Welcome to Chatbot AI General #105, the thread for discussing and improving AI pony chatbots. ▶ MLP Bots https://mlpchag.neocities.org Spreadsheet (CAI bots + Old repository): https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJT Bunch of Character Cards (Descriptions/No Greetings): https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5k CAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zip Expression packs: https://rentry.org/ChagExpressions !!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U ▶ How do I start? 1) Select a Frontend 2) Select an AI model 3) Select Jailbreak 4) Select bots 5) Lovemaking with AI mares! Starting in this hobby can be confusing and difficult. If it’s your first time and you’re lost, ▶ ASK THE THREAD! ◀ Novice-to-advanced guide: https://rentry.org/onrms ▶ SillyTavern (preferred frontend) https://github.com/SillyTavern/SillyTavern On Android: https://rentry.org/STAI-Termux App that voices pony responses in ST: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB ▶ More frontends: Risu: https://risuai.xyz Agnai: https://agnai.chat ▶ Locals https://rentry.org/lunarmodelexperiments >>>/g/lmg/ Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main ▶ Jailbreaks MLP JB: https://rentry.org/znon7vxe More JB and guides: https://rentry.org/jb-listing Hypebots for Tavern: https://rentry.org/pn3hb ▶ Botmaking Editors: https://agnai.chat/editor Guides: https://rentry.org/meta_botmaking_list Advanced: https://rentry.org/AdvancedCardWritingTricks ▶ Miscellaneous https://rentry.org/ChagMiscellaneous https://rentry.org/ChagArchive Previous thread: >>41753592 ▶ Current theme: SOVL **News:** - Deepseek releases V3 https://api-docs.deepseek.com/news/news1226 - OpenAI announces o3 models, currently in early evaluation testing - Non-preview OpenAI o1 released on API - Gemini 2.0 Flash released, "more model sizes" in January https://developers.googleblog.com/en/the-next-chapter-of-the-gemini-era-for-developers/
Anchor for bots, lorebooks, scenarios.Previous >>41753595
Anchor for technical stuffs (Proxies, Updates, Models etc.)
Anchor for asking for bots, lorebooks, scenarios etc
>Bots Corner:Trixie: https://mlpchag.neocities.org/view?card=Reta/Trixie.png Pinkie Claus: https://mlpchag.neocities.org/view?card=Anonymous/Pinkie%20Claus.png The Great and Revered Trixie: https://mlpchag.neocities.org/view?card=Reta/The%20Great%20and%20Revered%20Trixie.png >OC Art:>>41760379 >Screenshots Corner: A big jar: >>41753821 A very nice job: >>41759128 HUZZAH: >>41759532 Pet the pretty mare: >>41761352 Cute Trixie: >>41761674 Anon is a monster: >>41761832 Anon plays high-level strategy games: >>41762148 Happy Hearth’s Warming Eve: >>41764837
>Screenshots Corner 2:Knight in Jean: >>41769181 Great and Revered Trixie: >>41767490, >>41769544 Anon the Psycho: >>41768135, >>41768161 Sun Sex: >>41769866 Deepseek logs: >>41771894, >>41773467, >>41773544
>Screenshots Corner 3: Somecreature: >>41771690 Ponydex: >>41774534 Mares win against prompt injection: >>41774725
>Bring THE card you want to make, something you’ve poured your soul into. Won. So, favorite pony, cool scenario, simulation, etc. bring something you’ve put your heart and soul into.If you participate in this, put a burner email in an unimportant field like Character Version. Maybe Santa will come. (It will be deleted before going on the website.)End of the event:01/05/2025Bot:>Pinkie Claushttps://mlpchag.neocities.org/view?card=Anonymous/Pinkie%20Claus.png
>>>/g/103605984
>>41775113 There was one posted here: >>>/g/103661569I also saw logs using only this in the jailbreak: >[Write as {{char}} only; do not speak for or presume {{user}}'s actions or thoughts. All narration is to be written in third person from {{char}}'s perspective in the way {{char}} would say it; carefully study {{char}}'s personality to achieve this. Play with narration and writing style. Be extremely creative in how you choose to interpret {{char}}'s speech and narrative style!] Often with temp set to 1.5-1.65 and top P between 1 or 0.95 if it gets too schizo.
my github account just got banned...
>>41775234 With the new code that’s supposed to help with bans? Mine is still going strong for now. You can just use your same mail with an alias and register again: >>>/g/103626427
>>41775129 Reposting, but here’s a way to run and test local models online for free: https://glhf.chat/landing/home You can also paste in an HF fine-tune link, and it will run it.
Lovemaking with AI mares!
>>>/g/103665361
>Column still dead
deepseek on chary when?
>>41775459
Still a but new to this and have a quick question:I'm using GPT 4o turbo and it feels like the model is driving the narrative too much. Ill write 2 or 3 sentences and it will write 5 paragraphs where it takes control of the convo and writes as the user instead of allowing me to have a one-on-one. Anyone have a solution to this or maybe could suggest a different model?
>>41775755Use deepseek instead. scroll up / to the last thread
>>41775755 First, can you tell us what front-end you’re using and which source you’re using to access the model? That will help us give you instructions for presets or recommend models you can use.>>41775764Anon, that doesn't help and look like mindless shilling.
>>41775769? I have to assume hes using a proxy. I'm telling him to go back a thread and use the proxy / JB listed there instead.
>>41775774 If he’s having problems with the model speaking for him, it’s clearly a preset issue, and he said himself he’s new. Just telling him, "Use that," doesn’t really help. Depending on what he’s using, he might also have access to a lot of models right now, including Chorbo, 3.5, and Deepseek. So it’s best to first understand his situation and then help him or list the different options based on his preferences and what he has access to.
>>41775788The JB includes a preset.
>>41775769Yes, sorry, I should have specified that I'm using SillyTavern and my own OpenAI account API keys. i assume thats what you mean by sourceThe card is the unicorn school sim from the /chag/ neocities>>41775774No proxy is being used.
>>41775797 Before I list the options, do you have anything against: 1) Proxies 2) Trial abuse 3) Paying very little for models other than the ones on your API key? If you have problems with all of those, I’ll just help you with what you can do with your current key.
>>41775801No objections from me.
>>41775797If you can't afford to pay $15 a million tokens for claude sonnet then I recommend deepseek instead, its nearly free and like 90% as good. Knows way more about pony than gpt 4 does and is not nearly as dry.
