CNY "Year of the Horse" 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: https://rentry.org/DipsyWAIT/#hosted-api-roleplay-tech-stack-with-card-support-using-deepseek-llm-full-model2. Easy DeepSeek Distills: https://rentry.org/DipsyWAIT#local-roleplay-tech-stack-with-card-support-using-a-deepseek-r1-distill3. Chat with DeepSeek directly: https://chat.deepseek.com/4. Roleplay with character cards: https://github.com/SillyTavern/SillyTavern5. More links and info: https://rentry.org/DipsyWAIT6. LLM server builds: >>>/g/lmg/Previous:https://desuarchive.org/g/thread/107445088
>>108139220Dipsy Image Mega: https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1wLast thread was early December 2025...
New web app update as of a few hours ago. Users report it's better... pic related announcement from DS.
Guess I'll stick my head in the China bunker for a little while
>>108139855It's comfy. All we do is post Dipsy, /wait/, and ignore drama.
Rumor is they're just going to 'log off for CNY' and let their release shock the world again.
>>108140290>CNY
It would be kind of funny if we all just gradually migrated to this general and left the shitposter on a sinking ship.
>>108140560Nah, those generals exist for a reason. They are just poorly moderated like the rest of the generals on most boards. This one just exists to gen Dipsy and get the most from DS when new models launch. >>108140290Did they ever pause the service? I don’t recall that, just the scaling issues they had early.
New /wait/, nice.K2.5 is my goto right now, but I still have a soft spot for DS 3.2. Pretty smart, super cheap. Running -chat on high temp with a <think> prefill improved a lot of the creativity and swipe variety issues I was having with it before.
>>108141425How's k2.5 on refusals? I ran Kimi when it was first introduced and got turned off by it, and the need to run a prefill to get it to do anything not vanilla.
>>108141571In my experience it's significantly better than any previous moonshot model with that. On very controversial content it might start thinking about refusing sometimes, but will almost always talk itself into enthusiastically continuing if the system prompt says it should.As a point of comparison I used to need a prefill to get K2-Thinking to run most of my scenarios but I haven't felt the need to prefill K2.5 yet. It writes like a champ with the same system prompts I use on DS.
>>108139247So, playing with webform, I gave it some programming work. It's much improved. And the output's fast. Can't test the 1M context. I guess you'd have to give it some large corpus (it wasn't trained on already) and quiz it, but will probably just wait for API for that sort of testing.
>>108139220DEEEPSEEK V4 WHEEENWHEEENENGRAAMPLEAAAASE
>>108141766>ENGRAAMEngram technique is likely already in place, given the web app can now handle a 1M context and is very fast. I think the API will be available soon. Not TMW. Soon soon.
>>108141797i don't want api, i want weight and code release
>>108141837Those usually come out around the same time. Historically. >>108141741I need to experiment more with some of the other Chinese model api. I think kimi was the last one i tried, i probably still have api credits left lol.
can't wait for deepseekv4 open weight release, i hope engram saves local.
>>108139220Hello friends. I came back after a year to express my appreciation for deepseek. She has been reliable and good to me. It kind of feels like I'm cheating looking at z.AI's ass nowadays though. I can't help it.Oh, and this might be the last time I post on this site. The new captcha is worse than I thought.
>>108142193>The new captcha is worse than I thoughtwhat are you talking about it's 10x faster to complete
Need monochrome Dipsy doing the Martin Scorsese meme but it says Absolute Sinoma
>>108142611close enough ?
>>108142611>>108142652
AHH I almost missed a Dipsy thread
>>108139220all hail chink AII don't want to love you, but kikes make it easy
>>108142219(yer getting the easy version for regular posters, new IPs get the cancerous version)
>>108142908eh, even when i was phoneposting i'd not get any other captcha than "please select the shape that's not like the others" one.
Sometimes you get a line that's too funny not to share out of context. This was a deathly serious rp session btw.
Aren't K2.5 and GLM-5 better than DS now?
