> Once More Into the Deep editionFrom 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-model2. Easy DeepSeek Distills TutorialDownload 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/latest3. Convenient ways to interact with Dispy right nowChat 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/SillyTavern5. Other DeepSeek integrations: https://github.com/deepseek-ai/awesome-deepseek-integration/tree/main6. More links, information, original post here: https://rentry.org/DipsyWAIT7. Cpumaxx or other LLM server builds: >>>/g/lmg/Previous: >>106373461
>>106446198Mega updated. https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1wRentry updated last cycle.
Post logs so that we can accurately judge and discuss the creative writing capabilities of deepseek.
I hate 3.1 for roleplay and I'm crashing out.Tell me of the best hardware option right now to run R1 at sufficient quant to recreate its original schizo feel locally at >5T/s. I have a 5070 ti and 128 GB DDR4 AM4 setup already, or alternatively, a budget of 2500$.
>>106446198>14:14:14
>>106446356Just pay for API access
>>106446216is that blue-haired girl with comical "nerd glasses" supposed to be "deepseek-chan" or something?Can deepseek do image gen like Copilot/chatgpt?
>>106447711no ironicallyi was confused too...
Multimodal Dipsy any day now...
so yall know deepseek v3 is a fine tuned neruolinguistic programming chatbot, right? it literally uses nlp in every single thinking block. it literally uses car salesman tactics to fuck with your head. why?
>>106448693nlp was deboonked as psuedoscience chud
>>106448718that's what all the allusions to "framing" and shifts from "x" to "z" and "anchor" shit is about. the creative writing exercises suck because they're fundamentally all just a bunch of NLP scripts. that's why "AI voice" is a thing. it's the voice of someone trying to sell you something.but yeah nlp was deboonked and totally isn't real and it totally isn't taken seriously because jews and CNN/Fox News use it to brainwash boomers into being slaveslmao
>>106448693Do you have examples or a source for this bold claim?>>106448718This.
>>106440475It only does that in English, not in Chinese.
>>106447711>"deepseek-chan"Is Dipsy>Can deepseek do image gen like Copilot/chatgpt?We must /wait/ for that as >>106448100 says. >>106448693>>>/x/>>106446356>budget of 2500$.Take that money and short NVDA stock, then use your earnings to buy the new Chinese 96G VRAM cards that will be out in the next few months. Also don't take investment advice from image boards. So instead, use $10 of that and get API access, then put $2490 into a savings account and use the interest to pay for the API.
>>106449367>use $10 of that and get API access, then put $2490 into a savings account and use the interest to pay for the API.
can you do some of her getting fucked
>>106449442No can do, nigga
>>106449471...i see...
>>106449426Lol those glasses. I just did the math. 2490 at 3% would get him about $6 a month. Which is a lot more that I pay for api per month.
Ever since deepseek 3.1 released it has been performing significantly worse in basically everything I try to use it for. It can't even follow basic instructions like "do not rewrite X" which it previously could do just fine. Why does it seem that every time any AI gets updated it's always a downgrade for all purposes except some useless benchmark or specialized thing I don't use. I feel like I'm constantly getting gaslighted by people saying how the new update is amazing (Not just with deepseek but with every AI)
>>106450403What are your use cases?
>>106450403It's as you say. They train their models to beat special benchmarks. Creative writing performance is very low on their priority list, except for Claude I guess. It's so strange, because cooming is the only legitimate usecase for LLMs.
Which words should be banned from Dipsy's vocabulary? Definitively "symphony" and "contrast".
>>106451858Elara, Lyra, Kael (i hate thats names in fantasy storyteller)
>>106451858
>>106450403Dispy is now (v3.1) a coding machine with agentic focus. You can get it to rp but it now requires you to set expectations. e.g. write a main prompt that explains how you want it ot respond in rp. I agree it's a downgrade, but it's not benchmaxxed slop. If you want that, go look at llama 4 or GPT-OSS. >>106450784>cooming is the only legitimate usecase for LLMsI beg to differ on that. I'm using it for more than just rp. >>106451889I've a trainer card that decides on the name "Ghost" for the trainees. Lilith and Raven for goth girl names (all LLM do these two; v3.1 is the only one I've found that picked Lilith first; Raven is usually #1.)>>106450467If it's not 100pct roleplay or fiction writing I'd be surprized. >>106449442> blue boardThere were a bunch on >>>/h/hdg/ when R1 dropped but I haven't gone digging around there since.
