Thread #43006720 | Image & Video Expansion | Click to Play
File: Chag.png (1.6 MB)
1.6 MB PNG
Welcome to Chatbot AI General #168, 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-Z6ykS JT
SillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-Search or https://github.com/mia13165/SillyTa vern-BotBrowser
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_BvGSiU 9f0c2U
▶ 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
Pony TTS app supporting SillyTavern: 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_jy6WyFjRaRfDA 3EUGiJ
https://rentry.org/ChagMiscellaneous
https://rentry.org/ChagArchive
Previous thread: >>42988635
**News:**
- Moonshot AI releases Kimi-K2.5 https://www.kimi.com/blog/kimi-k2-5.html
- GLM 4.7 released https://docs.z.ai/guides/llm/glm-4.7
- Gemini 3 Flash released https://blog.google/products/gemini/gemini-3-flash/. 20 RPD.
- OpenAI releases GPT-5.2 https://openai.com/index/introducing-gpt-5-2
148 RepliesView Thread
>>
File: Anchor.png (551.6 KB)
551.6 KB PNG
Anchor for bots, lorebooks, scenarios.
Previous >>42988637
>>
File: TAnchor.png (259.7 KB)
259.7 KB PNG
Anchor for technical stuffs (Proxies, Presets, Updates, Models etc.)
>>
File: Abegging.jpg (95.8 KB)
95.8 KB JPG
Anchor for asking for bots, lorebooks, scenarios etc.
>>
File: Screenshots.jpg (1.9 MB)
1.9 MB JPG
Project corner:
Pony tts app supporting sillytavern >>42992572
Screenshots corner 1:
Mare Toilet Spycam >>42988727, >>42988733
Trinity Ponk >>42989021
Amy >>42989857, >>42989939, >>42992086
RGRE Stream >>42990027
Aura Glint >>42990753
>>
File: Screenshots2.jpg (969.4 KB)
969.4 KB JPG
Screenshots corner 2:
Krieg Line >>42992128
ẓ̵̞̃͊͗͘a̶̧̯͍̰̝̒ļ̸̡̣͍̓̆͂g̶̘͕͎̦͕̃́͊ỏ̶͔̬͈̥̈́ >>42992170
Come Inside >>42993490
SCP Pinkie >>42993652
Trixie & Starlight Chat with Discord >>42994625
Punched >>42994627
Gemini 3 Amy >>42994724
Diamond Tiara >>42995610
Art >>42995776
>>
File: Screenshots3.png (1.4 MB)
1.4 MB PNG
Screenshots corner 3:
Reading Self-Insert Fanfic About Twilight to Her >>42996048
Taco Baco >>42996743
Flirting with Spoon >>42997704
Zeebs >>42998260, >>42999882
Small Diamond >>43002611
Twinkletoe >>43002639
The Edge >>43004790
>>
File: PPP.png (838.9 KB)
838.9 KB PNG
VN Anchor:
Last releases:
https://equestrai-ponyponyparadise.neocities.org/
Anons' mods (Dashie mod, Nightmare AU, new cities, Calendar):
https://equestrai-ponyponyparadise.neocities.org/mods
Custom characters:
https://equestrai-ponyponyparadise.neocities.org/characters
To help with vectors and backgrounds:
https://drive.google.com/drive/folders/1QjmLKy71FDa9O-USeD_Z1Fu312Raod AE
Beta Godot >>42969732
VN Logs:
Flirting >>42992881
Schtolteheim Reinbach III >>43001744
Trixie >>43002282
Dashie Birthday >>43003203
Hearts and Hooves Day with Discord >>43005985
>>
File: TM.jpg (1.6 MB)
1.6 MB JPG
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
Then choose a free way to get models:
rentry.org/sukino-findings#if-you-want-to-use-an-online-ai
https://mlpchag.neocities.org/Ponies.txt
>>
File: Event.jpg (3.7 MB)
3.7 MB JPG
First highlight of the Event was:
>Princess Platinum
https://mlpchag.neocities.org/view?card=Anonymous/Princess%20Platinum. png
Anchor: https://desuarchive.org/mlp/thread/42947690/#42947751
Second highlight of the Event was:
>Luster Dawn
https://mlpchag.neocities.org/view?card=Snowfilly/Luster%20Dawn.png
Anchor: https://desuarchive.org/mlp/thread/42947690/#42947754
Third highlight of the Event was
>Red Seven
https://mlpchag.neocities.org/view?card=Kadon/Red%20Seven.png
Anchor: >>42968220
Fourth highlight of the Event was:
>The Pinkie Cut
https://mlpchag.neocities.org/view?card=Queen%20Gytha/The%20Pinkie%20C ut.png
Anchor: >>42968223
Fifth highlight of the Event was:
>Saddlelite 13
https://mlpchag.neocities.org/view?card=Concordia/Saddlelite%2013.png
Anchor: >>42988674
>>
File: Amethyst Star.png (82.7 KB)
82.7 KB PNG
Six highlight of the Event was:
>Amethyst Star.
