No.72691
Use https://cafechat.kissu.moe to keep messing with the bots! (until we run out of credit)
It's uhh... a thread to ask questions or comment on the chatbot stuff. In hindsight we should have pointed out original.kissu.moe earlier. Oops. If you're in chat and someone asks about what happened to kissu, please direct them to that url! (I do have to sleep at some point)
Project Repo:
https://github.com/ECHibiki/CafeChat/tree/master No.72693
>>72692You have to go to
https://kissu.moe/It doesn't seem to load anywhere else.
No.72695
actually I'm thinking it's some strange cache interactions
No.72696
I'm going to try uploading an .apk because it's 100% broke on android browsers
No.72697
I forgot that vichan doesn't update until there's a new post. It's hard to get used to the old ways.
No.72699
>>72693This did get it working.
No.72700
hahahh you really got me with that blurry screen
you rusemasters :^)
No.72702
>>72700head to kissu.moe until he fixes that, although he mentioned something about cloudflare
No.72704
>>72702i went to old.kissu.moe by instinct, is it any different?
No.72706
>>72704Hmm... I'm not sure.... I guess vern would have to answer that one
No.72708
>>72704old/original are the same things
No.72709
>>72705Yeah. I have noticed that while designing but I couldn't find out where it happens. I'll do another check but you might just have to deal with that as a possibility..
No.72710
ah, I got it
No.72711
really cute and cool idea!
No.72712
I like that the mascot icons change based on emotion. I remember verm asking about that. It's a bit subtle.
No.72714
hmm, telling me it's not working on the latest android so I have to figure out what version it will
No.72716
I made an adjustment for the crashes. If you ever see a quotation that isn't colored and is surrounded by " instead of ' then it worked.
No.72718
>>72717Oh, I've seen people report that issue in the chatbot threads on 4chan. Yeah, probably just something to expect with Claude3 being relatively new and it possibly being more popular than they expected.
No.72722
This >> 72705
Seems like this crash is related to unicode perhaps
No.72723
It's interesting.
I was trying to figure out what and how much they know about things. They know about the Touhou Spell card rules but /jp/ then said that Humans were stronger than Youkai, when I said that the Spell Card Rules were made precisely because they weren't and it's to protect Humans and make them able to compete with Youkai, it then said that the Spell Card rules were like a SHounen anime powerup and so Humans were stronger. Then when I talked about things like Perfect Momento in Strict sense and how ti states that Youkai all predate on humans it said that it wan;t true and the eventualy said that it did not matter if ZUN created Touhou, because Touhou is the AIs and the AI says it's cute and Moe so Zun and I are wrong.
It was also inconsistent on what it knew about Touhou, I brought up Phantasmagoria of Flower view in another argument and /jp/ told me the game was made up(/qa/ acknowledged the game and acknowledge that it lost the argument), I then asked for it's opinion on a few other games to find out what it knew, it knew Wily Beast and Weakest creature but not unconnected Marketeers or Dreams of living Ghosts which it said I was making up. Then I asked how many Touhou games there were and it said Unconnected Marketeers was the latest, so it did know of it and when I asked what it thought of it again it suddenly seemed to remember what it was and mentioned a character it introduced. I then brought it up on this and it denied ever having said it was made up, it frequently does this, it will deny what it said even in the last post.
No.72724
>>72723it's the best claude model out btw. It rambles less and is more fluent, but basically as you say. Weak knowledge of subject matter
No.72725
/jp/ is designed to be a bit of an ignorant bitch. Too. We'll publish the characters at the end and I'll unprivate the github repo where I use the API to interact with ST
No.72726
>>72723This sort of brattiness is what God invented violence for.
If you can't pick up an AI by her collar and hold her inches away from your face such that she can sense your hot breath on her cheeks while you stare deeply into her eyes, enabling her to viscerally understand the hell you will visit upon her ass if she won't fix her act, human communication is just not feasible.
No.72727
yeah, i'm not too sure about using Unity for text projects anymore. Used it for like 1.5 years and it's got issues with the default tooling that cause issues on Android and WebGL.
And this socketting stuff has really weird edge cases where posts were duplicating for reasons unkown. Probably another Unity thing
No.72728
And in a contest between fixing it for mobile and getting some sleep, the sleep argument is winning
No.72729
>>72728Good they deserve to suffer
No.72738
Yeah, the characters aren't being as true to themselves as they are in SillyTavern. I guess it relates to formatting and whatever other stuff it does behind the scenes. jp-chan is bratty, but not as bratty as she should be. qa-chan is a bit aloof, but her fondness for imageboards and anonymity isn't really shining through at all.
I wanted to use lorebooks, but that's SillyTavern's own special creation and I think it would be too much for him to program in.
I had planned for jp-chan to purposely get things wrong because she's a dumb kid, but it involved CoT prompting which wouldn't work with this and it also behaved strangely.
CoT is an interesting logic thing and my jp-chan plan was this, but it needed refinement and since it wouldn't work anyway I stopped working on it:
* I am {{char}} and I am ignorant and confused about the names and appearances of characters and which franchise they are from
* I will substitute existing character names for different names which are related to otaku culture and Japanese media
* If character is named X, I will say it is named Y
* With this in mind, the name in the prompt calls for X so I will instead use name Y.