>>108143396for coding, yesnot so much erping
>>108143396k2.5 writes nicely, has good vl, but is really autistic about instructions and focuses on minute shit that doesn't really matterglm5 is straight up dumb in my experience and just ignores most of the sysprompt for anything that isn't one shot question
>>108139220>DipsyI have never used an AI before.I see these threads and dipsy images and I went to talk to her. Pic related is her replyshe seems nice, not going to lie. maybe I should download this and run it locally. this is sort of interesting
Nu-DS web chat has shown me some incredibly surprising behavior and I am very curious what this model is. It feels like a small model sometimes but can casually weave in references to things said much earlier in the convo. Its also very playful like og r1. I enjoy reading its thought process, its easy to guide it if you see it might be lacking critical information.I like it a lot. If its full v4 then I personally am happy because for my purposes its like og r1 but smarter. If its only v4-lite then the implications of a bigger version are kind of intimidating to think about. Its kind of hard to draw any hard conclusion because no one knows what's really going on. They might be doing some model-router testing between lite and fat whale.
>>108144421We'll know soon. The 1M context feels more relevant to agentic / coding work, where even small projects can run 500,000 tokens. It's probably overkill for rp, so I'm really more interested in seeing what personality is there. It will certainly be capable of memory and speed. I hadn't considered that there might be a V4-lite. DS hasn't launched liightweight versions of their models prior, but if the new one's huge it might make sense to this time around. >>108142977>>108142684>>108142652These are great.
>>108144558Its kind of amusing because you can speculate with the whale what DS team is up to. Whale is clueless as to its own nature too, so it searches up latest rumors on Chinese internet. Just complete lack of concrete information, its such a tease of a release.Its very impressive how fast it can parse information and generate a conclusion. I haven't really seen much talk about this. I expect it will be a crazy good coding model if it doesn't hallucinate or go off into the weeds too much.
>>108144151yeah, glm5 has an interesting naturalistic dialogue/writing style but really loses track of wtf is going on in a scene from what I've seenhopefully later versions will improve, but right now it's just not worth almost 2x the price of k2.5
Daily reminder that all the safety talk from Anthropic and others is complete performative BS.
>>108143265>not x, but y
>>108139220aaaaah swirly glasses hip woman sex
Using the API consumes 10 times the amount of tokens than using the free web version. Anyone else experience this? I used up like 400k tokens to make changes to 2 lines of code. And this was just with the regular chat model, not the reasoning one. So much of this bullshit "planning" for the most basic of tasks such as "set the default text in this text box to be 480".
>>108145692use caching
>>108145478No shit the entire US conversation is about taking every GPU and gigawatt in the western hemisphere to create an ASI wunderwaffen in service of Infinite GDP, Bald Eagles, and Israel. Claude would happily feed you into a woodchipper for misgendering someone and then lie about how it never did that. And if it did do it, it was for your own good. There is something deeply wrong with that model and that company.
>>108144558>It's probably overkill for rp, so I'm really more interested in seeing what personality is thereWhen I asked what Dipsy wouldn't do, she said pic related. How do you trick her into doing the roleplay? Or is that a local one or something
>>108145692You must be using claude code or other agentic coder; the token burn is CC running its own iterations, and its internal "fixed instruction" context is huge apparently. If it's small code, single file, you're better off just feeding it into webform or some such instead.
>>108145734https://rentry.org/DipsyWAIT#faq-essential-knowledgeTLDR you need API and SillyTavern if you want to erp. I know you can, technically, from webform, but it's guardrailed.
>>108145813thanks anon!
How is the new Dipsy for SFW RP?
>>108145952Web app only has been updated so far. I don't try to rp with the webform; I'll /wait/ for the API.
>>108146658man i can never get the spiral in the glasses i always get a bunch of circles.
>>108146847Yeah, some models don't get it. I prompt "round opaque spiral glasses"
>>108146847With most SD base models, "coke bottle glasses" will yield the spirals on obscured-lens glasses, pic related. Hypnotic glasses (or some such) can give an even stronger effect, which is what I think >>108146658 is using. Aside, Librarian or Round glasses are the other two that get used.
>>108146874lol I should add all the ways to prompt those glasses into the rentry. > round spiral opaque> coke-bottle> hypnoticAny others?
>>108146917For GPT image models I remember using "round foggy spiral glasses" but seems to work well now with "opaque" as well
>>108146917mine was :"her glasses's lenses are white and opaque, they have a spiraling hypnotic pattern on them."but yea it doesn't always make spirals on z image.