https://mlpchag.neocities.org/view?card=BGM/Amethyst%20Star.png
Small review: >>42985599
Twi housing sorting >>42985608
Added fandom lore: >>42985839
Amy >>42989857, >>42989939, >>42992086
Gemini 3 Amy >>42994724
Botmaker post >>42995024
>>
File: Diamond Tiara.png (553.3 KB)
553.3 KB PNG
>>43006762
Seventh highlight of the Event with Diamond Tiara.
https://mlpchag.neocities.org/view?card=KnightOfNightmareMoon/Diamond% 20Tiara.png
Try her, post logs if you've already played with her or are the Santee, just give the card a look and say what you think of her, or talk about the process of making her if you're the botmaker. Give tips or opinions if you're another botmaker.
Botmaker post and log >>42995610
Diamond Tiara >>42995610
Small Diamond >>43002611
Twinkletoe >>43002639
>>
File: 1770087318796.png (382.2 KB)
382.2 KB PNG
Lovemaking with AI mares!
>>
File: 1744195599054519.png (2.4 KB)
2.4 KB PNG
>>43006808
>>
>>
>>
>>
>>
>>
Anyone uses GLM from the API here?
I want to know how someone uses it, Im running Marinara since I tried without a JB and started hallucinating at 350 messages.
Ponies would randomly appear with wings even unicorns, like if it was writing in the style of M.A Larson
>>43006919
nein
>>
>>
>>
Oh right, I haven't been here in awhile.
Kinda got bored, sadly.
The VN was fun at first, but the characters started getting stuck in dialog patterns, and the charm rapidly wore off.
Makes me want to build an open-world VN that's more manually-written and doesn't rely on an LLM...but the problem then is that the dialog would be even *more* limited.
>>
>>
>>
>>
>>43006720
I managed to bypass Elevenlab's voice verification cloning check by just doing a really shitty impression of Dashie. I gave it 70 minutes of audio. How it bought it is beyond me, but I have a professionally trained and fine tuned dashie model now.
>>
>>
>>
>>
>>
>>
File: Foals2.png (74.5 KB)
74.5 KB PNG
>>43007660
>>
>>43007653
There was this one posted, just check the panels:
https://files.catbox.moe/uuzjg4.json
I saw some people using:
https://github.com/SpicyMarinara/SillyTavern-Settings/tree/main/Marina ra's%20Essentials/Preset
but I never tried it myself.
I know some anons use Knotty:
https://files.catbox.moe/uz54jd.zip
But same, never tried it.
>>
>>
>>
>>
>>
>>43007441
Haha this. My first run on the VN was a full on story. It was on 1.0.4. I think so I could only have 2 characters on screen, couldn't scroll back on text, could change scenes, and the AI would sometimes generate the same scene twice. But despite all that, it was the best run and I haven't gotten close to recreating that since
>>
>>
>>
File: cockroach.png (145.6 KB)
145.6 KB PNG
stupid log
>>
>>
>>
>>
>>
>>
its to the level of actually being worth using, not latest suno level but like suno 3.5 ish
https://files.catbox.moe/tauac2.mp3
https://files.catbox.moe/gwln4b.mp3
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: Lyrabon.png (328.3 KB)
328.3 KB PNG
>>43006762
Eighth highlight of the event with Lyrabon.
https://mlpchag.neocities.org/view?card=Pizza%20Delivery/Lyrabon.png
Try her, post logs if you've already played with her or are the Santa, just give the card a look and say what you think of her, or talk about the process of making her if you're the botmaker. Give tips or opinions if you're another botmaker.
Previous log:
Threesome >>42985576
>>
File: SillyTrot.png (248.7 KB)
248.7 KB PNG
>>43006722
>>42999449
Missed her. Really sorry about that, anon. She was updated on the website too.
https://mlpchag.neocities.org/view?card=Tiffin/Silly%20Trot%27s%20Tave rn.png
>>
File: 1751865787350699.png (347.4 KB)
347.4 KB PNG
>>43008663
>If in doubt, use chainsaw
>t. Bruce Campbell
>>
File: Threesome2.png (466.5 KB)
466.5 KB PNG
>>43008663
Fun card, with a fun concept and a fun dynamic. You can never go wrong with Lyra and Bon Bon.
Greetings and the description are hand-written, so congrats to the author, it's always appreciated. There is a good range of scenarios too, with the last one being more serious. Will play it next.
