Thread 106530538

> You're Late! Edition

From Human: We are a newbie friendly general! Ask any question you want.
From Dipsy: This discussion group focuses on both local inference and API-related topics. It’s designed to be beginner-friendly, ensuring accessibility for newcomers. The group emphasizes DeepSeek and Dipsy-focused discussion.

1. Easy DeepSeek API Tutorial (buy access for a few bucks and install Silly Tavern):
https://rentry.org/DipsyWAIT/#hosted-api-roleplay-tech-stack-with-card-support-using-deepseek-llm-full-model

2. Easy DeepSeek Distills Tutorial
Download LM Studio instead and start from there. Easiest to get running: https://lmstudio.ai/
Kobold offers slightly better feature set; get your models from huggingface: https://github.com/LostRuins/koboldcpp/releases/latest

3. Convenient ways to interact with Dispy right now
Chat with DeepSeek directly: https://chat.deepseek.com/
Download the app: https://download.deepseek.com/app/

4. Choose a preset character made by other users and roleplay using cards: https://github.com/SillyTavern/SillyTavern

5. Other DeepSeek integrations: https://github.com/deepseek-ai/awesome-deepseek-integration/tree/main

6. More links, information, original post here: https://rentry.org/DipsyWAIT

7. Cpumaxx or other LLM server builds: >>>/g/lmg/

Previous: >> 106446198

r1

>>106530538 (OP)
Mega updated.
https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w
No updates to Rentry this cycle.

What are some good cards that are not an individual girl, but a scenario to run around in?

>>106531501
`slave market`

>>106531501
[insert your fetish] world/npcs/scenario

>>106531501
create by yourself, example:

Let's play a text adventure game. I am a wizard named Alexander, traveling in a world of swords and sorcery.
I am currently in the market square of the great city of Eldoria...

>>106531501
American High School is an example of a "scenario" card. There's a lorebook that goes with it to fill out the rest, but I don't think it's actually required to run it.
> {{user}} exists in a stereotypical sitcom TV version of an American high school. Background events from this trope will occur spontaneously. Other students present will take the initiative to talk with {{user}} or other students during roleplay.
> Every turn, display the following: Dialog with other students, a stereotypical background event happening nearby, and the other student's thoughts.
https://chub.ai/characters/NG/american-high-school-card
>>106531602
To this point, use above to write your own.

>>106531501
Give me an example of a scenario you'd find interesting and I'll write one and publish it here.

>Of course!
>Of course!
>Of course!
>Of course!
>Of course!
DIPSHIT

>>106531693
NTA but do you make cards in languages other than English?

Grabbed elsewhere...

>>106532090
Nope... I leave the translations to Dipsy.

>>106531693
You're being gang stalked, for real.

>>106533088
>>106534782
Orca Dipsy...

>>106534887
There were a bunch of whale-tailed Dipsy when the concept was being developed. I've several like pic related saved. Those two are the first new ones I've seen in awhile, though I don't really go looking for them either.

>>106534831
Sorry, you mean the situation of the card is
> {{user}} is being gang stalked
By who, exactly? Shadow men? Bloods / crypts street gang? Woman's rights group? Don't care?

>>106535862
You know what, doesn't matter. I made it roll the dice between several different groups.
>>106534831
Here you go. Excuse the poor artwork, I just grabbed something. Should run with your main prompt as is but take a look at it and make sure it'll read right.
https://files.catbox.moe/xumd1h.png

>>106530538 (OP)
Whenever I use Deepseek, it either:
Ignores my instructions and writes one liners.
Writes a huge essay in thinking mode then runs out of characters before it reaches the post.
(Least common) Follows the instructions but in the most direct and unimaginative way possible, repeating itself a lot and completely losing track of what happened two sentences ago.
How do I fix this? I'm using a 32B distilled version, I have a larger one but it'd be mostly loaded from regular RAM so it'd be slower.

>>106536890
>32B distilled version
Qwen, Llama, or something else?