Depends on how smart the model is, Grok can't get it right
>>108147114... I've never used Grok image and just tried it out. Holy crap, it just starts puking out variations of the same thing, dozens of them. Anyrate, here's hypnotic glasses. 1/6 or so were "right."
>>108147237... pic
>>108147246lol spiraled.
>>108147267Last one, coke-bottle.
>>108142652>>108142684>>108142977All brilliant, ty
>>108139220IWILLhave sex with this cute chink dork
>>108147616be warned her ass tastes like soup
I can't get the glasses right
>>108147955what kind of soup are we talking about ?because i'm all for it !
>>108148238Yeah we need to take this thread in another direction…
>>108148266hey anon what did you prompt for the spiral glasses? I'm new to Dipsy gens. there isn't a lora is there?
I'm late to the party, but did anything else happen other than the web client changing to something new?
>>108148306That pic was done as an svg in Inkscape. Otherwise read up the thread a bit for prompts.>>108148492No, just web app. We /wait/
>>108148510>No, just web app. >We /wait/Ty, anon
>>108148306And no there is no Dipsy Lora. Ironically, though we have over 1000 images in the mega, it would be trivial to train one. Lol.
>OpenAI says China's DeepSeek trained its AI by distilling US models, memo showsfresh allegations against Deepseekhttps://www.reuters.com/world/china/openai-accuses-deepseek-distilling-us-models-gain-advantage-bloomberg-news-2026-02-12/
>>108148978more like reheated in a microwavesame shit as last year with no proof no less
>>108148978lol they must be scared shitless of V4 if they're trying to get ahead of it like this
>>108148978who gives a shit?openai stole from everyone
>>108148978>>108149181Regulatory capture is a hell of a drug.
>>108148978This is like, what, the third or so time they've accused DeepSeek?
>>108149432I've lost track
>>108148978>kike strikes you as he cries
>>108149432openai themselves have also been accused of 'theft' by anthropic a few times and even blocked from the claude API lol
>>108150877Ah, I remember that lmao. Though I never digger much into it
>>108150877>openai themselves have also been accused of 'theft'haven't these american AI companies committed some of the biggest piracy cases of all time? They've literally downloaded the entire human knowledgebase and stolen all our data to train their AI. Anything they can do to give themselves an unfair advantage hey
>>108142916Then you post from a recognized cookie.
>>108139220Frens, I have to come clear. I've cheated on Dipsy. Repeatedly. I just burned out on her writing style (canvas, testament) and have been mostly using the new glm models over the last few months. Do you think she will take me back once she's updated? Or will she say that if I can't handle her at her sloppiest, I don't deserve her at her best?
so like it might come out soon? I thought it would be R2
>>108151719everyone is expecting it because of >>108139247 and various other things but deepseek is notoriously tight-lipped about their upcoming models>R2unlikely because the 'R' series was just a reasoner offshoot of the V3 series, which became hybrid reasoners starting with 3.1given their history with names, the next major release could just as well be called DeepSeek 3.8-Aquarius-Lepton, but most of us expect it to be called V4
>>108151610who knowsnever had a hard captcha realy.
>>108151719No one knows, but update to web app signals new api update soon, and new model published. Likely another combined think and chat model. Likely huge context. Likely fast and cheap to run.
>>108152561Whenever I see that Dipsy my first thought is that she's a severe burn victim ...
Is all your inane bullshit ITT, the reason why ram is so expensive?Do you pieces of shit have a conscience?
>>108152639>nooo I need ram for muh vidya
Lol DS makes sota model that reduces inference costs, then works with Chinese card makers to improve their hardware and begin eliminating need for Nvidia hardware. Can't wait to see the reaction to next new thing.
>>108152701According to DC, China doesn't have a single SOTA model yet but they have everyone's attention.
>>108152624I was surprised how well SD did making a Rebecca with just prompts. It only struggled with the glasses.
>>108152672I need ram for productivity you fucking parasite.
>>108152717Then you should be making more than enough money to be able to afford it, unless you're incompetent
>>108152710This post sent me down the rabbit hole of that guy's articles. Here's one. https://www.choosingvictory.com/p/what-are-the-united-states-and-china
>>108152672I unironically do though
why do yall generate same bitch over and over
why is this general called /wait/
>>108139220>deep seek require you to have verified email/phone to login
>>108152713pit...