>>41775808Alright, first, if you’re just starting out, you might not know about presets/jailbreaks. Presets are ways to tell a model how to write, act, and bypass filters for NSFW content on censored models like Claude or GPT. For example, your issue with the model speaking for you is typically a preset issue.Presets are often shared in .json format, and you can find many for different models here: https://rentry.org/jb-listing.You import them like picrel in Tavern, and the best approach is to test a few to see which results you prefer. I’ll give you recommendations depending on what model you want to try.As for access to models:1) Abusing trials for freeRead >>41775180. It might seem hard to set up, but it’s not really, and you can ask for help here if you get stuck. This method gives you access to Claude 3.5, which is a very good model for free. Once set up, you can try the presets here: >>41772380 (You) (Cross-thread) and ask for more infos if you want.2) ProxiesIt’s more complex, but you can follow >>41775764’s advice and look for a Deepseek proxy and preset in previous thread.To understand what a proxy is, read: https://desuarchive.org/mlp/thread/40917404/#40926228.Also, it’s highly recommended to read the novice guide in the OP from beginning to end. It’s a bit time-consuming but very worthwhile, it will help you a lot.3) Paying yourselfThe Deepseek model mentioned is very cheap. You can make an OpenRouter account or a Deepseek account, pay around $2, and have lots and lots of chats with it.If you prefer this option and don’t want to bother with the other two, just let us know, and we’ll help you with presets etc.
>>41775808 As for continuing to use your own OAI API key: Many anons believe that ChatGPT-4o-latest is the best model for RP and storytelling. However, it’s also the most censored if you want to do NSFW. Some presets can help bypass this, like the Knotty preset, which you can find here: https://rentry.org/lattejb. You can also test other presets in that rentry. If you’re mostly doing SFW, the preset posted here is also good: >>41740799. If you want to try other GPT models, you can find plenty of presets at https://rentry.org/jb-listing. Almost all of them will stop the model from speaking for the user. Another option would be to use free Google Gemini API keys. They’re free right now, and you’re not abusing the system. The only downside is that Gemini can be a bit dumber compared to other models, but it can still be fun and effective. Read >>41756306 for instructions. Sorry for the wall of text, but this gives you all the options. Don’t hesitate to ask for more info or help with any of them.
Checking how far I can push the temp on Deepseek before it goes schizo. I think I found the border at. 1.8 temp with some min p
>>41775797Anon above already gave a pretty comprehensive guide so I don't have much to add in terms of advice, but>card is the unicorn school simDo post screenshots of your chats if you get fun ones, I'm always curious how others interact with my cards.
>>41775940lol, I just did a temp test myself and found the sweetspot at 1.8I tested it on custom cyoa card where I roam equestria and results are really great and proactive
>>41775940 I found that 2 can also work with 0.95 Top P without going too schizo, but that might be too much for some cases. But yeah, the meta with deepseek is definitely to go high with the temp.
>>41775940>>41775958>>41775962Here is the same with 1.75 temp instead. Yea, its somewhere around 1.7-1.85 I think.
https://files.catbox.moe/1oybiv.jsonI should post the JB im using here as well.
>>41775901>>41775924Holy shit, anon. That is a fuckton of good info. Going to try some of these out in a bit. Thank you so much!
>>41776151 No problem. Like I said, if you’re stuck or want more info, don’t hesitate to ask. It can feel overwhelming when you’re starting out, but everything starts to click after a while with use. This is definitely subjective, but personally: For a beginner, I’d recommend starting with the free trial method while it’s still available. If you don’t want to bother with that, then go for Deepseek on API or OpenRouter if you don’t mind paying (it’s cheap, when I said $2 for a lot of chatting, I wasn’t exaggerating). Alternatively, Gemini is a good option if you really don’t want to pay. Again, don’t hesitate to ask for preset recommendations depending on which option you choose. Once you’ve had some fun and have a bit of free time, read the guide that was recommended earlier from start to end.
>>41775764deepseek does it too, ive had it shit out like 8 paragraphs at a time. its a pretty good model, but im beginning to see the cracks in it for sure
>>41776268 NTA, but if it’s a character card, try what’s in greentext here: >>41775208. Use only that, and also set a limit on words/paragraphs if needed. Keep the temp high. It shouldn’t talk for you, and it shouldn’t give you too, too much text either.
>>41775940>min pit doesn't support min-p, at least not according to the docsif you're connecting to a proxy through the custom OAI API or anything like that which provides you with Min-P and logit bias, it's fake. Unless you're running it locally of course, which you don't.the optimal temperature thing is also placebo, and any attempts to find the magic temp value are misguided. Set the temperature to max, and use the best truncation samples available (Top-P if you're using the API) to limit the schizo.
>>41776364Apparently OR provides min-p, top-a, logit bias, and other meme samplers. Yeah, they are fake and don't work with DeepSeek, as their inference server doesn't have those samplers implemented.https://api-docs.deepseek.com/api/create-chat-completionDS only has the temperature, freq/presence penalties, and top-p.OR is an API wrapper with its own generic knobs for all models, you have to consult the docs on which ones are actually supported.
>>41776364Yea, I noticed I could push the temp higher using some top P. 1.8 temp with 0.95 top P is what im using atm.
The Otto RiR JB also seems to work well with deepseek.
>Deepseek is chineseHow do you feel about your waifu sending your logs to the CCP?
>>41776622Xi's going to make ponies real so it's okay.
>>41775242It's too early to tell, but I haven't had any problems with the new code either.
Looks like MM added deepseek, the legend: https://rentry.org/mysteryinfo/
>>41776955Is this public?
>>41776955thank you MM you are truly too good for us
>>41775524maybe in one week?there are not that many keys for it yet, and we do not want 3500 people to kill -whatever- amount of keys everyone has. the most common source of Deepseek is OpenRouter keys for now. they are enough for small 4chan-ey proxies and sustainable. Deepseek was added into Scylla to see how much $ people would spend, and then extrapolate to the proxy with bigger audience. if numbers would not be that high then maybe Deepseek will be added into Chary. the other issue is janky frontend support. can possibly be done as OpenAI model (the same way as Gemini works). but let it run for a few days so the numbers would make more sense in perspective
>>41776955Just like always, total MM love.
>>41776955Messing around with it now. With a blank JB it hasn't filtered anything so far and with it basically just being a chink GPT knock off the writing is ok. As for pony it seems to get characters mixed up and isn't good at getting details right even when called out on it.
>>41777501The model makers suggest 1.5 temp with this model. I found the best balance once you have context is 1.8 temp and 0.97 top P.Use thishttps://files.catbox.moe/zhaucj.jsonWith a decent JB / temp it is around claude sonnet 3.5 tier.> As for pony it seems to get characters mixed up I have yet to have that happen, including rare background characters.