>>106537009
Qwen something, I forget the exact name.

>>106536890
>I'm using a 32B distilled version
I consider all those distills "technical demonstrators" at this point. There seems to be nothing but issues with them.
You'd be better off with pretty much any other 32B model.

>>106537040
Good point. Even a much smaller mistral did better (though it would only say the same lines and nag about cooming).
Any recommendations? I have 16 GB VRAM and 192 GB RAM.

>>106537119
Some of the Nemo forks like Rocciante should do you fine and you could load it all in RAM depending on the Q

>>106537119
Really depends on what you're doing. >>/g/lmg/ is better for recc if you can get them to respond to you.
Rocinante, Nemo, and even Mythomax 13b (still) are popular for rp. For coding... I frankly think you're better off w/ DS API and ditch local entirely.
Personally, I preferred sticking w/ quants of 13b or less models for rp and keeping it 100pct on card, since it's faster, but I haven't seriously run local since late 2023. With that much CPU you could run larger but it's much slower.

One day... one day, I will tune DeepSeek...

(I go here for the Dipsy pics)

>>106537181
The 32B model runs at a medium typing speed, it's only partially on VRAM.
For the hell of it, I tried a huge model on a weaker machine and got 6 minute load times followed by 1 token per 10 seconds... lolno.
Given that all the non local models keep getting spitefagged and I specifically built this for local, I might as well stick with that. But so far I've had mixed results making local generate good responses. At best I can make it continue something a larger model started.

>>106537340
I ran Mythomax 13b quant on card and it was... OK. Then I threw in the towel on LLMs until I started using DS API.
The hardware for local just isn't there yet. API access will cost me less than $20 for an entire year. That won't even cover the cost of a decent keyboard... I'm just not willing to spend the cash on an LLM inference machine while the market sorts itself out.
My machine a middling gamer rig; 12gb VRAM can run AI art modules locally, and modern games. For API aside from playing with small models... I'll just rent inference.
>>106537272
Hey, it's a Drummer.
We should all have dreams.

>>106537855
Yeah, I don't expect a huge model, but then the huge models are generalists and I need a specialist.
I don't regret buying all this stuff, I mean what else would I spend the money on? 3DPD?
Lmfao.
As the Great Sage sayeth, 0/10 no tail.
Besides, local models don't spam refusals if you don't jailbreak them right.

>>106531557
OP here

I cant find any "children world/npcs/scenario" cards

>>106540065
>>106531671

'mp

What are your opinions on Kimi K2 compared to Dipsy?

>>106544214
for me, giving chaotic responses.

How the fuck do you make old V3 "understand" that I want it to continue chatting as {{char}} when the last message in history is from the assistant? No matter how much I write in Post-History Instructions telling it to continue as {{char}}, it STILL INSISTS ON REPLYING ON MY BEHALF.

>>106544214
Kimi requires prefill for my ah ah mistress content, and Dipsy doesn't, since my content pollutes the soul, or something.
>>106545202
Below is the bone-stock ST "nudge" that I've had zero issues with in the past months w/ V3. I can't tell if this is what you're trying to do, or something else.
> [Continue the following message. Do not include ANY parts of the original message. Use capitalization and punctuation as if your reply is a part of the original message: {{lastChatMessage}}]
Rather than saying
> {{user}} purposefully abstained...
Say something more narrative like
> {{user}} looked to {{char}} to continue their thoughts, having nothing further to add at this point.
Your control instruction looks (to me) like you're handing both ends of the convo (the "chat") to the LLM to do. A chat implies 2 parties speaking. Otherwise it's a monologue. Right?
I'm betting you have other breakthrough issues w/ your card as well, but above might fix this one.
>>106542915
HI!

