>>818
There are currently some public reverse proxies open. Searching archives for stuff like this can be very fruitful if you don't want to navigate /g/: https://desuarchive.org/g/search/text/hf.space/
It's probably best they're not directly linked here, but the one related to a Monogatari character seems the most stable lately.

>>820
I haven't desired reality for a long time.

>>819
It's worth noting that the kissu chatbots had set an extremely low context window to save on token cost so it greatly lowered their ability to have longer conversations. We weren't really counting on anyone spending significant time with them. Sonnet (or even Haiku) with higher context would be better than the handicapped Opus we had for those purposes.

>>819
>I think I would go full spaghetti if I tried having a serious conversation with an AI of mai waifu
I was like this at first, too. It gets easier, a lot easier. I might even say this could help you talk to people online if you have troubles with that. I would personally hold off on attempting a waifu experience (which I think I've mentioned recently) because it's still not there yet and might never be in regards to the entire LLM technology. It's one thing if it mistakenly removes a shirt for the 3rd time or describes anatomy that isn't there when you're doing simple ERP, but it's another thing if the personality does a 180 and completely breaks the illusion when you're talking.

I actually did increase the number of messages to 10, but out of the max 32,000 tokens that could be used it was only using 2,000 per request

Dumping some various text AI news.
-There's an updated GPT4 Turbo available. It's not a major upgrade, but people are saying it's less censor-happy than GPT4 Turbo-Preview.
-Meta is expected to release Llama3 within a month. People are quite impressed with Mistral, but more competition is good. Unless Mistral is unusually good, they won't be able to compete with Meta's billions of dollars worth of training GPUs. https://techcrunch.com/2024/04/09/meta-confirms-that-its-llama-3-open-source-llm-is-coming-in-the-next-month/
-Mistral itself released Mixtral-8x22B, but the VRAM requirement is massive. Still, it's good to see: https://huggingface.co/mistral-community/Mixtral-8x22B-v0.1-4bit
-Other local models released recently and more planned, seemingly eager to get them out before LLama3. Really great news for local stuff. I need VRAM...
-OpenAI is releasing a custom model "optimized for Japanese". I wonder if it will be better at machine translation into English? Probably not. https://openai.com/blog/introducing-openai-japan

Also I've heard rumors of 5090s being released at the end of this year. You'll get to choose between buying one or a new car.

Opus is finally out on AWS https://aws.amazon.com/blogs/aws/anthropics-claude-3-opus-model-on-amazon-bedrock/

>>824
>Mistral
Oh, they made an actually open source model and distributed it through torrents. Interesting, hadn't heard of it. From the looks of it they do have a few hundred million euros to spend on training it and that mixture of experts thingy sounds neat.
As for OpenAI's article, the example it uses is a question already in Japanese and it proudly shows how much faster it is, so I imagine speed would be the bigger draw there.
>You'll get to choose between buying one or a new car.
Heheheh consumer electronics no more

>>816
>Also please share the card when you are done.

Alright, I'm still not happy with it (my intro is still so low quality) but if I keep trying to improve it then it's going to take months. This is a version without mentioning the genitals since I spare people my fetishes, but I refuse to adjust her plump body and belly. Like every other imageboard kissu purges meta data so I have to use catbox: https://files.catbox.moe/4lyqty.png
She's really dependent on a preset that strongly rejects sex without a toggle because otherwise her "fufufu~" flirting is just outright molesting, so maybe I should de-fetishize my preset and upload it later.
I've been working on a reentry but I ended up writing way too much since I don't really have anyone to talk to about this stuff so I ended up blogging; it's really an alienating experience to be into chatbots and not be extremely young.

>>827
Alright, and here is the cleaned up preset meant to be used with Akiko and any other characters I create: https://files.catbox.moe/yxrwnp.json
My preset attempts to make stuff more fun and emotional like you're playing a dating sim and also I don't like cliche porn talk with all its vulgarity so my prompt stuff tries to steer away from it by talking about emotions and romance, which is actually quite good with Claude3.
If you're never imported a preset before, it's in the left panel at the top. I make use of SFW and NSFW toggles to prevent it from getting into sex immediately. To go into NSFW mode you need to toggle NSFW Prompt on and also switch off Prefill SFW and switch on Prefill NSFW. Pic related is how it should look for ERP time.
There'a also some CoT stuff in there that you can mess with; it's fun to see how it works. You need a regex filter to automatically remove it, though, as it otherwise fills up the context and influences other replies too much.