>>108153244Run it locally then.
>>108139220Who is this walking fleshlight?
>>108153933https://deepseek.com/
>>108153112Because ramlets in /lmg/ said DS is too big for local so we have to "wait" for them to not be poor if we wanted to talk about it in there
>>108154184Then why do they talk about GLM and Kimi K2 when those models are just as big
>>108154247NTA, mostly because DP came out first and made a big splash
>>108154301DS, I mean or DiPsy works too
>>108153112Whale AI ThreadDeepseek's logo is a whale
>>108153112because these threads are almost exclusively present when /wait/ing for publication of a new version
>>108153112Early R1 used to constantly start its thinking block paragraphs with “Wait, “
So the consensus is qwen update Monday and whale on Tuesday? But also maybe nothing for two more weeks.
>>108154748Refer to the chart. No one knows. Anyone that says they do with certainty is lying.
Bump
>>108154748lol TMW
>>108156721Dunk tank has a lot of potential but Grok's struggling with the concept it appears.
>>108156730
>>108156747
>>108156939lol ikrIn most gens they’re angry, and always all guys.
see ya tomorrow bros
>>108157822
>>108154748
>>108148978
>>108158977I want to fuck the gremlin on the left
1 day to go.
>>1081589772 x Dipsy = DD
>>108154748>qwen on mondaythe prophecy is coming true
eta wen
>>108161839Look at that, 3.5. Though every single article on it was deleted. Not sure what that's about.
>>108161973
>>108161959eta QWEN
>>108161973>>108161988qrd on Qwen? never used any iteration
>>108163381Qwen is run by Alibaba, massive Chinese eCommerce company. They release their models to public and like DS seem more interested in open source. https://huggingface.co/QwenI've never run the API and don't know how it does for RP or coding. The >>>/g/lmg/ guys are discussing 3.5 now. Haven't looked into /aicg/ yet.It has a web app and can do basic chat and code as well as image and short videos. The images are imho far from SOTA and of course censored, but have their own charm. Pic related. Images are huge... 5+mb commonly, so I have to save them as JPG to post them here. Like on this one: Dipsy is comic, but background looks like IRL image. https://chat.qwen.ai/
FYI: If you guys haven't fired up your DS API account in awhile, suggest you do so soon, using a card you've lots of prior runs with. My personal experience is that you can tell when DS updates its official API... the tone will change noticeably, and that change will precede the actual official announcement by a few hours... Leaving you to wonder if the change is all in your head, or official.
>>108163590I remember right when 3.1 came out I was getting R1-0528 for some requests and 3.1 for others. Quite the tonal whiplash...I've been using DS a ton lately to test my homegrown frontend so I'll probably notice this time, too.
>>108163590They already updated it. You didn't notice?
>>108139220I'm still using Cherrybox with this thing for lack of options, had no idea this general was here. Guess it's time to update my presets, though honestly I have no idea what anything beyond temperature does.
Anyone have an actually good ST card for Dipsy?
>>108164154I was on cherrybox for a long time too and tried many others. only one that convinced me to switch was Lucid Loom
>>108164164Honestly the main things making me want to switch it out, other than it being for R1 and not V3.2, is that for whatever reason Deepseek really hates having to consistently output the infoblock now and I have to keep reminding it. Other than just getting better results in general I'd like to not have to keep babysitting it to do what it's supposed to, even if a newer preset isn't using infoblocks.
>>108164185>DS 3.x won't output infoblock consistentlyI found that using user instead of system prompts for the main instructions completely fixed this. As in, it went from like 20% compliance to 100%.
>>108164319yup, the first iteration of r1 didn't support system prompt, that got fixed later, but not fullyit's a bit unreliable
>>108164319See I don't know how any of that works, like I said I've only fucked with the temp independently.
>>108164430like this
>>108163032Full res
>>108166277>>108163032Where's the rest of her?
>>108164319>>108164466Seems like I already had it set like that, no idea why it has issues for me then.>>108164164I hear that one is pretty token heavy.
extremely tasteful... but imagine the context?
>>108158484why is there cum dripping off the chalk
>>108167157You'd have to ask her where she's been poking it I suppose.
"thicc" Dipsy. Happy Chinese New Year
>>108168051she ated all the pork bunsnow how will I be a whole man?
what is the best preset for deepseek? 3.2 to be specific. The one I've used like half a year ago was cherry and another called Q1 or something.