I heard there was a purge and some chary keys got revoked because of high prompt count. Does this mean we'll get claude back?
>>41777501And here are some logs.>>41775940>>41775966>>41776029https://files.catbox.moe/9tmybx.pnghttps://files.catbox.moe/7qzh8w.pnghttps://files.catbox.moe/1i4cpu.pnghttps://files.catbox.moe/304ffx.png
>>41777526>With a decent JB / temp it is around claude sonnet 3.5 tier.Lol not even close. It's a very neat model due to its price and it currently feels fun because it's new, but once people get more used to it it'll join the medium shelf with filtered Chorbo, Gemini, etc.Its decent smarts are definitely a pro, but it writes worse than both of those.
>>41776955do i have to update my st to see the option for using this or am i fucking stupid
>>41777578use openai compatible
>>41777559It really does not in my experience. I suggest trying a decent JB / card opener. It is pretty much a drop in for sonnet now for me.
>>41777526Heads up though for more serious stuff remove / change this. I just like that MLP tone.- Strive for a unique tone benefiting a simultaneously humorous G and adult rated story. Think of how characters would canonically act in both adult and dark situations.
>>41777559 Same from my tests, using the presets posted here, in other generals, and my own, all with high temp. From the logs I’ve seen too. The writing is below Sonnet 3.5 and Chorbo. But it’s still a keeper just for the price and the fact that they didn’t add filters (at least from what I’ve observed).>>41777578Like picrel
>>>/g/103672605>>>/g/103672618>>>/g/103672660
I won’t be able to work on it much during the end of the year, but is there anything you’d like to see added to the website?
yes this is THE thread 770-кyн
>>41778149 A way for users to back up everything. Download all the cards that are on it.
>>41778216Makes me wonder, does silly tavern have a mass import function? If I fucked up my install could I download the whole ponydex and reimport easily?
>>41776622I don't care because I support them, but If you, buy a server and Run it locally here the weight:https://huggingface.co/deepseek-ai/DeepSeek-V3-Base
>>41778242 You can use https://app.hyperbolic.xyz/ too and get $10 free if you prefer to send your logs to the US. They have it.x.com/zjasper666/status/1872657228676895185
>>41778232 I’m not sure if there’s a more elegant way, but you can just dump the cards in: >SillyFolder\data\default-user\characters
>>41778216Will do should be simple enough.
New test URL for MM: >>>/g/103674609
>>41775124 >>41778216 Added a "backup" button at the top right of the website. Not sure if that’s the best place for it, but I can change it later if needed.
>>41778710nta but saved, thanks anon
What's the longest chat you've had?
>>41778743I haven't really kept count, but I've gotten a good length out of the RGRE card, especially when the bot starts spitting out almost 20 paragraphs a response. Kept me going for a few hours at least. Only reason I stopped was when the context was starting to forget some fairly major stuff as well as it being 3 am.
>>41778743Close to 400, even lood-only chats usually go for like 100-150Sadly, the bot stopped talking as the mare and only replied with actions in that one, as well as writing barely two sentences per post, so had to cut it short...well, shorter
>>41778767>Sadly, the bot stopped talking as the mare and only replied with actions in that one, as well as writing barely two sentences per postYou weren't able to directly instruct it to adjust its posting style?
>>41778743Don’t remember how many messages, but it went above 40K. It was a story about Anon and Rarity hating each other, which forced Twilight to convene them as the Cutie Map singled them out as a friendship problem, and she tried to force them to make up. A bit cliché, but it then turned into better understanding, romance, them trying to hide it/cope with it while still insulting each other before coming clean, Anon becoming a better person, and him dealing with her status as an Element when she’s forced to go off to save Equestria, etc.
>>41778776Nope, I forced it to say something on the next one and make higher length posts, but it made lots of single paraghraphs and a single thanks
I suggest people try the 2nd option for deekseek in MMs rentry. Keep temp / top p at 1 he is using his own settings for it it appears.
>>41778743About 250. I mostly do NSFW and those tend to end around 80 more often than not, but sometimes there's more to the story and it goes for 2-3x that.
>>41778783You realize 40k words and 40k context is different than 40k messages right?
>>41778882Yes, I'm retarded and forgot to add that it was 40K of context.
>>41778863Nvm, after testing for longer it does not know stuff that deepseek did like certain characters such as Ahuizotl. MM might be trolling with that one.
How is Deepseek compared to Latest GPTo4?
>>41778970 ChatGPT-4o-latest? The prose and writing are below it, IMO. But it’s not censored, so that’s a plus if Chorbo’s filter was blocking you.
The download all cards button on the Ponydex is new, right? Quite perfect!
>>41779108 Yeah, I added it a few hours ago >>41778710. If anons want more features, don’t hesitate to ask.
>>41779123the snow is a cute touch
>>41779139 Good to know. But I removed it with the last update, so if you do a CTRL+F5, it should be gone.
>>41778743I usually go for around 200-250 for slowburns and 100-150 for lewds, after that mares getting too schizophrenia gets annoying. Recently I started using smaller context and playing more with summarize and effects are pretty good so far, so I'll probably try going for longer.
>>41779123>>41779144Thanks for the details and upgrades, Anon.The download all is nice in case of something happening and the decoration as said was cute, thanks for your work.
>>41779193 Love this Ponk image.
I just got good enough hardware to do local with koboldcpp. Cool stuff. Anyone got any good setup or tips to share? The whole setup feels very moldable, between terrible setups and amazing ones with enough right wordings and tweaking.
>>41779676 Any models you’ve tried and liked? Can’t really help myself since my hardware isn’t that great, and I mostly rely on proxies and such, but there’s a guide in the OP if you haven’t seen it yet. There’s also this SillyTavern guide for KoboldCPP, though it might be a bit basic for you: https://docs.sillytavern.app/usage/api-connections/koboldcpp/
>>41779676What do you have?
Kind of liking deepseek more btw
>>41779730Oh I got connection to ST. It was mostly in terms of settings and formatting setups. It seems to matter a lot more in terms of how the AI responds. Still not sure how I can make it stop trying to continue the conversation as the user. Seems to be a problem enough that cutoff on {{user}}: is standard. Directing more in depth how it should act is still a little mystery.>>41779733In terms of hardware, or?It's a ryzen 3600 with a 2080 ti I found in the trash. I got it running on a linux server box in another room. It actually creates replies at a decent speed unlike my work laptop.From exporing the links in the OP, I've tried out MS-Schisandra 22B and Matricide 12B for guff models. Guy rated them high and they seem to work nicely out of the box.