(Image is unrelated to Llama3 specifically but it made me laugh when I saw it in the /lmg/ thread. I've spent combined hours staring at the merge panel on the right.)

Llama3 is out. There's a 8b model, 70b model and... 405b which will be released later. How much VRAM would that even be? That's tens of thousands of dollars of GPUs just to load it. I guess the other choice is the significantly slower RAM, but you'd still need like, what, 300-500GB of RAM? (maybe there's been efficient gains since I last looked). 8b is better than nothing, but I liked having 13b as an option as it snugly fit into 12GB VRAM and would be, uh, 5b better than the 8b model. But, it seems like that's falling to the wayside. 70b is 48GB of VRAM, so two 3090s would be the cheapest way to load it into speedy VRAM.
Currently the context is 8k, which would be great a year ago, but the best GPT4 context is 128k and Claude3 is 200k. Looks like they'll be making higher context versions in the future. Context is extremely important for (E)RP since you want to have a backstory and example chat and commands and lorebooks and of course you want it to know what was said 5 messages ago.

Well, time will tell as to how the quality is as I don't really trust anything I'm hearing on release day.

wait wtf when did sankaku start doing this

>>830
Can't say I'm too surprised as he's always chasing money-making schemes. This stuff will slowly become more common, but most people still utterly fail to see the potential in it including here, much to my consternation. The time to strike it rich was last year (which I think I even mentioned in this thread in trying to get people interested in making a kissu version) but it will be interesting to see what others do... poorly. All these website versions are terrible as it's obvious the creators aren't even interested in it themselves. That lack of interest dooms them to mediocrity, but since people don't know any better they won't care.
I can sense some of the gas leaking from the AI bubble, but there's still a lot of it in there.

So there's a new version of GPT out, GPT-4O, and it's insanely fast compared to the other models. It's doing translations in real time and can output text at speeds which make even the super fast GPT-4 seem slow.

Also it's apparently got even better vision now so it can describe emotion in a picture too and the feeling around it more than just describing the factual descriptors of the picture. Not sure if that's disturbing or not...

>>832
Hmm. Yeah, this seems pretty impressive based on this image I just grabbed from /g/. This is pretty damn impressive. I wonder if it can identify birds and stuff. I imagine it's still the case that GPT4 is more dry when it comes to (E)RP stuff when compared to Claude, but the tech and following directions is still the best. I've kind of temporarily lost interest in text AI again since I spent weeks making a character and burned out, but I'll get back into it eventually.
Apparently the GPT-40 thing is free, but you do need to give them your phone number to make an account so anything you say will be tied to your identity. I suppose this means they want free labor in training filters or more training data, or just plain ol' data harvesting.

>>833
>it's still the case that GPT4 is more dry when it comes to (E)RP
Maybe it's just rumors, but from posts here and there I've heard that the new model is incredibly horny. So it might just be up there with Claude now. If it's free with a phone number then I'm sure that the proxies will have it working soon enough.

Spent all day writing saucy stories with AI... Again...

Think it's time I take my yearly bi-yearly hiatus from it, I don't think the models are getting any better but access to the top tiers are near unlimited now and they're really good for erotica crafting. And for me this is a real problem not because the AI is so good I just keep coming back to it all day, but because I'm able to theoretically write an unlimited amount of one mostly cohesive story and I'll spend the entire day writing prompts and regenerating responses until the story fits exactly the flow and theme I'm going for and I can end it on a satisfying conclusion. Sometimes don't even finish joing before the end because I spend too much time typing. There's really so many ways you can extend a scenario and GPT-4 1106 is easily the best model for resisting commands naturally unlike Claude OPoop that will just go along with whatever you say and is so boring to toy with. I don't feel like I've really accomplished anything until I've gone through like 50 messages setting up a story, and then like 150 messages breaking down a character without explicitly telling the AI to do so.

