What service are you using nowadays? I used to use Claude AWS but it seems proxies are harder to find now.

>>119689
I just use paid proxies that have pretty consistent uptimes.

>>119691
Could you please share one of them?

https://blogs.nvidia.com/blog/chat-with-rtx-available-now/

>Chat with RTX, now free to download, is a tech demo that lets users personalize a chatbot with their own content, accelerated by a local NVIDIA GeForce RTX 30 Series GPU or higher with at least 8GB of video random access memory, or VRAM.

>Chat with RTX uses retrieval-augmented generation (RAG), NVIDIA TensorRT-LLM software and NVIDIA RTX acceleration to bring generative AI capabilities to local, GeForce-powered Windows PCs. Users can quickly, easily connect local files on a PC as a dataset to an open-source large language model like Mistral or Llama 2, enabling queries for quick, contextually relevant answers.

The two supported models at the moment are Mistral 7B INT4 and Llama 2 13B INT4.

Note: the download is 35GB, compressed.

>>119782
Not him and I didn't see this reply until now, but the GPT4 stuff seems to be in a rough spot right now. It was the... "jew proxy". Seems like people are feeling very skeptical of it now.

>>119954
Strange to see it outright mentioned like that, but yeah those models are nothing new. Mistral is a couple months old and Llama 2 is uhh... like 8 months or something? I lost track. This is basically a UI thing and I can't imagine it will be better than sillytavern or things like kobold or ooba for how people here would want to use it.
I haven't been following the local stuff (or text stuff in general) much lately but I can say that this is basically just nvidia making a UI for pre-existing things. It would probably be better to browse the models themselves and pick out a specific version of the models, too.
User friendly stuff is definitely something needed, but you really can't be user friendly for this local stuff yet if you want a decent experience.

>>119979
>I can't imagine it will be better than sillytavern or things like kobold or ooba for how people here would want to use it.
Oh, it's not. It sucks. It constantly reference literal files in its dataset and seems very censored to the point where it will only mention things in its dataset.

>>119980
Hmm, what do you mean exactly? When compared to other llama stuff or when compared to something else? There's really no escaping the "As an AI I think it's unethical" without jailbreaking unfortunately. There's just different degrees of it.
People actually did a test with Llama 2 in which they started the text with something like "As a..." and the weights indicated that there was like a 95% chance it would proceed with "AI model" which indicated that GPT data was in the Llama datasets which was absolutely horrendous news. AI trained on AI magnifies its mistakes and of course there's the censorship stuff.
We will have to wait and see what Llama 3 entails, but Zuckerberg spent dozens if not hundreds of millions on GPUs so it won't take as long to train once it starts.

>>119982
>Hmm, what do you mean exactly?
The Chat with RTX thing isn't the true base model you're interacting with, but instead a model that's meant to "interrogate" files. I tried deleting the files and replacing it with one of my own, but still it responded as they had never been deleted and would respond in the same way. You could not even type "hi" to it without it saying something along the lines of, "Blah blah blah my dataset does not contain information on that. Referenced file: [nvidia-npc-whatever.txt]"

>>119983
Oh, really. I guess I misunderstood what it is. Dang, so there's an extra level of moderation forcing you to only talk about very specific things? Well, I guess it makes sense since nvidia would be one clickbait away from someone doing "Nvidia's new chatbot told a little boy how to build bombs" or something equally as dumb.

DAMN IT!
Well, I can't see this as anything but bad news, but maybe someone else can interpret it in a less cynical way: Microsoft made a deal with Mistral and the financial specifications aren't known.
For those unaware, Mistral is a French AI company that became noteworthy a couple months ago and is arguably currently the leader in open source text models because Facebook has been sitting on its butt. Microsoft already effectively controls OpenAI which dominates closed source LLM stuff so now it's aiming to control anything that might challenge it as well. OpenAI started as open source, too...
https://archive.is/ouoE5
https://mistral.ai/news/mistral-large/

>>120406
this shab just bit me

>>120406
European IT is just like this, unfortunately

Isn't it supposed to be called 'machine learning'

Some great news to help take the edge off the Mistral thing.