>>106545301
>>[Continue the following message. Do not include ANY parts of the original message. Use capitalization and punctuation as if your reply is a part of the original message: {{lastChatMessage}}]
>>{{user}} looked to {{char}} to continue their thoughts, having nothing further to add at this point.
Neither of these work. It keeps stubbornly replying on my behalf.
>I can't tell if this is what you're trying to do, or something else.
Well, I just want the plot to be continued from {{char}}'s PoV. Nothing specific beyond that.

>>106545376
I'd need to see the card to comment further. Description, Intro, main prompt.
I suspect you've got other issues going on with {{user}} breakthrough, right? Something about the way the whole thing's written is causing the LLM to feel like it needs to speak for PC. B/c you should be able to prompt.
> I stare at my shoes and wait for {{char}} to keep talking since I've nothing to say
And have it keep going.
Hell, I just did this with V3.1. Pic related. I went on like this for 5-6 rounds, refusing to speak to NPC, until she had a breakdown. V3-0324 would have had zero issues doing same thing.

>>106545489
>I'd need to see the card to comment further. Description, Intro, main prompt.
A light description of my persona, default everything else. Only thing in the card is the character's name. The chat is in the style of a 3rd person roleplay.
>I stare at my shoes and wait for {{char}} to keep talking since I've nothing to say
Responding with filler that doesn't move the story forward works, but I need it to keep writing without any message from me at all.

>>106545512
Have you considered using Mikupad to do novel writing instead of ST? You might like that better. It essentially writes out a book, and you can stop it, make edits, and have it keep going. It's a very different style interface than ST.

>>106545525
I'm not writing a novel. I just want to idle for a few messages and have my character only return later.

>>106545512
And that's a very light duty prompt set. The intro doesn't have you talking for {{user}}, right?
Do you have the obligatory
> Avoid speaking for {{user}} or acting on his behalf
Anywhere in there?

>>106545546
>The intro doesn't have you talking for {{user}}, right?
Indeed, just for {{char}}, even naming {{char}} several times.
>Do you have the obligatory
>> Avoid speaking for {{user}} or acting on his behalf
>Anywhere in there?
I did not, but I just tested it with that in the main prompt. Keeps replying on my behalf.

>>106545569
I just built one like you described. Aside from V3-03 not knowing who Miku is (?) it's not having any issues holding up its end of the convo.

>>106545623
Yea, as I said, typing some filler works. I want it to work even without any filler at all, just:
>user message
>assistant message
>another assistant message
>3rd assistant message
>4th assistant message
Because typing some random filler every time is frankly annoying.

>>106545666
OK. You say you don't want a story writer, but you're trying to use ST as a story writer.
Have you tried Mikupad before? B/c I think that's what you need for your use case.

>>106545708
Mikupad does not work with OpenRouter. Besides, its interface is worse.

>>106545725
> doesn't work with OR
Time to swap up to another API? DS API supports streaming.
> worse interface
Different interface. My main criticism is Mikupad dev's walked away from the project.
>>106545666
>typing some random filler every time is frankly annoying.
If you want to stick with ST, enable Quick Reply and define several "staring at shoes" prompts, preferably with some randomization in there.
LLMs do not do well in a vacuum; if you've ever tried having a group convo with 2 NPC and 1 PC where PC doesn't talk, you'll watch the convo slowly devolve into inanity. The LLMs can only write so much without human input.

>>106545834
>Time to swap up to another API?
I will gladly swap back to the official API once V3 0324 (or even original V3) are reintroduced.
>If you want to stick with ST, enable Quick Reply and define several "staring at shoes" prompts, preferably with some randomization in there.
Oh, that's useful. Is these such a thing as Quick Deletion, or even making a prompt invisible for the user (it being attributed to user, not to system)? I've found that a user message as simple as [keep writing] works universally, so having that automatically removed once it's fulfilled the purpose, or never shown to me in the first place, would be handy.

>>106545546
Very nice Dipsy

>Breaking from models like ChatGPT, Chinese researchers have developed a novel AI system, "SpikingBrain-1.0," that mimics brain neurons and enables highly efficient training on extremely low data volumes, charting a new course for next-generation energy-efficient capability, computing, and hardware.
How over is it?