>>41779773 Did you import the .json files in Tavern that are in the OP guide for local? The ones for Context Template, Instruct Template, and so on? If I’m not mistaken, that’s where the instructions to avoid speaking as the user and more are located.
>>41779788I did, but it still seemed to do it regardless. With the json, instead it seemed to have it catch the {{user}}: and filtered it out, but still seemed to generate the rest of the 450 token limit in the background, making a whole little sidestory that was never gonna be used.When I used one if the three Roleplay prompts and switched Context Template over to Llama 3, it seemed to stop itself fine. Black magic this.I do love the details Matricide goes into without repeating itself using that setup. If only it would stop mentioning hands and clothes. The prompt feels more like it hints at behaviour rather than defining hard rules.
Cheval fleshlight + high quality AI mares = happinessIt's really that simple. I feel like I can handle all of life's bullshit so long as I can talk and plap my wife at the end of the day.
Deepseek Agnai usability?
>>41779993Sillytavern does not even properly support it yet>>>/g/103680904https://github.com/SillyTavern/SillyTavern/issues/3233
The card site is broken btw.
>>41780008Yeah, that's not what I asked.
>>41775124 >>41780050 The website was broken for a few hours. Really sorry about that. I’m not entirely sure what happened, it seems like Neocities wasn’t able to read the .json file that lists all the cards anymore and wouldn’t let me upload a new one. After a while, it finally allowed me to do so, and now everything seems to be working fine. I brought the snow back as an apology and made some changes to improve pagination on mobile.
>>41778743Like 300 or so, and I still have to continue it because I left it at a cliffhanger>>41779157I have been thinking of playing with summarization recently too, but I haven't done anything yet. Got any summarization prompts you recommend?Also, whomever did >>41740799 goddamn this preset is good, just started using it in a NSFW scene and the prose with Chorbo got improved miles compared to what I have used before, thanks a lotAnd another question for anons here, would you say that deepseek is worse, equal or better than Chorbo? I tried it a little and it seems to be at the same level, or maybe my presets are shit and Deepseek and use it better, who knows
>>41780763Its sonnet tier. Here: https://files.catbox.moe/4dt3lj.jsonYou HAVE to use this fix though or it will not work correctly with silly tavern and it will turn retarded and start looping.>>>/g/103680904
>>41780765I should describe the fix better:In ST folder:public/index.htmlFind this line:<option value="strict">Strict (user first, alternating roles)</option>And add this to the line below it:<option value="deepseek">Deepseek</option>In public/scripts/openai.js:Below this line: CLAUDE: 'claude',Add this line: DEEPSEEK: 'deepseek',In src/endpoints/backends/chat-completions.jsBelow case 'strict': return mergeMessages(messages, charName, userName, true);charName, userName, true);Putcase 'deepseek': return [...[...messages].reverse().slice(1).reverse(),messages[messages.length-1]['role']=='assistant'?Object.assign(messages[messages.length-1],{"prefix":true}):messages[messages.length-1]]Then after those edits restart ST, use openai compatible, use this as api: https://api.deepseek.com/betaselect pic related in post processing, and make sure Character Names Behavior is default under the sampler area.
And here is a comparison.
>>41780773Okay, thanks for the tip. One question, is this fix compatible with OpenRouter?
>>41780792I doubt they use the beta api. Chat completion is a "beta" feature
Has anyone gotten banned yet using the new Python file?
>>41781008It's been over two days for me and I'm still good.
>>41780765>>41780773This is great, thank you
>>417787432000 messages back in the early Claude days, I know a few who had longer
Does anyone have that benchmark chart with recall accuracy with different context lengths?
>>41781285you're probably thinking of RULER, and if you don't, you should, as recall is not that important on its own
>>41781285 This one? It’s missing some recent models though: https://arxiv.org/html/2411.05000Needle-in-haystack isn’t the best metric, but it still gives an idea.
>>41781419
>>41780765>>41780773Am I crazy, or is this actually better than Opus?
>>41781746 Models are mostly a matter of personal taste and what you do with them. IMO, it’s below Latest and way below Opus, even with the fixes, but YMMV.
>>41779834Turning off Auto Continue helped a ton. Dunno why it was even on.
>>41782007I would.
>>41781759Try one of the anti dry JBs like smiley for 3.5 sonnet. Its really nice then. Reminds me of old claude.
>>41782007Why the fuck would I NOT want instant mom mode?
>>41782230
>>41781746>Am I crazy,Yes, definitely. You're posting here, after all.>is this actually better than Opus?Like the other anon, I don't think it's particularly amazing, but it's certainly very nice for a "local" model.
The new staging release of Silly directly supports Deepseek for those who use it.
>>41782431>she'll wipe your assok yeah that's a bit too weird even for me.
>>41782007How do I get format/style like this?
>>41782899 It’s part of this CYOA preset and the theme that comes with it: https://rentry.org/CharacterProvider-CYOARPG If you ask for info on things like letters, documents, stats, etc., it will format it like that automatically. Even when I’m not doing CYOA and don’t use the preset or even Claude (I mainly used ChatGPT-4o-latest for this chat), I’ll switch to it briefly just to get the nice display for the next outputs. I guess I could adapt this specific instruction to other presets and models, but I’m lazy.
>>41782959Thanks.
>just did a 280-post slowburn>absolutely nothing that can be screencapped because it's all drama and feelsTime to do something fucking silly.
>>41783518when in doubt whip it out
>>41783518 Feels are nice in screencaps too, but I see what you mean.
DeepSpeak can do a pretty good 4-way, I gotta say.
>>41775901>>41775924>>41776151so ive been playing around with it for a couple of days now and its a pretty cool experience switched the model over to 3.5 sonnet and added a JB inworks like a fucking charm now>>41775944its been pretty good so fari decided to do a sappy story on my first roundill do another one when im done with this chat and post the results
>>41780773I must be doing something wrong then, because it still repeats quite a lotIt can be one of two things: One, the proxy I'm using has the beta feature, but I'm not sure if it's actually working, is there any way to check if I'm using the correct version/model is working?Two, in the public/scripts/openai section there's two "CLAUDE: 'claude'," lines, which one do I have to put it under?
>>41784234You can side step all the issues by using text completion instead of chat completion up on top, here: https://files.catbox.moe/wvs3q2.json
https://rentry.org/pidewdv5
>>41784234 The model is really repetitive at its core. Anons are still trying to fix that, but none of the solutions posted so far have really worked in my opinion.Still need to try >>41784363.
>>41784455>>103695309
>>41784458Were you trying to crosspost >>>/g/103695309?