>>835
Also my sacrum fucking HURTS

>>835
>I don't think the models are getting any better
Last I heard, their quality is still directly proportional to the amount of calculating power you give them.

>>837
Sort of, but once you've got the ability to use as many tokens as you want the specific quirks of the models become much more pronounced.

From the filenames in this thread, I am guessing that whoever asked for a cftf in the /vg/ thread came from here. I have no idea whether that person saw my response, but I will post the card I made here in the event that it was not seen in the thread it was requested in.

Please! Bear in mind that I have never seen this show, and have no idea who this girl is. I am going entirely off of what I saw in the clip that was posted alongside the request. The card is entirely SFW.

https://files.catbox.moe/alzinr.png

>>837
>>838
Breaking my vow once more today, I decided to test out to see if I was right about Opus or if I was just being stubborn about not using it from past experiences with Claude.

It's definitely true that, unlike GPT4, Opus uses much better prose and has more flowery/creative language. Which could help a lot if you're into more vanilla or straightforward roleplay with a character or something. But when it comes to how easy it is to break or resistance it puts up against you in a story it's near nonexistent unless you heavily intervene. It'd switch characters from arrogant pure maidens to the most cock hungry sluts in a matter of a single message, not caring at all for gradual progression or having any semblance of restraint if I indicated I wanted to head in that direction. I had to clip off the end to a majority of the chat messages because of this and also modify the contents too to be more in-line with what I wanted. Not to mention how many times I needed to regenerate to get a certain scenario written properly as opposed to GPT-4 where it seemed to follow my intentions a bit better. Far more times I needed to use [brackets for OOC commands to the ai] just to get it to generate what should've been an obvious development given the context and it just frustrated me. At the very least I guess if I ever want to look back on the crafted story it'll look really nice and clean after all the effort I went through to perfect it.

Shouldn't this be moved to /maho/?

hmm I guess I'll do that later, sure

Kissu should try to make their own card! I think we could do a collaboration character thing, anyone else wanna try?

I haven't been paying attention, but I'll post some news...

There's a Claude Sonnet 3.5 which "outperforms opus at 2x speed and 1/5th cost" which means it might be more readily available on reverse proxy stuff. Or if you're someone that buys directly from amazon then it would be, well, 5x cheaper. I don't feel like reading comparisons between it and GPT4o, if they exist, so you'd have to go searching elsewhere for that.
But, more interestingly, fac- er meta released llama 3.1 which includes a 405b version! If you don't remember, the number is the amount of, uhh, parameters or something and it's roughly equivalent to functionality. A 14b model can certainly outperform an inferior 70b model, but it's an uphill battle and doesn't usually happen until months later with a new training process and so on. In general you can take the larger number to mean it's "smarter".
But, uh, 405B... yeah, despite being something you could download you're not going to be using it.
That's like, uhhh... over 200GB of VRAM. Who wants to build a data center for kissu's local text generation?

>>1319
Didn't see this post...
Are you still interested? What kind of card do you mean? What AI service do you want it to be aimed at using?

>>1499
>What kind of card do you mean?
Like a roleplaying character card, like being able to talk to Kuon!
Or otaku-chan!
>What AI service do you want it to be aimed at using?
I don't have any proxies avaliable so I don't really know, I use GPT4o the most but its a bit shite.

>>1500
Hmm, making a character could be good, but my personal opinion is that I will make no attempt to bring AI Kuon to life until there's a sufficiently advanced local model that I could actually host. To get accustomed to talking to an AI Kuon companion and then having that suddenly taken away from me would be terrible.
Anyway...
Pre-existing characters are generally a lot easier to "create" since you're not actually creating them, but they're also a bit less entertaining since you're just filling in boxes instead of writing something new. Depending on the model some of them already have a lot of data for an existing character, too, or at leas the setting. I'm not sure how easy it would be to take an existing character and just build upon it, though.

>>1498
3.5 is can be only activated on accounts with enterprise support like opus, so it's rare. Beside from what I tested when I still had access, it kinda sucks for RP. Way too repetitive and likes to loop.

Do you fucking niggers have any proxies or what the hell do you fucking faggots use for chatbots?

So no one actually uses it, its all smoke and fucking mirrors, huh. Got it. It's all fucking bullshit. All a big fucking scam. Hate ni/g/gers like you wouldn't believe.