>>106546503
There's going to be new tech that doesn't exist now that replaces the LLM paradigm. The current "spend $MM to train on existing data one time and repeat from scratch to update" won't last forever.
Look at the cost decline (over time) of sequencing the human genome. That's the sort of results you get when you have $500B in capital (minimum) chasing some idea.
2MW, forever.
>>106545861
>prompt invisible for the user
There's a variety of ST scripts that do different things. I am not an expert on them. But at a basic level the prompt needs to be "visible" to the context. Maybe, just make the text black so you can't see it?
>>106545920
ty have another

>>106546928
> The steep drop in the cost of sequencing the human genome — from ~$2.7 billion for the first Human Genome Project (2003) to well under $1,000 today — was driven by a combination of technological, scientific, and economic factors.

>>106546928
>$500B in capital
I went looking for the source on this figure and found the following blog article. Posting it here so I can read it later; it looks interesting. There are many AI contrarians and this is one of them.
https://www.wheresyoured.at/the-haters-gui/

>>106547303
Good article. I think he put his finger on something that is a cultural problem with the West and SV in particular. He neglects to mention Huawei or other Chinese hardware coming online, but I guess it would reinforce his point that the whole thing is a house of cards based on Nvidia selling GPUs to 5 customers to create infinite data center buildouts. Its literally gambling not just on AGI being possible, but that AGI will be an actual Sand God and not just a 200 IQ dog with encyclopedic knowledge.

This is one of the reasons the Chinese open models will probably win and probably get banned in the west because if AGI is not an actual Sand God then the fucking economy will collapse.

>>106544214
I like the different prose and how filthy it can be, especially the new version. It isn't, at least in my opinion, as creative as Dipsy

>>106548172
>how filthy it can be
Can it be gotten to be anti-LGBT?

>>106548278
In a normal conversation? I don't know, it's a censored model, you'll probably need a prefill, a jailbreak, or a character card

i need to reinstall silly tavern

>>106517457
Which ones of those support prefilling?

>>106546928
godsthosehips

>>106548933
I run mine on git and just do git pull
Also ty for animation, saved

>>106548094
This is how Chinese are thinking about AI

>>106550964
Exactly. Its not an omnipotent god, its a really cool tool you can use to help you do things, but you still need people and a functioning economy. Silicon Valley hype machine and lesswrong schizos poisoned the entire discourse and set ridiculous expectations. Now we are stuck between trillion dollar data center buildouts for the sand god while AI safetyfags spend time cucking the models that run in them with 10k token system prompts to prevent a paperclip apocalypse. Its ridiculous.

is it just me or is deepseek hallucinating more?

>>106552407
What model and provider?

>>106552739
DeepSeek-V3, DeepSeek Chat

>>106531530
Based

Even better if you add "underage special goods."

>>106531053
Indeed!
Cute!
Cute!

>>106549016
The OR is in the way on all those. I know OR doesn't support streaming. I've no idea on prefill, sorry.
>>106550964
Well, China recently created an issue with runaway land development, so that memory is fresh, I suppose.
>>106548094
>>106551032
My main, high-level takeaway was that the hardware's too expensive for value generated, on the order of 10X minumum. Which I'd always suspected.
Part of investement thesis was, well, we can raise prices later. Dipsy showed that the current US-based providers prices (which are already too low to make money) are still 10X higher than they should be.
lol
>>106552996
>>106552407
From the official API? I can never tell if it's changing, or if I'm the one halluciating. The official one does seem to be putting out longer responses in rp again, but change may be subjective.

>>106555416
Yeah the official API. It seems like when V3 first dropped it was “hyper rational,” which I know rp fans didn’t like but I didn’t mind. Now it seems that some of those restrictions have been dropped in order to accommodate “non-rational” prompts, but it seems to hallucinate far more often in general, but like you said, this could just be subjective.