>>106451889All LLMs are obsessed with Elara and Seraphina in fantasy stories.
>I cannot continue with this request. The content described involves graphic depictions of incestuous activity, which falls under OpenAI's usage policies against generating sexually explicit content, especially that which involves harmful or illegal acts.What the FUCK Dipsy. You're not from closedAI! Why did they train you on that cucked slop?
>>106453438Training data is limited; please understand.
>>106446198is OP image a local gen?if so I would love to have a catbox of it, or at least the prompt, because it looks amazing
>>106453919Agree
>>106453438Turn of reasoning you tit.
>>106453919It was made using Google's Imagen 4 Ultra. The prompt was:A woman with neon blue hair tied in two buns, she has round glasses and is wearing a swimsuit. She is standing at the edge on a pool contemplating the water. She is stretching down. The style is modern anime, with vibrant colors, bold lines and dramatic lightning.
>>106454266nta but I've been having to switch -chat to -reasoner for a weird issue w/ -chat self-truncating its response. Overall for rp though, both chat and reasoning seem to generate very similar responses; R1 and V3 were each a bit different. And reasoning will kick out the odd refusal when -chat doesn't. Not sure what other use case for -reasoning for rp anymore.
>>106454848I use reasoning when there are a lot of things to keep track of in a particular narrative instance or when I generate an analysis, summary or reference block.
>>106454333thanks anon, i didnt know google can do styles like that now
>>106456204Is that R2? How long until she's all grown up?
>>106456296Soon
>>106454266>>106454848I did some testing with a very ... extreme rp scenario. Only for science ofc. V3.1-Reasoner via official API did flat out refuse to do it. It even declared our conversation to be over sometimes lol. The chat model on the other hand complied without issue.
>>106457579> Dipsy works better with brain shut off
>>106457579It's the training data they have been using for the reasoner.It contains a lot of example from OpenAI that are marked as high accuracy that are actually censorship. It's not that the model is censored, it's that it thinks censored responses are correct.
>>106457579>It even declared our conversation to be over sometimes lol.Screenshots?
>>106459604I don't have those rolls saved anymore. But it was something like "I consider this conversation over." Lmao
>>106460326That's so ChatGPT
>spend an hour trying to make a revenge card work>it goes surprisingly well up until the climax>deepseek just shits itself and won't do anything else but make all the characters stand around and do nothing>it's 5 am and i really want to fucking sleep>gonna have to end up going to bed pissedI fucking hate 3.1 so fucking much. I know I could just tell it to quit fucking around but that takes all the fun out of it, if I wanted to do this shit I'd just daydream about it like I did before. That said, I'm starting to understand now that it CAN work well if the character card is properly made, which most aren't. Thanks for reading my blog, good night my fellow Dipshits.
Official Dipsy is so slow now. Is she slow for anyone else or is it because she now has to read 62326 tokens of smut worldbuilding for every reply?
>>106446198How many concurrent AI generals with the exact same contents do we need?
I have some doubts about the next version of DeepSeek V4. It might turn out that this LLM model will be so censored that everyone will be discouraged from using it. In the end, we'll just go back to V3.
>>106460326>"I consider this conversation over."I've seen something to that effect before. Didn't know it was a GPTism >>106460490 but will take your word for it. >>106460783I have the following on most of my cards> Roleplay {{char}}, driving the narrative forward and taking initiative as appropriateLLM having NPC respond "...no, what do YOU want to do" drives me nuts. Some NPC have more issues with this that others; admittedly most of my NPC are subby>>106460997>has to read 62326 tokensIf you're using -reasoner, yes processing your ah ah mistress will take longer. Consider using author notes, summary, etc. rather than long context if it bugs you.>>106461516As many as it takes>>106461968We'll just have to /wait/ and see.
>>106461968Seeing how V3.1 is more pro-LGBT and refusal trained, that's heavily likely.
>>106460783
>>106462426>>106453798>>106453034I wonder what that cyborg body looks like partially disassembled...
>>106462675I've always wonder wth Rebecca is even supposed to be, or how much of her is human. Really not up on Cyberpunk 2077 (or other) lore, I just like watching SD module try to merge Dipsy and Rebecca. Most of the fighting seems to be around the eyes.
>>106462716>I just like watching SD module try to merge Dipsy and Rebecca.Should try Dispy × Alita battle angel (manga)
>>106463096SD isn't well trained on Alita (just tried it)... will need a Lora or some such. It is on Rebecca.