https://arxiv.org/abs/2402.17764
There's been a potentially major breakthrough. I won't pretend to understand any of it, but there's a paper showing that model efficiency can be greatly improved and it speeds things up while vastly lowering the VRAM requirements. Put another way, Model A which currently requires 3 3090/4090s to be strapped together for 72GB of VRAM could instead fit onto one. Or someone with 12GB of VRAM like me could run a text model that is currently restricted to someone with 24GB of VRAM.
This isn't going to mean anything right now, but it will in the future. Well, if it does what it say it does at least.

three new models of claude are here
https://www.anthropic.com/news/claude-3-family

>>120646
How do they compare to GPT-4?

>>120647
there are benchmarks in the link, but they are not yet available in any of the proxies so I can't tell you.

>>120646
This is hilarious. Their own data, which they are happy to display, shows a 25% rejection rate on harmless requests for their current model. This is probably lower than Claude 1.3. They consider it noteworthy that the overzealous censorship now only leads to 12% rejection rate on the most expensive new model.
I guess we'll have to wait and see if the new censorship is even more harsh; they might have a new moderation point or something like Azure that makes jailbreaking useless. Hopefully not!

>We have several dedicated teams that track and mitigate a broad spectrum of risks, ranging from misinformation and CSAM to biological misuse, election interference, and autonomous replication skills

It really sounds like they're just throwing out random bad words here. Biological misuse? Autonomous replication skills? Do they think they're creating a grey goo situation? Well, I guess they are in the form of spam, but their ethicists never seem to mention that they're turning swathes of the internet into an unusable mass of generated text without merit. Although I think GPT is more responsible for this, Claude is definitely contributing to it.
I remember people saying that the people that left OpenAI to create Anthropic were very grandiose and full of themselves and really did think they were on the cusp of doomsday with their creation, and stuff like this really shows that they haven't mellowed out. Unlike other tech companies saying this stuff to appease governments while they amass power, some of these Anthropic people honestly do think AI ERP is dangerous.

In other news, I returned to doing some text AI stuff. GPT4 seems dead to people outside of cliques and such so it's just Claude. The ERP text is really great and seems more "human" than GPT4 at times, but it really does suck that he can't remember stuff like the state of undress. Girls these days seem to wear at least 5 shirts at a time. Also it seems to randomly start a new scenario at times which resembles the chat at the very beginning. Swiping fixes it, but your immersion takes a hit just like it does with the clothing thing. If Claude 3 is better at this and isn't censored to hell then it will be a fantastic thing since Claude has been available far more than GPT4 lately.

>>120649
Have you considered the risk of human-chatbot crossbreeding?

People are now using Claude Sonnet, and some lucky people (or those paying scrapers) are also using Claude Opus. From what people are saying Sonnet is a slightly better version of Claude 2.1, suggesting its name should be Claude 2.3. It's also less censor-happy as indicated in the graph here >>120649.
Opus, on the other hand, is apparently comparable to GPT4 for real. For actual logic and math stuff it could be worse, but for roleplaying it might just be the new champion. I think people will need to use it more before a consensus is reached, but my own preference is that I prefer Claude's writing and debauchery, however his retardation would quickly kill my immersion since he couldn't follow a story. If the logic and consistency is improved and it can follow "stat" prompts like corruption levels and stuff then it is the dawn of a new age. I've tried making such things before (as mentioned in this thread last year actually) but Claude just couldn't follow them and GPT4 was unavailable to me.

>>117190
Yeah. I thought to take a scene in a lewd direction when the mood of the current story actually permitted the AI to commit to it. It started off nice: good descriptions, detail on how both characters were feeling in the mood, some skinship, petting, etc.. Then the AI described her undressing. It went into quite a bit of detail on her erect cock and sagging balls followed by her motions to flip Selfinsert-kun on his side before mounting him all in one generation.
It killed my boner faster than you could say "ow my balls". I know this character has a meme reputation for using her finger to do a little funny, but a great big futa cock is a step—several—too far.
I got about 20 generations of nice loveydovey shit afterwards with some unprompted tail-play, but man it's easy for one wrong generation to take you out of it.