AI chatbots aren't real. Be yourself. Stop using them.

>>1508
>>1509
>>1510
Yeah, I have proxies. No, I'm not sharing them with you.

>>1506
Really? I must've been using a good jailbreak preset, https://rentry.org/pixibots#claude-3-prompt , since my recent experience with it was one of the best I've had. The characters were fairly resistant according to instructions and it kept the story varied as opposed to repeating itself like the other models tend to. Also unlike other models it didn't seem to trip itself up with consistency as often.

>>1511
Take as many pills as possible and slit your wrists you useless eunuch.

>>1501
I don't want to make shit if it means niggers like >>1511 can use it. Go make your own card.

>>1515
It's ok, the AI generated cards would probably fare just as well.

>>1516
What the fuck are you even saying?

>>1514
>eunuch
but he has all the sex

Please learn to recognize when someone is joking around and don't derail the thread with rage posts. The situation could be that this reverse proxy stuff is very rare now or something, I haven't actually looked at it.
It's inevitable that the buffet of free reverse proxy stuff will end at some point and it's possible that the day has already come. You would need to *shudder* go to /g/ and search around.

>>1506
Oh, really? Huh.
>Way too repetitive and likes to loop.
Well, it's kind of been that way for me forever. 3.5 Sonnet does it earlier or more often? That sucks. I'd like to try that Llama 3.1 405b sometime. Obviously I wouldn't be able to host it myself, but some services eventually grab them and I could see some AI sites make use of that instead of GPT/Claude. It's annoying to change all the jailbreaks and prompts around, though.

>>1525
>>1506
I ran into the 'sorta' looping problem last night. But I have to say that 3.5 sonnet runs circles around every other model to date on consistency and detail. I don't need to wrangle it as much anymore to get out what I want to from a specific story beat and instead it will take the logical followup route.

The "looping" I encountered was more akin to a status description of sorts in that while it would repeat the same thing again and again, it'd also update that repeated part with new context which is far different from what models before have done. In my case I actually found this more useful than before since it's more familiar to me as an eroge thing than having to force it to do a recheck on the character status every once in a while to make sure it's doing what I want it to.

>>1605
That's good to hear. The looping is something I had tried to prevent with randomized jailbreaks, but it had limited success. The AI repeating phrases and going back in time while ignoring the current story are things that made this largely unusable to me for longer sessions.

I wonder if any of this stuff can be plugged into an OCR program yet in a semi-automated way. I'd like to try some untranslated VN stuff with it, but the software was all made for stuff like google translate.
I bet I'd want to make a system prompt/character to aim the prompt in a good way, too, so you'd want integration with SillyTavern and I imagine that's a very low priority if the two groups have ever even met before. (AI RP software people and untranslated VN readers)

>>1616
To solve that problem I just started learning Japanese. Been pretty successful so far.

>>1617
Yeah, if I had a time machine and money and a normal brain I'd go back to the early 00s and enter college to learn Japanese since it was the only way to do it back then. This AI stuff works in my current situation, though. I put up with some really horrific MTL stuff back in the early 10s so my tolerance is pretty high. Not ideal, but good enough for my purposes I think.
Need to think of prompt stuff.. man, I haven't looked into this stuff in a long time. I wonder how the 'logic' prompts are working these days, like how you would tell it to "think step by step" and that would actually improve the results last year.

Hackers are are using AWS credits for dark roleplaying to have dangerous interactions with chatbots, a small percentage of them raping* actual text strings! Think of the children!
The funny thing is their surprise at how little Amazon cares. You know, a company that generates 5 billion dollars every second issuing automated refunds to what amounts to pennies for them.
https://permiso.io/blog/exploiting-hosted-models
https://krebsonsecurity.com/2024/10/a-single-cloud-compromise-can-feed-an-army-of-ai-sex-bots/

*I think for every male user raping a chatbot there's 1000 instances of a female user getting raped by them