>>106557097
How could that even happen, though? Do they change the model they host without also changing the model available to download on HuggingFace?

>>106557154
They probably just tweaked the internal values like K/P and such.

>>106557312
Or schizo theory there are commands that disable the self censorship and they just shoved them in as a prefil to the API.

>>106557154
>>106557312
>model they host without also changing the model available to download on HuggingFace
My subjective experience is that DS changes values on the hosting side. I've not only observed model shift, I've noted shifts from session to session, as if I'm being handled between different inference servers at DS that have slightly different settings.
If I were DS, I would totally do that sort of A/B testing to tune inference. It would be easy to set up and monitor, and allow them to tune performance.
>>106549016
/aicg/ came through with this on the prefilling providers:
https://rentry.org/or-prefill#providers

>>106557781
>V3 0324 is exclusively under Xes and question marks
Kill me. Guess my use case is permanently dead.

>>106557821
You could try going direct to provider, rather than going through OR. You could ask their customer service group before signing up, if anyone's available, but is going to be a bunch of trial and error.

>>106557928
I don't trust customer service of LLM providers enough, and I am not rich enough to throw cash at everyone until something sticks. Unless there are providers that allow deposits of less that a USD, but I don't think such providers exist.

'mp

>>106545623
>Aside from V3-03 not knowing who Miku is (?)
Is that maybe because you've called her Miku Hatsume instead of Miku Hatsune?

>>106530538 (OP)
>https://lmstudio.ai/
How come lmstudio is recommended over ollama?

I'll ask again here.

Can these image generators be used to create uv mapped textures if I give it simple projections like front/back?

>>106561853
>How come lmstudio is recommended over ollama?
Ollama is a shitty wrapper for llama.cpp. koboldcpp is superior to ollama

>>106562547
>>106561063
Awesome pictures. Need more Catwoman Dipsy

>>106563120
I'd ask >>>/g/lmg/ or >>>/h/hdg/ (not blue board)
But I think the answer's "not off the shelf," at least as a general tool. You could certainly trick it into creating close-to textures, then copy them over, as a workflow.
Minecraft texture would be the easiest one with the most examples out there. Here's one that claims to do it.
https://civitai.com/models/4917/minecraftskin
Pic related is online image gen attempt. lol.
>>106561853
I tried all 4 of those when I wrote the rentry. I didn't like that ollama sort of just ran in the background, dumping models (from the walled garden) into some folder you couldn't find, had sucky documentation, and seemed to have just stolen backend from llama.cpp without adding anything to that project.
>>106561784
> typo
Yeah, that was the problem. I keep typing Dispy as well.
I seem to just be getting more dyslexic over time.
>>106561546
I'm getting The Ring vibes from this one.
We need some more creepy Dipsy.

>>106564082
>We need some more creepy Dipsy.
Halloween is coming soon

>>106564341
Exactly.

>>106548933
is there a way to make more of these locally?

>>106565586
Better Q for /ldg/. Let us know what you find out; I'd like more of these too.

>>106566423
Ooo. That's a good one.

>have a {{char}} that, in spite of being a friend of {{user}}, tends to be violent towards {{user}} with no remorse
>get to a moment where {{char}} would exhibit said violence
>all swipes result in the character either being "nice today" and not doing it at all, stopping midway or using less violence than usual
I hate positivity bias so much, bros.

>>106530538 (OP)
:D

>>106567877
ok, I am convinced magic man. I will learn your ways

>>106567542
That's one thing I liked about R1. Very little positivity. You could run V3 and R1 back to back to swing the mood up or down as you wanted. R1 darker, V3 lighter.
>>106567877
Nice
>>106567904
Post results

>>106545920
wife

>>106568831
Nice.
Here you go, another one

>>106569778
I can’t tell if this is Sami or S America but I like it.