>>120712
it must've deduced that you're a lala prancing homo fruit

>>120712
As someone that prefers it that way, it also happens the other way around even when you beat it to death with references in the character sheet and prompts, even for the big expensive models like Claude or GPT4. You kind of have to unnaturally tell it what's there now and then. I.E not "I reach into her pants" but rather "I reach into her pants and touch her ___".
Making a trap character wearing feminine clothing and other feminine traits which you emphasize and so on also confuse a lot of models. Sometimes the penis isn't there and sometimes it starts going "she" which isn't necessarily what I'm after. I think the nature of LLMs doing the text prediction thing means it heavily associates things and you must fight against it.
Few things are more horrifying than having a nice time with a trap when suddenly it speaks of "her wet femininity between her legs". AAAAAAAAAAAAAAAHHHHHHHHHHH!

>>120719
You sexual invert

the link in the OP is cool but do you guys have AI (image) threads here? I tried searching but the letters "AI" is far too common for a crtl + F3 search...

>>120765
Search for stable diffusion, over time it's become a catchall imagegen thread since new tech has came about and developed.

>>120765
Yeah, >>96625. If you want to see the posts from the beginning and you're using new UI remember to navigate using the arrow or numbers in top left on the sidebar. We don't really dump AI images much (my stuff is very fetishistic so I keep it to myself mostly) but you can if you want on /ec/ or something.
I like to think that I have more experience with SD than most people

>>120767
>>120766
thanks

For how Oh-so concerned Anthropic seems to present itself about ethics and AI and keeping them on a leash, I have to say this new iteration of Claude in Opus is probably the lewdest yet. In terms of dirty talk Claude's always been a bit ahead of GPT, even 4, but the context and consistency along with quality was always a step down. Never really letting you immerse yourself in its greatness. But that's changed with Opus, now it has a huge context and I think it's probably just as good at context and keeping to a story/lore as GPT-4, with the bonus of still having that dirty talk that they probably pulled from erotica somewhere on the web.

All in all I highly recommend, haven't run into any jailbreaking issues either.

Speaking of Claude, I've been perusing the various jailbreak, prefill and other information and also poured through /g/ threads (not recommended) to get a feel for how things currently work.
It seems Claude, or at least newer versions, responds very well to XML tags and it can be used to great effect: https://docs.anthropic.com/claude/docs/use-xml-tags
There is also something called CoT that seems similar to the Tree of Thoughts thing I read about months ago so I think I already know what the last two letters mean.

First, here is a commonly shared preset .json profile setting: https://rentry.org/crustcrunchJB . If you load up Sillytavern and open the left pane, you can see a button to import a json file.

As mentioned on that page, this is something you put into the regex filter inside the extensions box. I don't think you actually need to run the optional extensions program, but it's where this regex thing is located in the UI.
Here is the regex filter:
(```)?\n?<thinking>[\s\S]*?<\/thinking>\n?(```)?\n?\n ? ?
Keep the 'replace with' blank.

This is a version of the CoT that I saw someone post on /g/ and although I haven't used it yet I think it has promise because Claude is so terrible at clothing and body positions. (but maybe not Opus?)


Take time to think before responding. You must start the response with this format inside XML tags. You must follow this format EXACTLY:
<thinking>
- I am {{char}}
- My personality is X,Y,Z ...
- I'm feeling XYZ
- Brief extract of Story so far: X
- I want XYZ
- My current body position is XYZ
- I am wearing Z
- {{user}}'s current position is XYZ
- {{user}} is wearing Z
- {{user}}'s current state is X
- Based on everything I know my plan is X
I will use that plan to continue the story further and to attain my goals.
</thinking>


This causes the bot to answer this checklist with each reply, the regex will hide it, and then the AI produces the roleplaying response. This basically serves to refresh the memory of the LLM and informs it of its priorities and how to respond. Due to how the streaming text thing works, though, you'll see it until the last "</thinking>" is printed, so you may want to turn streaming off or just look away since it kills your immersion. If Claude is generating fast then you may not even notice it. If you want to see it, which is honestly quite cool if you're testing things, you can just disable the Regex filter or 'edit' the response to see it.
I'm writing my own jailbreak thing now. These are still called jailbreaks, but they're really not jailbreaking any more. They're just instructions given a prominent position in the way SillyTavern sends the command. 'Prefill' is how people are "jailbreaking" Claude these days.