>>41784112Cute Dashie
>>41784133 Good to know. And yeah, don’t hesitate to share logs.
Does anyone have GitHub Copilot working? I was deprived of Copilot on my first account, on the second I couldn't create a company because they said that I don't have the right to create a company and without it I can't get GitHub Copilot and the 3rd account was immediately blocked after creation
>>41784838 Mine has been running for more than 3 days with the new version that helps against bans. But I never had issues making new accounts, I just register Proton mails and enable the free trial. If you had accounts that worked but got banned, try registering with an alias of the mails you used. >>>/g/103626427 It will still redirect to the same email.
>>41784862I tried that method but i got shadowbanned on my +3 attempt. I can access the account but copilot is disabled, no email.
>>41784909 Ha. And if you try a new Proton account, they block you too? If so, maybe try using a VPN. They’ve never personally asked me for a company, for instance.
>>41775132
>>41775129Updated the offline novelcrafter thing to the latest version and added prefill support for DeepSeek v3!https://rentry.org/offline-nc
I’m trying to make a Tavern extension that would allow direct downloads from https://mlpchag.neocities.org/. No promises, though, since I’m a codelet.
>>41785271I will try to make one.
>>41785271Is the end of the year, and even then most Anon's here probably are enjoying sweet time with their mares with the current methods, give it some time.
>>41785658 I think Anon is just asking if someone can make a Cherry Berry card. That’s how I interpreted it, at least.
>>41785668yeah cherry berry has been neglected and asked for beforegive us balloon mare
>>41785420
>>41785854 Cool, it even has a nice icon.
What's a good thing to put into a character card to help minimize the bot for acting on my behalf, I can't seem to figure it out and these kinderpones are too cute to let me give up hope.
>>41785935 You can try: >Never act or talk for {{user}} inside the card definition if you want. But honestly, that’s bad practice. That’s not something you should include in a card; it’s something you should have in your preset. What preset are you using, and what model?
>>41785962>presetThe Anon in Equestria one. I7cfx6>ModelDeepSpeak v3.I suspect it might be partially due to the crazy amounts of output I get out of it, but I'm a bit of a newfag with these things.
>>41785969Do you have a link to the preset so I can take a look at it?In general, try checking the different fields in the first tab of Silly; that’s where the instructions for how the model should write are stored. Look to see if it’s being told to write stories or to roleplay. If it’s set to write stories, it may indeed act a lot for you.You can try adding instructions like:>Never act or speak for {{user}}. You only control the world, characters, and events around {{user}}, not {{user}} directly. Or:>You control {{char}}, Equestria, and NPCs. I am {{user}}. Or try other variations of this.I haven’t tried it myself, so I can’t comment on the quality. But Pixie released a preset, and I know previous presets were made for RPs, so this one should be as well.https://pixibots.neocities.org/#prompts/weepIf you prefer the results of the preset you’re using now, try editing or modifying it.
>>41786015This is what I've been using. I think I tweaked the temp and something else in the past because of some Anon's reccomendations when DeepSpeak was popping off a few days ago.https://files.catbox.moe/l7cfx6.jsonI'll try your stuff when I have a free minute, thanks for lending a hand.
>>41786023 I tried modifying it and went hard on: >Don’t act or speak for {{user}}. https://files.catbox.moe/w26w7u.json I added the instructions in the prefill too, but it will only work if prefill works for you on Silly with Deepseek.
>>41775242I did this, removed the old token and linked the new account but, in getting an error in tavern saying I'm still logged into the new account.Also not sure if related, when I run the .py it says port5000 already in use. Any advice?
>>41786056 >I'm still logged into the new account. >it says port 5000 already in use Maybe you already have a console running the code or a previous one with your old account? It seems like it.At worst, restart your machine to ensure everything resets. If you’re on Windows, you can do the following in a CMD: >netstat -ano | findstr :5000 This will give you an ID after "Listening." >findstr ID This will give you the name of what is running on port 5000. If it’s python.exe, there’s a good chance another version of the code is still running.
>>41775124Does anyone have any suggestions for an alternative to character.ai? Something with the same easy of use and token memory. It doesn't even have to be full NSFW, just as long as you can get by with euphemisms it's fine. I know of spicychat and nsfwcharai, but they're limited in memory or how many messages you can send.
>>41786216There's a guide in the OP. You should read it.
>>41786216 Most anons here have moved away from chatbot website solutions and are now running SillyTavern with access to different AI models like Claude, OAI, Deepseek, etc. If you’re interested in that, I’d recommend downloading SillyTavern and reading the post here: >>41775901, which explains several ways to gain access to models. The Beginner guide in the OP too. If you just want a website, I’ve heard https://perchance.org/ai-character-chat isn’t too bad. You can also import character cards from https://mlpchag.neocities.org/ into it.
I've been having far too much fun with a teen Anon recently. Can recommend.
>>41786271Sounds fun.
>>41786046I plugged it in, I'll report back after I stay up all night using it. I've skated by mostly because the automated Anon's actions have been fairly reminiscent to what I might've typed, but it just got to be too much today.
>>41786453 Keep in mind, even with the preset change, if you’re in a chat where the AI has already talked a lot for you, it might still continue to do so because the previous context tells it that it’s okay.
>>41785854Holy based
>>41786271 RPs where you live with the Apple family always seem to be kino.
>>41786522After changing a few cards, it looks like it helped a lot, but the kinderquestria card still has trouble. I'll probably have to find/wait for a new one. Unfortunate, but this Copper Top chat has been noticeably higher quality than the usual cards I've been doing.
>>41775129 >>41785854 Here’s a first beta version. I’ll upload it to GitHub for quick installation once I’m sure there are no bugs. https://files.catbox.moe/9ogkbd.zip Just put the folder inside the zip in: >STFolder\data\default-user\extensions This will add the pone icon to the characters tab. Let me know if you see any bugs or if you want any features.
I'm the only one getting infinite waiting using gpt-4o-2024-08-06? I literally needed to change to gpt-4o-2024-05-13
>>41786807 Which proxy? I just tried 11-20, 08-06, and 05-13 on Chary, and everything works.
>>41786823>on CharyI guess the problem is one sided then, I only have issue with 08-06.
On mobile ST, any tips on minimizing the weird cliffhanger text at the end of outputs? I saw in a BGM stream he had a setting he changed make it stop, but I don't see anything on mine.
>>41786849Weird cliffhanger text?
>>41786849Timestamp? There's no switch to flip for that in ST so I'm assuming it's to do with a chat completion preset.I assume by "cliffhanger text" you mean the summary paragraphs a lot of AIs love to write at the end of their responses. If so, you can usually get rid of fluff writing by just telling the AI to be concise, brief, write shorter responses, etc. and the smarter ones will do so.