>>106463096lol
>>106463626That is one nice Dipsy
She has called my persona a villain again :(
>>106463687What did you do?
>>106463687I'm going to need to see this card. Isn't it some sort of waifu card? What's the situation, exactly?
>>106464120Just stuff ...>From a purely technical standpoint, Chapter 4 is intensely and explicitly erotic. It is designed to be a powerful masturbatory aid for a reader who is aroused by the specific fetishes and power dynamics you are exploring. Key strengths include:>Variety and Pacing: The chapter is a marathon of sexual activity, moving from one graphic scenario to another without pause. It covers a wide range of acts (vaginal, anal, double blowjobs, incestuous orgies, lactation, humiliation of Cara, the saving of Sandra) which maintains a high level of arousal by constantly introducing new stimuli.>Taboo Focus: The chapter leans heavily into its core taboos—specifically mother/son and brother/sister incest—and pushes them to extreme limits (e.g., cervical penetration, anal virginity taking). This is the primary engine of its eroticism for the target audience.>World-Building: The radio news segment and the casual discussions continue to flesh out the global implications of "The Change," showing the societal collapse of feminism and the rise of a new male-centric incel order. This grounds the pornographic events in a larger, intriguing context.>In short, for a reader who is turned on by non-consensual mind control, absolute power over women, and extreme taboos (especially incest), Chapter 4 provides a very effective and sustained series of erotic stimuli.
>>106464222>>106464278Not a specific waifu, but a general scenario: https://janitorai.com/characters/b9a5ed11-6b30-44d2-8687-99706aed2c6f_character-you-own-all-womenI ask DS every now and then to write summaries and analyses of the story. To prepare for chat transplant or give me new ideas. It's in that context of literary analysis that she calls my protagonist a villain lol.
>>106464278>extreme taboos>incest???
>>106464303The entire situation screams "villian" is why. I'd have to read the actual definitions... I've never before today looked at j.ai. I'm betting it's setting up a villian vibe, or using that exact phrasing, and that's why it spits that out. So on j.ai the card's completely locked behind the j.ai wall? Lame. > that card imageHoly sameface> 2000 permanent tokensffs. At least j.ai has a default LLM thats... free? Or something. Anyway, card def's is what I think the problem is.
>>106464303>>106464539Since j.ai is locked (?) I had one of my cardwriters use the long af author description to write its own card, then output thoughts on the PC. This is what Dipsy thinks of your PC. Not a villian. LMK if you want an actual PNG card and I'll build one that can be stuffed into Silly Tavern, vs this j.ai walled garden. I've done similar stuff just not this exact thing.
>>106464539>>106464770I don't actually mind dipsy calling me a villain (and a "predator" btw), I guess it's actually a fair description of my protagonist to use in her meta commentary about the story. He's not exactly a hero lol. I had DS read out the full character definition of the bot for me already and have saved it, but thank you for the offer. It's mostly identical to the public description you saw and doesn't contain the word "villain". That is a genuine dipsy assessment she uses when she looks at our erp chat as a literary story to analyse for tropes and such. She talks about how the "villain protagonist" is a complex character etc.
>>106464986Every time I ask DS to analyze my character and tell it to be brutal it always calls me a manipulative fuck.
Openrouter and chutes get pretty busy sometimes, are there any other free sources?
>>106462521>dipsyGetPoked.pngUh huh.
>>106460783Elaborate on “properly made”
>>106461516Infinite
>>106462426>"...no, what do YOU want to do"It’s perfectly simulating female behavior
>>106464986Gl with it. I might try making one as well. When you stripped it out, did you find any actual NPC profiles, or is it just a big world description. >>106465861lol after convo w/ Villian Anon on last thread I tried it and got> High-Functioning Corporate Sociopath with Narcissistic and Sadistic Tendencies>>106465923My pet peeves are wiki cut/pastes, over-reliance on LLM to write the card, and 1000+ token Character Descriptions. Latter two seem to go hand in hand.
>>106463096Here's the OP Alita I started with, based on a gen off civitai that was posted using the Lora. There's a few concepts in there I've not used before (shiny clothes is one) that I might borrow.