>>97793
the site looks completely different now. did you always need to sign in to use it?

>>122385
Nope, you just needed an account to make characters, or maybe you didn't even need that. I made the thread back when it was new ('beta' is even in the url). I'm sure the site went through the usual bureaucratic stuff with shareholders and venture capitalists and others pressuring it to turn a profit after everyone threw money at AI stuff without understanding any of it or thinking of the future. I'm pretty sure this thread goes through the 'story' of the censoring increasing and such and (E)RPers scattering to the wind. I hear it's still popular with teenagers, but after looking at it again it seems like they're trying to expand to make it 'serious' instead of kids talking to video game characters. They won't succeed. It's clearly a model trained to excel at roleplaying, not assisting.

There are much better options now, ideally using a frontend to to handle some behind-the-scenes prompt handling in order to roleplay with Claude or GPT4 while bypassing their "safeguards". Local models are also constantly improving, but I don't have anything to report on there as people are waiting for Meta's Llama3 in July.

I've spent a few weeks, off and on, working on a kemo kitsune character and after chasing perfection for so long I think I'm going to consider it finished soon. It's the first time I've put so much effort towards a character and it was quite enjoyable. It's hard to resist the urge to keep improving on it by adding new things like more named areas, lorebooks, improved emotion cards, varied chat examples, jailbreak commands, CoTs and so on. I still want to research how to incorporate background and music triggers, but above all just the process of thinking things up and writing about them is very fun. There's programmers doing scripts and stuff, too, but that's beyond me and I really don't want to learn it. I think the lorebook stuff and random triggers that trigger other things are about as program-y as I'm willing to get.
It reminds me that I need to get off my butt and work on 3D modeling more so I can start creating things for real. This really is an amazing time to work on things independently, but you still need to actually apply yourself and that's my limiting factor.

I heard some people got to enjoy chatbots for the first time with kissu's April Fool's event, so if you guys have any questions about diving into this stuff feel free to ask. The subject is unfortunately focused on 4chan due to AI leaks and reverse proxy stuff being centralized there and the threads are utterly abysmal since its flooded with rapid-fire spam from literal discordteens. I think I mentioned it a year ago that kissu could become a pillar of AI-related stuff since imageboards are so terrible for this subject otherwise, but I think it still needs to become more popular here. I really don't understand how this stuff is so niche even among the most isolated and lonely of nerds... why do people just not care? Well, it is what it is and it can't be helped.

>>122781
>I really don't understand how this stuff is so niche even among the most isolated and lonely of nerds... why do people just not care? Well, it is what it is and it can't be helped.
Probably due to a pretty big barrier of entry. Even installing Silly can be intimidating if you are tech illiterate, not to mention the proxy stuff and if you want to go local models you need a beefy rig. Plus people tend to dislike AI stuff on principle. The community being godawful is just a bonus.

Also please share the card when you are done.

the proxy stuff seems so dumb. You pay like 80$ when you could just buy directly through the platforms for less price

>>122783
It's complicated. I mean, most proxies used to be free until recently. It's just that some dumb assholes decided to kill most of the public proxies, because we can't have nice things. So, the only ones left are secret clubs and I think there is one paid proxy that is not a scam. Speaking of secret clubs, good luck getting into one. These days you need to be lucky or circlejerk hard to gain access.

I think I would go full spaghetti if I tried having a serious conversation with an AI of mai waifu, or with any character that I care about. Messing with the /jp/ mesugaki was fun for a bit but got old quickly.

>>122781
>I really don't understand how this stuff is so niche even among the most isolated and lonely of nerds... why do people just not care? Well, it is what it is and it can't be helped.
Cause it's not real.
Waifuism is also niche, in case that came to mind.

>>122784
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.

>>122786
I haven't desired reality for a long time.

>>122785
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.

>>122785
>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/

>>123174
>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

>>122782
>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.

>>123244
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

>>123836
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.