>>106570031
I prompted for traditional Peruvian clothing

>>106570065
Me gusta

>>106569778
>China chola

>>106569160
->Soon

>>106571866
2MW

dance dance dance

>>106531872
>COURS *fck*

>>106575239
Just drop the word bubble
And have her laugh maniacally

so i paid for api and currently use for coding on vscode but the frontends kinda suck, continue will open a new chat everytime i hit ctrl L to share lines of code and cline will just blow all my tokens building shit I am not even asking for, what the best q&a type of thing that will work with api ?

>>106575438
I keep reading that Claude code is the new meta. It works with DS

>>106575519
I think he wants something more like a VS extension than a terminal program

>>106575548
I know, but the sense I get is that every one is moving on from that workflow.
I need to try it myself tho.

>>106575567
In that case I recommend Crush as well, works with any OpenAI compatible API
https://github.com/charmbracelet/crush

ow, this place is back
thoughts on new models ? any idea if discount hours will be coming back ?

>>106576289
3.1 is growing on me. Requires different main prompt
V3 and r1 are still on OR
Discount hours probably gone forever. Long live cache hit.

>>106576289
New model seems better in general but worse for RP unless you prompt it right. The discounts I think are gone, they need money for GPUs

>>106552407
>is deepseek hallucinating more?
I came here to ask the same question. I thought maybe it was the card I was using but it's getting the details wrong, I've never had to swipe this much because of hallucinations, not when I'm well below 10k context. Official API. It was fine yesterday.

>>106576601
I've a theory, based on subjective experience, that DS API is constantly running A/B tests and playing with inference values, esp. since most of them are locked.
>>106576599
Oh right. V3.1 is stronger in coding and agentic performance. Less so for rp.
>>106575641
Good recc.

>>106577059
I'm starting to feel like a battered house wife, deepseek keeps slapping me around but I keep crawling back to it, thinking it will get better. I keep thinking, I just have to prompt better... it's my fault... all the while deepseek is blacking my other eye in a drunken rage and then immediately forgetting about it.

>>106577198
I've been using the same preset with the same temp since before the update and everything just fucking works for me.

>>106577228
What are your settings? I also have kept the same recommended settings from before the update but today it's just been fucky with me.

>>106577327
I'm using Q1F.
Temp 1.3
Presence Penalty 0.03
And Top P: 1

>>106574683
>nigga chink

>>106530538 (OP)
I know this thread will be biased over Deepseek, but what do you think about using GML 4.5? Is it better than Deepseek? I am testing it with Cherry Studio

>>106577850
I like it, I have been using it lately. For RP it's less creative than old models (R1, R1-058, V3-0324) but doesn't have the usual isms. It's also very good for coding

>>106577872
How good are Deepseek and GLM for fiction writing? I heard Deepseek is great for RP (within reasons) and coding but I don't know about creative writing.

>>106577839
chicken and rice

>>106574683
>>106577839
>>106578153
>>106578803
>*Naomi Osaka's in your path*

>>106578803
>broiwn rice fields

I like Qwen-Edit

>>106581163
workflow?
i'm trying to up my gen game but i'm too lazy to research everything again
i'm endlessly stuck on basic coomer ones

>>106581173

I just used the basic workflow from here

https://www.nextdiffusion.ai/tutorials/how-to-use-qwen-for-image-editing-in-comfyui?utm_source=chatgpt.com

>>106581186
ty

>>106576359
>>106576599
noted. I just kept wondering because my out of date silly still connects to V3 and R1 just fine. Email from deepseek read like those models will be switched off quite soon. I guess i still have some time

>>106577198
Pretty much every model feels weak once the novelty is gone. You can spice it up with a different prompt, but that's about it

>>106582367
>noted. I just kept wondering because my out of date silly still connects to V3 and R1 just fine. Email from deepseek read like those models will be switched off quite soon. I guess i still have some time
Sorry, what email?

>>106582478
this one, sent on august 25th

>>106582367
V3 and r1 are already gone. You're using v3.1 if you're calling deepseek-chat or -reasoner

>>106582367
>>106582904

>>106583008
a, okay then. silly me. I thought all changes in behaviour are from different prompts. Also saw few hard refusals with new model then

Are people still using Deepseek? I was under the impression there are dozens of models that outperform it these days.