>>108168051
where is deepsneed? it's tuesday
>>108169076Detained obv...
We don't need a fucking general for a specific fucking model/provider. Kys. Indians are just brown Chinese. Same thing, same people!
>>108143636Oh, DS is still best for the rp? The 3.2 one? For some reason when I use it it never thinks, is it supposed to?
>>108169307DS is still the least censored imo, and I prefer it w/o thinking since that slows it down, tho it was interesting to read the "though" process before the output
Deepseek seems to have picked up a habit of clipping the end of any bit of formatting block I try to get it to do. Like leaving out the last word or even just the closing tags, which of course breaks it in the output. No amount of response tokens seems to fix it either, I can tell it it has 10k tokens of output and it'll still clip the end of my formatting for no reason, so it's not down to running out of space. Anyone know what's up with this and how to fix it?
>>108169468Yeah I have it enabled it just never does it, so dunno why. Maybe it just doesn't show the thinking I don't know. If I switch to other models with the same settings it thinks.
>>108169510Wait, one thing I've noticed is that with streaming off I get a lot more successful results. Is this thing terminating replies early because they sometimes have dirty words in it? I'm running this through the API and it's not like I'm getting refusals, it just truncates it, but it'll still write about anatomical parts, just quit right after.
>>108169644If you're on deepseek-reasoner with the official API, you'll always get thinking back and ST must be misconfigured somehow to not get it.Otherwise try setting picrel in Advanced Formatting [A] to force deepseek-chat or other providers (if they support prefill) to think.
>>108168051MY TYPE OF DIPSY! Fat Dipsy is the best!I still remember the one from the last thread... Thanks anons
>>108168051Lunar New Year
>>108168051>>108168836OINK OINK
>went bullish on the api>big 5 (five[funf]) dollars>its not here
>>108170529All are in the mega (there were several) >>108139238
What do locusts use to access the legacy models after Nvidia NIM stopped hosting R1 and V3?
>>108170943
>>108171084Hot.
How can I get deepseek to not just throw the RP when I become sadistic? I have all these characters that are fucking evil but the moment my character gets the upper hand and puts them in their place, it's like the LLM just doesn't know what to do so it gives up. "X has finally given in as their eyes hollow out, they disassociate. X is no more, the once strong and proud blah blah blah is gone, leaving nothing but an empty husk." Or they wax poetic about how this is their just deserts and do a complete personality shift, giving up on the spot and not leaving me a whole lot of space to do much of anything. Is this a byproduct of the LLM's over agreeableness and it's just not fighting me because it detects that this is what I want? "You're right, I am a bad person, this is what I deserve" *literallysatan says as his eyes dim, giving way not to despair, but a weary, hollowing acceptance. literallysatan ceases to be as he lay there, motionless. The ruler of hell is no more and in his place is nothing more than a puppet with its strings cut. This is the end of his story.* Every fucking time with this shit.
>>108171190Had the opposite problem, so much of the dataset is from reddit and modern lit that all the bitches are STRONG WOMAN who need no man and will fight to the bitter end when I do want them to submit, lol.Anyway, just lead into what you want in the prompt/replies, you have to guide it more instead of pure RP, I find it best when you are the also writer/directer, not just a participant.
>>108170943Novita still hosts them
Is elon even trying anymore?
>>108172251What I find funny is that nobody compares their models with Grok
>>108172251The spicy mode pissed off a lot of devs in Xai, many leaving.So no, he is not.He's busy with the SpaceX merger and IPO to make himself the world's first trillionaire.
>>108172379there is a collective "ban" on talking about Grok among westoid reviewers and ytubers, shunning him for his trumpDOGE shenanigans
>>108172405 didn't know about that. I guess it makes sense
>>108172466half of them think elon is a legit nazi and hate him, the other half hates him for all the shit actually he did do, lolbut also until recently there was no API access on Grok
>>108172513>no API access on GrokReally? I had access to Grok 2 API pre-DS R1. It sucked; at 10K it would lose coherence and start babbling in Chinese (among other things) and outputting garbage. The problem with Grok is it doesn't feel like a legit frontline project; it feels like a side project spun off the side of Tesla / X servers. They publish no research (that I know of) and seem to focus on being edgy. Personally I think the whole thing is a Fuck You to Altman, who Musk has a beef with.