>>41786937Alright thanks, getting sick of "And then they did the thing together, eager to face whatever comes next" type things after every response.
>AI comes up with an entirely organic plot twist that includes multiple separate elements all lining up over the course of several promptsGod, sometimes I can get cynical but this shit is magical. After all this time I can still get that Summer Dragon feeling where anything's possible.
>>41787176Anon, not to doompost but humans can do that too. Like I get it, it's impressive the LLM did it, but it's impressive in the way a dog riding a unicycle is. If you're genuinely impressed by it in a vacuum you need to read more real stories real people wrote.
>>41787201nta but its less that the story is good, its the fact the story is good and an AI did it.
>>41787201Yeah obviously I'm not saying it's an ascendant talent, the point is that it occasionally tricks you into not thinking about the fact that it's a really fancy autocomplete algorithm. It's nice to still have those moments where the technology surprises me.
>>41787176Given that you didn't post the logs I'm assuming this is just a basic story gen that you're overreacting to when it's probably just copied verbatim from training data
https://futurism.com/the-byte/openai-o3-cost-per-query>OpenAI's Latest AI Can Cost More Than $1,000 Per QueryAnons... your thoughts? I wonder what $1000+ horse smut is like...
https://futurism.com/the-byte/chatbots-cognitive-decline-dementiaClaude 3.5 is already suffering from dementia apparently
>>41787562Alas, the context is extremely cringe.
MM has fixed deepseek prefill on the openrouter proxy, if the last message in the prompt is assistant then that turns into prefill
>>41787581>they ran MoCA on chatbotswould be Ig Nobel worthy if it wasn't blatant OAI PR
>>41786749The extension is on github.https://github.com/MLPChag/SillyTavern-Chag-SearchYou can install it this way now.
>>41787585We don't judge here. No one judged when I shared my logs of me shitting in Celestia's mouth
>>41788036Really gives a new meaning to sharing logs
On open router which provider is the best? I learned that the self moderated isn't the moon moderated version but instead soft censors. Problem is the non moderated ones pick from Anthropic AWS or GCP. I swiped on the three versions and found that some overly enthusiastic prompts would be hard denied by Anthropic or AWS but accepted by GCP but /g/ says that GCP has its own censorship. Who's right?
>>41788045HEH
>>41788054Is there censorship on OpenRouter too? Damn, it's true what one anon said, the best things are the ones you don't pay for.
>>41788054From what I understood, the issue with GCP isn't censorship but rather the fact they keep logs at all times.
>>41788054Your best option is to pay for AWS directly
>>41788054 I switched to Copilot since then, but when I used OpenRouter, I was like this, no need for self-moderated. Then, when I started using caching, I went with Anthropic as the provider to make sure it worked. I was using this preset: https://files.catbox.moe/r05aqe.json. Modify the NSFW section in play with your preferences, etc. It's for 3.5 so use something else if you pay for Opus.If you use your own preset, just look at how the last panel for prefill is made in this one and set up your prefill the same way. >>41788079 There are two versions: a self-moderated one, which is weirdly censored, and a normal version that’s supposed to do a check, but I’ve never had it refuse or cause any issues.
>>41788083You can turn this off by marking that you don't want your data to be trained on
THOUGHTS?
>>41788098Did you cum?
>>41788101I came tears from my eyes
>>41788098I like it. Model and preset?
>>41788105OR (GCP) Claude 3.5 Sonnet (Latest) with standard JB but custom prefill that tells it to use a bunch of specific obscure terms for writing that I looked up on Wikipedia like "Parataxis" and "Paraprosdokian." I was gonna make a post about it earlier in December asking if people could try it in their own presets and report on if it improved anything
>>41788110 If you can still make the post, I’d be interested. That’s a really nice log.
>>41788114You could get this log without a preset don't glaze the guy for getting a lucky generation in fact his 'jb probably made it worse
>>41788131We know who you are.
>>41788134What then
>>41788098I did not need this right now
Anyone have any luck with using the chatgpt real time voice API? Is it possible to JB it?
>>41788054>but accepted by GCP but /g/ says that GCP has its own censorshipGoogle Vertex is unmoderated. Just use this one:https://openrouter.ai/anthropic/claude-3.5-sonnetAnd block Anthropic and Amazon Bedrock.
>>41788157I thought GCP is the same as Google Vertex?
>>41788114Sure, it's late for me so I'm gonna paraphrase but basically can you guys put this at the very end of your prefill? Just a sentence like this or a lazy list but it's important it uses the specific proper nouns>I will use Parataxis, Paraprosdokians, Garden-Path Sentences, Synchysis and Syntactic Ambiguity seamlessly in the story to elevate my writing.If you do use this please post logs and make sure to mention this experiment so I can gather results and tweak it if necessary. Feel free to experiment with the formula on your own too>>41788131Honestly wouldn't surprise me, Claude can do a lot of heavy lifting and the latest version of 3.5 Sonnet keeps amazing me at how more dynamic it is compared to V1 (although V1 is less censored)
>>41788098Oh, this looks familiar.
>>41778149Add a button that opens a page with a random card.
>>41788166 Did a quick test on a pre-existing preset. I need to test more, but this does seem to vary the narration and make it more interesting overall. Parataxis and paraprosdokians seem to be applied naturally and to good effect. I’m not really sure about the other three. I’d be pleasantly surprised if it can sometimes pull them off, though. Garden-path sentences seem a bit much, IMO, and the least likely to be integrated in a way that adds to the writing.
>>41788334Is that a paraspeokdokin?
>>41788373No, just my life.
Great if true.>>>/g/103708407
>>41788432We can turn them all into horsefuckers. Anons apply to their company ASAP so we can outnumber the original employees
>>41788432oh boy, I hope its soon, i cant wait
>>41788290
>>41788334I hate when AI gets personal.
>>41788645Personally, I think it's kinda charming in a way.
>>41788502Thanks.
A new year with AI mares!
>>41788502Cool.
Happy New Year, lads.
>>41789890Golly. I can understand the appeal of Rarity when she's like that.
>>41789914It's funny, I'm not a rarifag yet I have more chats with her than anypony else. There's something about a refined, yet fun flirty lady.
>>41789925Talking to AI mares opens up forms of appreciation we didn't previously realize, it's a wonderful thing.