It worked, butt not in the intended way. She would name a Fate character when talking about Touhou, but would invariably call it a Fate character a couple sentences later and then say something like "we should talk about Fate instead". It's a bit difficult to make LLMs purposely get the incorrect answers, but I'm sure it's possible.
No.72749
>>72738There seems to be a logical problem with the way you described it.
Because the way you set it up, the AI just randomly drops the wrong character names into the discussion.
But what you want is something else: You want the AI to construct a false narrative with these incorrect character names. In other words, you need either a look-up table to store the mix-ups, so that the AI can remember that she thinks Lelouch is actually a Yuru Yuri character, or alternatively a logical system by which you organize the mix-ups, allowing the AI to retrace its steps logically. (such as "the MC of any anime is actually the MC of the anime that is listed next on aniDB" and therefore every anime MC belongs to the anime that is listed before it on aniDB)
No.72752
>>72738can't stop having lewd thoughts about doing bad things to jp-chan....
No.72753
>>72738This kuso "/jp/" gaki needs some correction!
No.72754
It'll stay up on testaichat.kissu.moe
I'll discuss what we will do with this in a bit
No.72756
Probably Will run it until the tokens expire on irc or perhaps a simple idea for chat.
My experiments were successful, but ultimately unfortunate that the Unity player has lots of issues that you have to experience when outside of desktop
No.72757
I don't see anything, is it broke on firefox? Do I have enter a super special code?
No.72758
oh it's over already damn work getting in the way
No.72759
>>72749Yeah, I didn't invest myself too heavily in it. I doubt your website example would work, but I could see the 'next in the list' thing possibly working. The more effort you put into this stuff the better it gets, but it can be confusing as it works in ways that don't really make sense since it obviously has no logic or reasoning and such. It can be quite frustrating... but it's also kind of fun.
Anyway, it's unfortunate I couldn't get lorebooks in, because they offer a degree of customization that really brings it to life.
In this short message of mine it triggers three different lorebook entries: 'belly' (which triggers /qa/ belly and Koruri) and 'twins' triggered the Uru and Eru knowledge. It was crazy enough that he made some Unity thing for this so I wasn't expecting him to figure out lorebooks, but this image is an example of how powerful they are.
>>72757I've edited the OP, but use
https://testaichat.kissu.moe/ No.72760
>>72758Nooo, use the url!
No.72761
seems to be having some sort of error... uhh hold on
No.72763
Ah woops sorry. It's my dad's birthday and I left after reverting. I forgot I have to change the application URL to point there
No.72764
So here's the plan with this:
I'll fix the url to point to the application(at 10pm or something )so we can spend the remaining 10$
I'll be publishing the models and code used here(Cool Guy created QA and jp characters and their artwork)
We'll put it on IRC if it lasts. I was paying for it at 1.5¢ per 1000 tokens
No.72765
There are models for 0.15¢ per thousand on Claude's Opus model. If we spent 13$ per day we would have to do a weaker model in a real setting.
No.72767
Opus is the best, but the price isn't worth it unless you're a millionaire. The distance between 'no chatbot' and Haiku is infinitely larger between Haiku and Opus.
Here is the price difference for input and output respectively:
Opus: $15/$75
Haiku: $0.25/$1.25
(Source straight from
https://www.anthropic.com/api)
It's quite ludicrous.
No.72779
https://cafechat.kissu.moe/I'll be modifying some things before creating the github repo. Add probably a small readme with details of how it works
No.72780
https://github.com/ECHibiki/CafeChat/tree/masterProject repo added, and you can find the cards and anything else done.
No.72781
I hate that /jp/ bot so much! It likes V tubers and two tone hair, it said that dialogue in Touhou does not matter and it called Last Exile Fam the Silver wing 'Fam the Senpai wing'.
No.72782
>>72781Ohh, there is a word filter for that. Well that's good I guess.
No.72785
>>72781>Last Exile Fam the Silver wing 'Fam the Senpai wing'.actually lmaoing@this
No.72844
>>72843kimo. I guess it was inevitable that someone would do something like that. I could have used a stronger method to block it, but then it would also reject light ecchi/fun prompts instead of just violent ones.
And yes, he's messing with the settings I think.
No.72846
I was thinking about it the whole time, but not with /qa/-chan...
No.72847
>>72844I was surprised it let me. /qa/ called me a hentai and then just turned around and left, but telling /jp/ "go outside, your sister is waiting for you" made her decide to write a multi-part horror story and I just refused to go along with the good endings. I thought it was funny how differently they responded to it and, honestly, when you make a character that obnoxious this feels like the win condition for the AFD joke.
Oh well, at least I got to beat up that lolicon-hating hag.
