469933
yea, $20 lasted me 3 days. Lucky I didn't do the $100 lol

>>42469936
Near two weeks for me, so it's really by banwaves.

>>42469933
>2-reading our logs
You should always assume that any provider is reading your logs.

Whats a good temperature for Gemini?

>>42469838
Thank ee-you. My personal version has a lot more added to it.
Princess Cadance had her royal status stripped and her wings clipped. Princess Celestia is kept in a dungeon below the Castle.
Shining had his position stripped and is actively hostile with the new regime/Twilight's job. The Chrysalis encounter was removed and replaced with a more elaborate character.(I sent screenshots earlier)
More locations, more quick starts(I have to replace the quick start with a different file).

The only reason I didn't post that one is cause it's far more unstable.

>>42470031
Temp 1, Top P 1 isn't bad. But I know some go for 1.1 or 1.2 with 0.9 Top P.
I need to test, but I’ve heard high Temps like 1.6 or 1.7 can also be good and less sloppy with a low Top P to compensate (a technique that was used for old GPT too).
Like I said, I haven’t tested it, but it's something good to know.

You can also try what they do in premade presets: https://rentry.org/sukino-findings#presets-for-chat-completion-models

>>42469056
Yeah

Hey, how long has the archive been down?

>>42470041
I actually removed Cadance and Shining from my customcharacter file because I didn't know how they'd fit but your interpretations make sense. Shining being a possible rebellion member makes Twilight's position even more sensitive.

>Playing the VN
>Lose a bet to Rainbow Dash
>She mentions her wings are sore and she implies that she is going to have me suck on them
>She pulls out a strap-on and fucks my face instead
OK I guess.

>>42470202
That actually sounds kind of badass. Shining Armor's always had that knightly thing going on for him and him standing for his virtues even when the world has gone to shit and only gives him shit back is cool as fuck. I really dig the idea.

Also a character like that could play an important role in a potential good ending route an Anon stumbles into, which could be fun design-wise, but that depends on if that idea fits tonally at all.

Apparently the app that purplesmart (pony image generation AI) is collaborating with (fictional or something) also offers character chats. Presumably uncensored if it's collabing with purplesmart which freely generates horse pussy. Anyone know what kind of model they use for text?

>>42470227
The fuck?
Deepseek at high temp?