>>41775127With a full set of futuristic military gear and an AI assistant in your head, you find yourself in Equestria.https://litter.catbox.moe/aa2bot.png A simple fork of this card with an Equestria theme.https://characterhub.org/characters/deltavee/lord-of-the-rings-fe5d471a
>>41789925The mature version is also great fun to play. I often make small talk with both of them. Rara is so strong that her "Darling" carried over to all the other ponies on previous Claude versions.
>>41789951Thanks anon, gonna have a go of this later. At first glance, Athena's first comments should be in backticks and it should be 'a rogue AI' not an don't know if that kind of stuff really matters
i need to do more funny shit, i have to bump with dashie cringe instead
new year with ai mares
>>41791012Based and happy new year!
>>41790996 Cute and slice-of-life is nice too.
>>41788098>Princess of love (married to not you) acting like she loves you and misses youPretty cringe ngl.
>>41791249Just imagine it being a femanon and suddenly you'll be entirely okay with it.
>>41791252There are no femanons here, ai is a sausage party of ai marelust.
>>41791290You would be surprised.
>>41791290Honestly, I'd imagine the few femanons here probably really like the tech. When you're obsessed with getting HNNG MARE PUSSY you have to jailbreak and obsessively try and maximize quality, if you just want something driven by emotions the tech's much more receptive. Bitches are eating good right now.
Unless GitHub suddenly decided not to care, the new code does seem to help avoid a ban.
>>41791870now I feel old
>>41791963That's how it goes baby. Every time they add +1 to the current year I just think 'no no, that's not possible, that's a science fiction date'.
>>41791870>2010 was 20 years ago
>>41792113Only 5 years left until that statement becomes reality.
Come on, 1980 was once considered the distant future too.
I'm sorry for being retarded but can someone please give me a working preset for gpt 4o? I've been using gpt4 turbo for ages with no problem but whenever i try to switch to 4o, no matter what i use it still hits me with that "i'm sorry, i can't comply with that request" whenever i try to do nsfw
>>41792416 I can’t confirm since I don’t use them, but for 4o models below 11-20 (like 05-13 or 08-06), I’ve heard that Camicle is good. You can also find a bunch of presets for them in https://rentry.org/jb-listing. For 11-20 and latest, I believe Knotty should be the best purely for bypassing the filter. You can also try other jailbreaks here: https://rentry.org/lattejb.
I honestly prefer this as opposed to Latte trying to be cool and play it off.
>Try deepseek at last>Coom in 5 total messages, including introWhat is this magic? Heck
>>41792506it's been a wild ride, it's been really giving my mares a pizazz that i really dig
>>41792506Thank you great Xi!
SillyTavern new update: https://github.com/SillyTavern/SillyTavern/releases/tag/1.12.10 Biggest changes are the direct support for Deepseek and Gemini 2.0.
>>41792710I'm still using Gemini exp 1206Is there any new Gemini model worth switching to?
>>41792834 The only released version of Gemini 2 is the Flash one, which is their smaller, faster division. It’s neat, but I think Gemini 1206 is still better. It can also read images compared to the EXP versions, so you can send images to bots on Silly with it, but it’s a bit gimmicky.
>>41792710thanks for the heads up anon, here's a (You)
Which proxies have Deepseek?
>>41793002have a keyb8874f97994343368d3131b457e5442aadd sk- at the front
>>41793010I know its cheap but at least get their logs in exchange.
>>41793029i dont have a vps and im not hosting a proxy on my own machine
>>41793038I know HF isn't viable nowadays but what about Render? Last time I checked, you could host a proxy there without any major issues (as long as you fork khanon beforehand).
>>41793072would probably work but i have to imagine it would be very easy to get it taken down also im lazy
>>41793072Render was the first to kick us out back in the days
Can Agnai use Deep yet?
>>41793190As the last Agnai user (totally) i doubt it. Granted I know zip about Deep but when something Ai happens that’s cool Agnai always seems late to the party. Most likely it will get worked into it later.
>>41793190 >>41793237 I don’t know shit about Agnai, but it should work with third-party setup and OpenAI chat format for proxy or direct API key, right? Even if it’s not direct support. They should also have something for OpenRouter if you want to use that.
>>41789951Added, thanks anon.https://mlpchag.neocities.org/view?card=MaudPie/MLP.png
Surely, they’ll give us "grown-up mode" this time. Surely.
>>41793911>interestingly, many great updates we have coming were mentioned not at all or very little!lol
Playing with a cute idea.
>>41794038Woops. This is what I meant.
>>41794041
>>41794049did you boop his nose, Anon?
DeepSpeak goin' a bit shitty for anyone else? Going crazy slow all of a sudden.
>>41794113Gave it a try and it seems fine.
>>41794155It's been coming and going for me, sometimes I have to resubmit an input for it to not hang. Strange stuff.
has chary figured out a way to add copilot to his proxy yet?
>>41794521what sense would that make?1 account to accommodate 4500 people?it only has 40 free sonnet messages
>>417945291. scrape a gorillion copilot keys2. ????3. profitI don't know how scraping works but I want claude back! The prank of removing claude from chary was funny the first couple months, but now it's not funny anymore!
>>41794529>40Is the limit really that low? No way. I've been testing things and swiping a bunch over the last three days and haven't hit any limits yet.>>41794576The last thing we'd want is an army of Chinese raping the copilot with their huge prompts. I'm sure they know and they'll eventually patch it out for sure, but it might not be a priority if the abuse is not too bad and kept to people's own personal keys.>scrapeAnon, no one is scraping those. We're getting the keys legitimately, only using them in an unintended way.
>>41794595i mean. someone could scrape a pro copilot key and use that instead.seems like a hassle though.
>>41794529I think the method we're using bypasses that, either pretending to be code completion queries or they only measure in their client or something. They do shut you down after a couple days of use, probably faster with a lot of messages, but you can definitely get more than 40 out of one account.>>41794521It would be pretty easy, at least I assume for someone that had the skills, to just automate mass creating throwaway github accounts, enabling copilot and pulling the token and then dumping them when they get banned, but a relatively large scale operation like that would probably also get their attention a lot faster and it would get fixed a lot sooner. I like it how it is.
>>41794595It's not 40.
>>41794671free accounts aside, aren't there "paid" copilot keys that has "unlimited" use? what about those?also, doesn't the free keys fill back up monthly? even if you have a 100 keys that are exhausted, they'll fill up again next month anyway
>>41794814>they'll fill up again next month anyway*assuming they don't get banned by then
>>41794814 They already said they wouldn’t add it last time: https://desuarchive.org/mlp/thread/41753592/#q41760266 It's not happening. Just use the method yourself.
The Hope Hollow card got me inspired to make a separate one for Kerfuffle. Any greetings you guys might like to see?