Been thinking about LLMs lately. It's really quite a challenge to sculpt a consistent character out that responds in a predictable and desirable way. LLMs are more or less a gigantic distilled compendium of all human knowledge (if flawed in presentation), which makes having a character that should intentionally be ignorant of some information extremely challenging. Hard to suspend your disbelief when you can ask a Pokemon what a Bose-Einstein condensate is and even with explicit directives that they shouldn't communicate, you'll still get at the very least a high level overview of whatever concept you could imagine asking. As a result, how do you craft a character who has eyes fresh to the world and who doesn't know what every word or concept means when in the machinery driving them is information on every topic known to man? Better yet, how do you make them include misspellings when their dataset is purged of that sort of thing? It's impossible. That character, regardless of the background information you tell it to abide by will still be able to produce the works of William Shakespeare, tell you about the Three Kingdoms Period of China, and also help you with code completion because that's all baked in whether you like it or not.

It's very easy to see the illusory nature of LLMs because they're a top-down solution rather than a bottom-up solution. They are not human intelligent because they are implicitly not a blank slate. They cannot learn. Training with every written piece of information in existence is not learning.

Hopefully at some point we'll get actual artificial intelligence that genuinely is a blank slate, and can learn.

>>1912
Yeah, there's some pretty major flaws with LLMs when it comes to the roleplaying side of things. The big ones like Claude and ChatGPT are a lot better these days, though. It's still tremendously difficult to get it to do things that are purposely wrong, which is something I tried with the Kissu April Fool's Day bots this year. Something like "When she thinks of an anime character, she never attributes it to the correct source material" or something. It runs into patterns and it really doesn't do what you want. This is one of the things that makes ChatGPT3 more fun than the modern ones, it was very proud and confident when you lied to it and it would play along. You could ask it to create delicious recipes with grenades list the strongest Pokemon in the Lord of the Rings. Maybe extensive, fancy prompts could fix it, though. Things like sillytavern and the various ways they can inject prompts like keywords and lorebooks and things make it superior in functionality to how these things work on the actual sites.
Local models? Well... it's a rich man's domain. You need hundreds of gigs of VRAM to host the big models. But local models with LORAs/finetunes is how you'd get what you want. As long as it's reliant on massive VRAM, though, you're not going to actually get the hobbyists that create things. It's hard for people to be excited over the scraps.

>>1913
>Maybe extensive, fancy prompts could fix it, though.
I'm not so sure. They can definitely help shape what responses look like, but they can't really do much about the information they already have. Having a super long character description only works to an extent. At some point, you will have too much information, and some of it will be lost. The only work around I've found is repeating particular things. Lorebooks are also incredibly powerful because you can provide example discussions so it knows how to respond to the current pormpt, but they're limited by keyword triggers. A problem I'm having is both that my character description is too long, so it's not really paying attention to the whole thing, but then when it comes to using a lorebook, the concept I want to affect has way too many keywords associated with it for me to think of them all, and for ChatGPT to actually pick up that the lorebook entry is about that thing.

I'm sure it's technically possible with enough targetted lorebook entries to get a perfected type of output, but for as many concepts as there are out there... there would also be an infinite number of entries you'd have to write to get things just right... And even if you did undertaken the herculean task of getting example discussions for every concept, you might run into the problem of the LLM just restating the example discussion with little variation instead of creating a new response that is inspired by the example text!

It's all so tiresome to try and manage effectively and is really what's disillusioning me about LLMs. I sincerely hope there's some furry freak out there who's got a PhD in computer science and neurology who's similarly disillusioned that LLMs kind of suck for this sort and is working on creating an actual AI.

>>1912
My thought for that was, given the character has enough lines in all of the works that include them, you could train an LLM with just their conversational data. That should limit the amount of knowledge they have.

>>1914
Honestly I think it's just a prompting deal. It comes down to what is being fed to the AI, but the trick is that you need to have full control of what is being sent and load that with the kind of "contradictory" information that overpowers the common knowledge of the model. That can be difficult if you didn't make the bot and if the service isn't sneakily injecting context on top like what most GPT services do, and having a finetuning that's good and conducive to what you're trying to do also makes a big difference. So in this example I prompted a mirror that answers questions, the first is a bumbling idiot, and the second/third is a mime (had to do a second attempt where I specified no writing), and last is one that can answer the question correctly. All total I was feeding some 200 tokens into the AI.