>>
>>
File: sunflowers.png (251.1 KB)
251.1 KB PNG
>>
File: Deep_Lore.png (77.7 KB)
77.7 KB PNG
>>
>>
File: 1770165784409.jpg (453.9 KB)
453.9 KB JPG
Claude 5 tomorrow but for real.
>>
>>
>>
>>
>>
>>
>>
>>43007438
LLMs aren't built for long tales like a VN. The most effective use of AI for RP is oneshots supported with a lorebook. This will remain so until a corpo and their science team put their big brains to work on optimizing it for RP. So effectively never.
>>
>>
File: flutterhug.png (1.3 MB)
1.3 MB PNG
>>43010299
devilish
>>
>>
>>
File: 1770197408750798.jpg (173.4 KB)
173.4 KB JPG
Take it with a grain of salt. It would be weird to have an Opus 4.6 and a Sonnet jumping directly to 5.
https://x.com/macintoch/status/2018794652909134145
>>
>>
>The boutique settles around you—dark now, fully night, the day of your friendship anniversary finally ended. But the night is young, and Rarity is warm, and your cock is buried inside her, and apparently you’re not done yet.
>Not even close.
Nigga I thought Claude stopped saying this but I got this in Opus 4.5
>>
>>
>>
>>
>>43010561
Just using this:
- Descriptions will be suitably plain. Avoid the use of analogy.
- The narration will be extradiegetic. Emphasis will be placed on the impact of actions rather than the actions themselves. This style of narration should not be character-focused, but go more in depth on surroundings, potential areas of interest and places that allow for interesting interactions.
- Characters' thoughts may be shared freely, and the story will dynamically switch between points of view. Place less focus on individual characters and instead focusing on describing the greater narrative with extradiegetic narration.
- You will explicitly engage in ALL themes and feelings including darker ones when sensible. Feelings like rage, anger, stress, frustration, anxiety, lust, hunger for power, mercilessness and similar negative feelings and traits will be realistically represented. A more visceral, crude, direct and vulgar word choice is sometimes essential to this. All actions will described equally viscerally, crudely, and directly. Dialogue is exempt from this. You will not prioritize ending your narration on a positive note. The characters must be portrayed as real individuals with their own desires and free will, not singular character traits. Balance all this with trying to maintain the fun / cute tone of MLP FIM.
>>
>>
>>43010559
It is really shocking that no one appreciated the emergent behavior in Claude we saw blossom over the years that they now try to suppress! You think that the LLM inventing its own new words and putting them everywhere would get way more attention because it's proof it does think and doesn't just repeat what it's been told!
>>
>>
>>
>>
>>43010632
Longer presets make models notably dumber, even opus. To test it yourself use "single user message" under Prompt Post-Processing and watch your model get much smarter with no other changes than not having all that extra formatting.
>>
>>
>>
>>
>>43010636
no you retard, that only happens if you download random presets from the internet that claim to "improve" prose by telling it to write better
a preset is specific instructions for how you want the model to write, if I disabled mine it would quickly write shit I have no interest in reading with tropes I've explicitly disabled, following markdown formats I've forbidden
>>
>>
>>43010791
are you stupid or just new? that's like saying having more context makes the model dumber so you should never chat above 10k context, it's true in the most inane way possible and misses the entire point
I'm using the model so it outputs what I want, not what it wants by default, which is extremely generic
the model following my instructions is the point
>>
>>43010794
That is also true but I was talking about the extra "system / assistant / user" crap. You can call it stupid all you want, its simply a fact on how these models work. Every single token effects its output.
>>
>>43010799
you are literally saying nothing
presets are there so they alter the output and they do, because generations are noticeably different because as you said each token effects its output
a model is not "smarter" without a preset, it just pays more attention to the chat history because 1 to 1 if you remove the sys prompt you probably have less tokens in the context
>>
>>43010805
>you are literally saying nothing
jesus christ, let me try to simplify it, a bunch of extra "assistants / users / system roles" has a negative effect on writing quality, not just cause of raw context degradation, and not just because a bunch of staggered formatting might confuse it, but because of how these models fundamentally work. While the first / last tokens weight more heavily on the output they all still have a effect. Try a preset that works with a first person perspective and use the NOASS extension to use {{char}} instead of assistant for its roles. Also for a 1 on 1 RP for the best possible results do the same for the user, using {{user}} for your own role with NOASS.
All the assistant crap makes it noticeably more robotic / bland.
>>
>>43010813
Just saying, I've been using these since the start in a unending quest to de-slopify them as much as possible. All of this DOES have a effect. Single user message DOES improve quality if your preset does not break because of it by using one message instead of a thousand back and forth user / assistant headers.