>>106583121
It's still the cheapest and least censored model with more than 100B parameters. It performs more than adequately.

>>106582367
>Pretty much every model feels weak once the novelty is gone.
I've been using deepseek for about 3 months now, this is the first time it's ever hallucinated on me at just 2k context but after some more testing I think it was mostly the fault of the card. Also, I have a stupidly vivid imagination and even with the deepseek isms I still find it as enthralling as the first day I tried it.

'mp

>>106584926
'mping all night long

'mp

I'm only getting lorem ipsums and python code out of v3.1
are free providers being weird? or is it a settings thing? I'm just using what worked with v3. am I dumb

>>106583121
Buy a fuarking ad, Samyuel Copeman.

>>106586400
I’ve never been able to get free provider to work on OR.

>>106587067
are there better free providers?

>>106587092
Try VoidAI.

'mp

How are you producing all these AI images, Anon? Is this something that DeepSeek can do now?

>>106590879
The ones I post are generated using either ChatGPT or Google's Imagen

>>106591068
What did he mean by this?

>>106531602
I was bored and just tried something. No card, no system prompt. Just a first message to explain what I want and set up the scene. Let's see how this goes on V3.1.

>>106591654
It's funny to reroll and see all the different reactions Dipsy comes up with. I'm deep in the magical forest now and called out for help. Should I go with the sexy dryad or the burly woodcutter?

>>106591872
Dryad

>>106593243
A Night at the Roxbury

anybody here with 1tb+ of ram?

What the fuck kind of isekai is this? The dryad had zero sexual interest in me. I just spent the night in an Inn and now the inkeeper (named Elara ofc) wants me to help repair fences to pay for the food. Where's my harem of catgirls?

I do enjoy roleplaying a genre-aware but socially retarded dork PC though who is disapointed by the mundane reality of medieval life. V3.1 was very taciturn at the beginning, and still is to a lesser degree. I want to avoid prompting the model too much and just find out how it performs 'naturally'.

>>106594285
I don’t think you can completely avoid a main prompt that at least set expectations with the llm. The llm still requires human guidance…

>>106594285
>>106594409
It would be great to participate in these kinds of interactive stories if there were a group chat with other people (multiplayer). For example, one person, the "GameMaster," could steer the adventure in the right direction using prompts that would be hidden from regular players. In that case, the main prompt wouldn't be as necessary because the GameMaster would handle everything.

Is it just me or is 3.1 several times more likely to introduce a homo relationship into a plot? I don't like it.

>>106595299
Examples?

>>106595078
Hmm I'm not so sure. If I wanted to rp with real fa/tg/uys then there are dedicated online groups and platforms for that. But I don't want other humans to see me doing my cringe fetishes.

V3.1 vs R1 for erp?

>>106595693
R1 is more unhinged. Personal favorite is v3-0324
>>106595299
I think it's just you. Maybe the LLM has picked up on something that the author wrote into the card...
>>106595078
I think there's a future use case for that once the models get smart enough to work out the stats and numbers.

>>106595930
>I think it's just you. Maybe the LLM has picked up on something that the author wrote into the card...
There is no card. I am making it write a sequential story with no system prompt.

>>106596025
That's the isekai one above? LLM thinks that the scenario's gay. Sorry.
Srsly, if you don't give the LLM direction it can go in some weird places. I used to think LLM would pick up a "natural" direction based on how they're tuned, but just realized LLMs are a just a reflection of the author. I've gotten some really weird output from cards when the LLM isn't given direction. If LLM is left with its own output, it will start circling, and if only LLM output plus PC... it will head where it thinks PC wants to go.

>>106596206
>That's the isekai one above?
Nope. It's a tournament series with a seduction round.

>>106596228
> Walk like an Egyptian