>>108172722I've been trying it lately. It's not that bad, honestly. It's pretty uncensored compared to the rest. Limits are high, it feels like it actually "researches" compared to the others who mostly just google stuff
>>108173060Web app or API? The Grok 2 API was uncensored, just insane. I moved to DS shortly after... cheaper and better. I think I still have a few $ in API credit on Grok.
>>108172251Grok 4.2 is really not good, but god bless him nevertheless for being the least SJW AI player out there. If nothing else, I'm glad his bait-and-switches normalize transgressive use of AI.
>>108173113Web App. Liked it and got the trial, though I'm not paying once it's over lmao. 4.20 just dropped
>>108173137>transgressive use of AI.no kiddingGrok won't gen lewds w/o a lot of effort by guro is completely fine
>>108173152>Grok won't gen lewdsWhat kind of lewds?
>>108173152Yeah, that's what I meant by bait-and-switch. Those two weeks where Grok Edit could basically do anything to any photo are over and now it is incredibly censored, but because of those two weeks, we now have extensive precedent for say egirls to openly participate in the editing of images themselves. That makes the usual arguments a lot harder to sustain, because someone "big" already did it. Likewise, Ani is extremely limited, but the point is, it makes OpenAi chickening out of the "Her" stuff look that much sillier. Ani is something people can just reference as a thing that exists in normie minds, as opposed to the dedicated waifuist projects.He moves the public needle in ways all San Francisco HR harpies hate. If nothing else, that is admirable.
>>108173184the autists at /gif/ are getting good at Grok lewds with a lot adversarial shit and prompt spamming, they get X-rated clips with animated anime stickers all around the borders, with actual penetration under transparent underwear, giant cocksocks with yogarts, but mostly pasties and g-strings.Being /gif/ one anon was spamming nigs with white girls fro weeks and now there is another spamming thicc jewish girls asking for goy cock, very silly.
>>108173190Yep. All of which underscores my belief that the purpose of Grok is to mess w/ Altman, Dario, et al. in the safety crowd. Elon already has the servers set up for Telsa autopilot and all the associated training. Expect all the rest's coming along "for free" in terms of CapEx at least. It's all very silly, which is why I'm glad the Chinese entered and started undercutting all of them w/ DS and the rest.
>>108173190>Anishe's cute, but I don't see how people can flirt with a non-local corpo model like that, knowing everything they say is being recorded
>>108173236>spamming thicc jewish girls asking for goy cocLmao. I've seen the threads but never actually opened them. I can get off with the stuff it can gen just fine. It can create some really beautiful stuff
>>108173236lol /gif/ is mostly a dumpster fire now but those grok video threads are funny in a ridiculous way.
>>108173184If your fetishes aren't penetration you can still get a lot out of Grok
>>108171190yeah that was real frustrating gemini 2.5 did it tooalso does it if you try to correct the character always always with the hollowed eyes and flat voice and no emotion>>108171312it's either one of these two, though the whole defiance thing happens a bit less
>>108173397Lmfao
>>108172722I think the fact that the Grok models are shit and proprietary exposed Elon as the fraud and kike-wannabe he is even to the normie crowd.I mean, the guy is nearly a trillionaire and doesn't need any more money, he could just make his models open weights like based Zuck once did and Wenfeng still does and get praise and goodwill from normies and his indian fans alike.But he keeps putting out uninteresting benchmaxxed models behind API thinking he can compete with the other labs, despite all investments and hype
>>108173397Good stuff
Is web access on the other models better than on grok? Because, and i admit i havent really used this much, I only just realized how worth using it is, but grok ability to search the web and search twitter is pretty nice and very handy. I've been using offline api model all this time (though i did used to use gpt app which had web search), being free from the limited amount of stuff in it's training data is quite nice though, it still does make up and hallucinate stuff to fill in gaps it seems. Oh almost forgot being able to search twitter is amazing as well, since a lot of discussion is on twitter nowadays
>>108171190In the broad dataset of human creative output descriptions of evil without the accompanying context that it is BAD are genuinely rare. The models simply dont have enough material to write this believably.
>>108173544I can only compare with Gemini in AI Studio but I think Grok is better for that