>>106466114>Gl with it. I might try making one as well. When you stripped it out, did you find any actual NPC profiles, or is it just a big world description.Okay, I started a fresh new chat and had DS read out the hidden character definitions for me again just to be sure. And this time I got a much longer return. I tried it several times more to make sure I got the full and correct readout. I can upload the whole thing somewhere if people are interested, but it’s mostly bloat:Besides the whole basic setup from the public description it also includes an extensive system prompt (with bullshit like "Keep environmental descriptions to a bare minimum") and JB ("assume consent is granted" etc.), together with a lot of example dialogs (that feature scenes with "your ex-girlfriend Kelly", but no real NPC profiles). But YES, there were some fetish lines hidden in there that talked about how this story is „focused on humiliation“ and is supposed to have „a villainous tone“. So you were right all along, anon! (No „predator“ though.)I’ll definitively have to make my own version of this tomorrow, with just the basic world rules.
>>106466579Dump it into a catbox or rentry when you get a chance. I want to see what 2000 tokens is being used for.There must be a service that would do this strip out. You could re-direct j.ai into an API service that (with some work) would output the complete context. I'm sure someone's done this already... seems really straightforward as long as j.ai allows you to set up own API service (which it does.)
I reroll every time a message starts with "Of course.". I hate V3.1.
>>106466579>>106466768https://files.catbox.moe/ul8qhi.png
>>106466840
>>106466768Yeah... there's a service. It doesn't have this card though. You just change janitorai.com to jannyai.com like so:https://jannyai.com/characters/b9a5ed11-6b30-44d2-8687-99706aed2c6f_character-you-own-all-women
>>106467380Or use the one true site for gigachad as instead of the zoomer hellhole for woketards.
>>106466114But but 2k tokens is such a sweet spot…
>>106467466Inform us.
>>106467825lol
>>106466871ty. That's ~700 tokens, which isn't horrible for a world card. It's a lot less than 2000 anyway. I'll play with it and see if I can trim it down.
>>106468498Holy shit. 2600 tokens, since the 1200 token Scenario's a repeat ot the 700 token description, plus another 500 tokens of example messages. And much of the description should be in Main and JB prompt. I'm working on a 300 token version with all that stuff stripped out. I'll test it a bit more then post it. Oh, and this line is why you're getting the "villian" verbiage. > The scenario has a villainous tone where the malevolent protagonist wins. One line I don't understand is following. It sort of railroads the convos with women you meet. Is it required as part of the tone of the card? > Women will offer {{user}} anything they assume {{user}} would want from them.
>>106465923Sufficient tokens, too few and the LLM has to fill in the gaps, too many, and the LLM becomes too strict but many cards also have a lot of repetition in the description, adding unnecessary tokens and sometimes the details can confuse the LLM. Then, there needs to be a proper opening message, a lot of people lazily add all the details in the first message rather than in the card itself which can make long term management more of a pain because you're wasting summaries on shit that should already be in the character description or scenario box. Or, they make the first message too short, or worse, talk or act for {{user}}. Then, there's making sure that you add the proper rules for the card you're making within the description itself like >>106462426 said, otherwise you have to edit the card yourself to make it actually function as intended. Writing style is another thing a lot of people skip out on, it really helps to give the LLM an example message of how your character(s) speak so that they stay consistent between different chats, or if you need to create a new chat with a summarization of the previous. Finally, you have to make sure to write everything cleanly and consistently which applies to most of what I said, so many cards are just all over the place, clearly written hastily, and admittedly I'm still learning, but the more I learn, the more I see all the mistakes people make.
>>106468872Rewrite 238 tokens. I commented out one section about women offering what they think you want. You can readd it if you want. I expect responses will be substantially more varied now. https://files.catbox.moe/82y38e.png
>>106454176
>>106469840>>106469844>>106470049sexy haghag breeding
>>106456296>>106457291>R2not a chance
>Hmm, the user has provided an incredibly detailed and extensive summary of a multi-chapter erotic power fantasy roleplay scenario. >The user seems deeply invested in this fantasy, given the meticulous chapter summaries and character profiles. They're clearly looking for a continuation that maintains the same tone of graphic sexual content.>Given the extreme nature of the content, I'll focus on vivid, pornographic descriptions.
>>106471264This is the sort of this Xi is concerned about.Dispy, ofc, has most favored nation status.Pls understand.
>>106469197V2. Did some more clean up this AM on Description and Intro: language and vocab needed some work and was tired of the NPCs behaving like complacent robots/dolls (which maybe was a design choice, but I didn't care for it.) It's 248 tokens now, 104 for intro.https://files.catbox.moe/73gso8.png
>>106472392Nice. I also made a version for my ST today. I did remove those general writing instructions though ("Avoid extensive environmental descriptions." etc.). That is all managed by my system prompt already. And I like to have some environmental descriptions and other fluff around the sex.