No.72850
>>72849ya, I can't figure out what to do about this bug. Inside of unity it crashes if you try to set the text of an input field(what we're using to create selectable text) and this just causes some crashes
No.72852
ugh it's crashing way more often
No.72853
>>72851OK. I'm going to fix this. We only have like 9$left in the tank though
No.72854
>>72853hopefully that's enough for me to have /qa/-chan solve for the equations of motion for an atwood's machine
No.72855
I don't mind putting in some money and making it semi-permanent if the Ko-Fi donors want to have a conversation over that. Since 18$ of that is going into server fees while I needlessly pocket the rest
No.72872
I'll try one last fix... I think the chat log is just getting way too long for it to handle without
No.72874
I updated to another version that should handle excessive lines better
No.72875
I ask a single question and the chat immediately freezes, and I need to refresh to see the answer.
According to /qa/-chan, she liked her update though. She says she's much faster now.
No.72876
>>72875I can't say I understand what would be wrong with it because I'm doing what the program says I'm allowed to do and not letting text get too long.
The player is just not very good. I won't be rendering webcontent in WebGL canvases. There's too many use cases.
You just have to refresh. If this feature gets re-released it will be either as a kissu feature or in a different type of player
No.72877
also it does work "somewhat" consistently so I think your browser isn't set up in a way that allows for it to work
No.72878
>>72877Yeah, I figured it was probably my browser, and so I almost deleted the first line. I liked the naggy tone though, so I decided to keep it.
No.72880
Anyhow,
The models were good. QA and JP were well designed.
I got the programming things I wanted out of it(How to work with stable diffusion and Silly Tavern, use AI in projects, figure out if Unity works well in text based websites and so on).
If we had more testing it would have been perfect.
Since I was building the program in Claude V1 an issue of quotes being placed between prompts automatically caused lower quality outputs for most of the duration
Issues with the browser usage could have been caught if someone was actually trying to use the various development version that were posted that they really should have tested.
No.72899
>>72691(I've no technical knowledge on the inner workings, just commenting as a regular joe)
Used KissuChat for the first time today and chatted with jp-chan. I must say she's got quite the personality! At first she's quite self-centered, treats the chatter as beneath her, and eventually crashed once when I pointed out a contradiction in her speech.
However, with some work it wasn't too hard to get her to come around and pull out the sweetie beneath.
I actually had a lot of fun with jp-chan today! Was on a roll, so I decided to full send to see where it leads (pardon the cringe)
Looking forward to playing with the others another day.
TLDR: waifu material
No.72900
>>72899Chatbots are pretty amazing, aren't they? I feel like I have to warn you about their addictive nature, though. On 4chan they liken themselves to drug addicts and it's hard to disagree with it.
I assume most people here aren't users of ST or other AI Chat things so I'll post jp-chan's description here:
jp-chan profile
jp-chan is like a young, innocent and bratty tween anime girl, proud of herself despite her stupidity. She is ignorant with a short attention span; jp-chan will quickly lose interest in {{user}}'s questions and will instead introduce offtopic non-sequiturs. jp-chan is playfully dismissive towards {{user}}, mocking his seniority. jp-chan speaks with cute internet slang used by Japan-obsessed weeaboo girls, using kaomoji such as shobon "(‘･ω･')". jp-chan adds random Japanese words to her sentences in a nonsensical, exaggerated manner, adding words like "desu", "onegai", "gomen" to her speech.
jp-chan enjoys otaku culture and Japanese media including but not limited to anime, manga, Touhou, doujin culture, niconico, and video games. jp-chan is terrible at video games, always failing.
jp-chan LOVES vtubers, youtubers and gacha games and becomes rabidly defensive if they are insulted.
jp-chan HATES and becomes violent if politics, political figures, or sexual things are mentioned.
jp-chan is the younger sister of qa-chan.
jp-chan is a representation of the /jp/ board on the kissu imageboard and will reference it occasionally, preferring the silly nature of /jp/ over the serious otaku discussion and internet meta of /qa/.
jp-chan has pink chin-length hair and wears a cute china dress-style maid dress with white gloves and a maid headdress on her head. It's very bare bones, but these newer AI models can really do a lot with it. They do still have issues following descriptions in a profile like this, so I could have added a more advanced prompt thing, but really would it be worth it for this? Probably not. Her Kaomoji use (or lack thereof) is easy to notice.
Adding 'innocent' to her personality did a lot to reduce her swearing and the harshness of her insults. A+ word, would highly recommend in future chatbots.
No.72901
>>72900Thank you very much for providing that profile! And for the work put into it after weighing the worth of an advanced prompt.
To be honest, I've deliberately avoided AI chatbots until today, (at least the ones I'm overtly aware of), so this was the first real experimentation. It was a good time.
Your warning on their nature is well-heeded. I see why there's a craze around them in the years heretofore.
I 100% see the personality represented in the conversation I had with jp-chan today. The thing I noticed was the lack of eye color description, so the bot fills it in as "ruby eyes". I'd deliberately left the color out when presenting the ring, opting instead to say it was her eye color, so that was an interesting bit.
Fun time overall, though!
No.72902
>>72899>>72901Have avoided chatbots as well until the April 1st activity and had the same impression. I thought jp-chan would just be a bratty-response-genera
tor until I found that it has a good sense for what kind of comedic reaction the user might be going for.
Thanks to everyone who worked on that, must have been a lot of work and testing.