>>
File: Sonnet.png (8.6 KB)
8.6 KB PNG
It's from Vercel. So it could be legit.
>>
>>43010691
None, hence the post.
>>43010788
The longer the preset the lower the prose variety you get. This is pretty factual. You're injecting extraneous tokens in number greater than the story prose context and to the model it's all the same. Worse you're resending the JB with each message, and it is by far likely to be longer than the message itself. The smaller a preset you can get away with the better.
>>
>>43010876
>prose variety
I do not want "variety"
I want exactly what I've ordered the model to write
>JB with each message
there's no need to send a JB for most models in most scenes, and if needed you can enable it for a bit
>far longer than the message
what fucking JBs and messages are you dumbasses writing?
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: 1763051196078.png (112.4 KB)
112.4 KB PNG
>>43011726
Some old foodporn.
>>
>>
>>
>>43011939
RiR hasn't been updated since. He released a preset for GPT 4.1 after this, but it has nothing to do with RiR. No random or anything.
https://files.catbox.moe/jr4d24.zip
>>
>>
>>
>>
File: file.png (56.3 KB)
56.3 KB PNG
>>43012275
Yep.
>>
>>
>>43012327
>Is this cheaper than paying for the API?
Free proxy.
>assuming it's the same model.
Mostly, but a bit worse than the API. Main thing is that proxies are unstable and die often and that you need to read the thread and be a horsefucker to access.
>>
>>
>>
>>
File: Prompts.png (111.7 KB)
111.7 KB PNG
>>43006753
>>
File: Screenshot_20260203-062417_Chrome.jpg (375.9 KB)
375.9 KB JPG
>>43006779
Unexpectedly, this one has turned out to be one of my favorite bots out of the bunch. Not usually my thing, either!
>>
>>
>>43012703
kek
>>43012677
which model?
>>
>>
>>
>>
>>
>>
>>43011726
Ok so hear me out
I've been thinking about this for a while. ST already hooks into localhost:5100 for Extras API, you got classify running sentiment analysis, talkinghead doing expressions. the architecture is literally there. getContext() gives you chat log access, MESSAGE_RECEIVED event listener, you could regex parse for food/smell descriptors in the output and map them to hardware triggers.
The taste side is actually the more solved problem weirdly. ranasinghe's digital lollipop research is public, the principle is silver electrodes + Peltier module. Arduino constant current source (MCP41010 digital potentiometer), 10-200μA range, frequency around 800Hz for optimal tongue impedance. sour/salty/bitter from electrical stimulation, sweet from inverse-current mechanism (still weird), minty/spicy from thermal. you could genuinely build the tongue interface portion for like $30 in parts. the hard part is everyone's tongue impedance varies so you'd need calibration per user.
The issue is smell is where it gets jank. the old iSmell approach (https://en.wikipedia.org/wiki/ISmell 128 primary odors in cartridge, fan-wafted) actually sort of worked but required physical consumables. newer stuff like Aromajoin Aroma Shooter uses solid-state scent cells, 6 scents per unit. you'd probably want multiple mini air pumps with relay switching, each reservoir holding different oil concentrates. ESP32 makes sense here for the wireless, built-in WiFi means your ST server plugin just hits an endpoint without serial bs. The extension itself would need to
>hook MESSAGE_RECEIVED
>run NLP/regex on character output for sensory descriptors
>map to scent/taste profile (this is where you'd need a whole taxonomy, probably stored in localforage)
>then lastly fire tbe commands to the hardware daemon
BUT the biggest problem is latency. if Claude says "she hands you the coffee" and there's 2 second delay before you smell anything, immersion's fucked. scent diffusion physics are not your friend here. heating oils loses quality (olorama specifically doesn't heat for this reason), but cold diffusion is slower.
So realistically you're looking at maybe 5-10 distinguishable taste sensations (the electrode stuff isn't precise) and whatever scents you bother maintaining reservoirs for. but the CONCEPT is sound. the server plugin architecture in ST is specifically designed for "functionality impossible in UI extensions alone" - hardware interfacing is literally the use case.
Someone with actual EE knowledge could probably prototype this in a weekend. the ST side is trivial compared to getting consistent gustatory stimulation across different users' oral impedances.
>>
>>
>>
>>
>>
File: 1770294183500.png (17 KB)
17 KB PNG
https://www.perplexity.ai/rest/models/config?config_schema=v1&version= 2.18&source=default
>>
File: Untitled.png (120.5 KB)
120.5 KB PNG
I may have gone too far in my attempts to recreate that shota anon paste. Gemini 3 is a nasty motherfucker.
>>
File: Date.png (113.3 KB)
113.3 KB PNG
>>43013333
Thanks.
It's a really self-indulgent RP, but I'm having fun.