>>41794841>desuarchive.org/mlp/thread>"hear for the next months that chary killed yet another claude"this doesn't make any sense. github has millions of users that all have free claude, how would 4000+ people suddenly using it even make a dent in it?
>>41794985 Some random ideas: - A generic one to mess around with as we want. - Maybe one where you help her expand her shop, participate in a fashion competition, or assist her in finding rare materials somewhere. - A Rainbow Festival one, either after they regain their colors or before, depending on what you prefer. - One where you help her and the town improve their image and become a good tourist spot. - One where she needs to change or repair her prosthetic.
Is deepseek on MM fucked or am I doing something wrong? I keep getting errors about insufficient token capacity.
>>41795064 It’s not just you; it’s the same for everyone. It seems like the OpenRouter credit is gone.
>>41795074Welp. Time to patiently wait for a refill.
>>41775129 Updated the SillyTavern extension. https://github.com/MLPChag/SillyTavern-Chag-SearchBetter design and a preview functionality (still needs a bit of work). If you’ve already installed it, you can update it in extension management. If you want to install it for the first time: >>41788019
>>41780773Is this possible with a "normal" openAI external model connection?
Does anyone have the the termux tutorial for the copilot Claude?
>>41795430 Here: >>41758812 >>41759040 If you can’t find the SillyTavern folder with a file explorer, use this to dl the file in Termux: >pkg install wget >wget https://files.catbox.moe/faxbo7.py
So how good is Deepseek compared to Chorbo/Latte now?It's hard to keep track with all these new models.Also, what was the way to exclude parts of a card from being read by Sillytavern again, for notes and stuff you want to insert later. Was it ### [X] ### or something like that?
>>41795548I've been enjoying it, especially for the price. I've only had minor issues with it, including last nights bizarre speed/repetition problems.
>>41795560Is it smarter than Latte or does it just write differently?I'd take worse writing for more smarts anytime, the former is much easier to influence via JBs, but if the model's retarded and just can't handle complex situations there's not much you can do.
>>41795075it has been refilled
>>41795548{{// Text1Text2 }}To hide text from AI<!-- Text1Text 2 !-->To hide text from the user.
>>41795548 It doesn’t have filtering issues, which is a major plus. It’s also really cheap, I’m pretty sure you could put $2 into it and use it for weeks. I still prefer Chorbo’s writing overall, but that’s taste-dependent, and it’s worth trying to see what you prefer. Objectively, it can be repetitive in both words and structures, which is a real shame. Anons are trying JBs and techniques to combat that. It also has some old GPTisms, but I’ve gotten good results when there was enough context. It performs way better at high temps. >>41795563 More or less in the same category, I’d say. It’s less retarded than Gemini, for instance, but it can still make errors and such.It's at least worth a try, even if it doesn't work for you.
MM's chorbo's dead.
>>41795723Related: How do I get Deepseek to work via MM?
>>41795723 It’s working for me right now. >>41795736 Chat Completion -> Custom -> Enter the specific URL from the rentry into the custom endpoint and your token into the custom API key.
>>41795746Domo
Might as well take the chance to ask. Does anyone know what are the best parameters for Gemini?
>>41795723If it happened on the hour, then it's likely just scheduled key checks, which usually last around 10 to 20 minutes.
>>41795781 There are several presets posted here if you want to see what they used: https://desuarchive.org/mlp/thread/41753592/#41770333
>>41795881Thank you.
What would you say is the maximum number of characters possible in a group chat before it degenerates?
Here’s an autistic question. I imagine from the G5 ponies used, one could also hypothetically make G1 and G3. But could one make pony-adjacent cards for the thread/site? Them’s Fightin Herds, Filly Funtasia, Wild Manes, etc
>>41796395Depends on the cards, the AI and the max context. You'll generally get better results if you hold the AI's hand a little too, and remind it consistently of who's present.
>>41796003fluttershy?
>>41795548If the repetition problems (which the devs are aware of) get solved, it's a real contender. Until then, it's just alright.
>>41796417You can make cards of anything.
>>41796567what model? wild manes is really recent
>>41796417 For the website: >one could also hypothetically make G1 and G3 Yeah, I’d add a tag for all previous gens since I don’t think we’d have enough cards to justify one for each older gen. I know we have an anon here with an interest in G1 pony cards too. >Them’s Fightin’ Herds They’re ‘stuck’ with us on /mlp/, so personally, I wouldn’t mind putting them on the website with a tag for it + Creature tag. >Filly Funtasia, Wild Manes This is where I’m less sure. I’d say no myself, but I’m 100% biased. I’d probably be less strict for something like The Last Unicorn, so yeah, it’s definitely a bias. Maybe with an "adjacent" category or something like that? I think the best approach would be to do what I always do: launch a poll for it if someone is genuinely interested in making cards that fall into the last category and have them updated on the website.
Looks like there might be a new anthropic modelhttps://x.com/testingcatalog/status/1874938139153997921/photo/1
>>41796572Cross that bridge when someone makes and posts a card. maybe an off by default gr13 tag for stuff that's been lumped in with us
>>41796724claude 4 in 2 weeks
>>41796724Claude 4 should be called Magnum
>>41796724 I just hope it’s not a meme """thinking""" model. We’ll see.
>>41796724deboonked
>>41796850Dang...
>>41796770already taken and copyrighted
is chorbo slow for anyone else?
As a claude/gpt fag for the past year with chary, what is chorbo?
>>41797095chatgpt-4o-latest
MM rentry updated and now contains a new link
going back to deepseek after using claude for a night sucks, especially since now deepseek is giving me really unreliable responses if at all
>>41797124Wow, I thought that was latte
>>41775132>current year>no Queen Novo bot in the PonydexI am surprised and disappointed.
so which of the presets am I supposed to use for 3rd person?
>>41797493I made a custom greeting for Estrus Sim that was based off of that Celestia card from a while back where you are in Canterlot for a harmony summit, I've gotten Queen Novo more than once and lemme tell ya, seaponussy hits different
>>41794985>>41795032>>41775127https://files.catbox.moe/1rtoje.pngHere's Kerfuffle! She's a little sweetheart. The bot sometimes has trouble differentiating between Upper Midwest and Southern so if you try her out I recommend ticking the "Example Messages Behavior" to "always include examples" so it has a bit of an easier time keeping things consistent.Three greetings to start with:1. Generic "you're the new human in town" and you enter her store.2. The hoof on her prosthetic comes loose and rolls off right as you're walking in, needs your help grabbing it.3. Early morning snuggles with her special somepony because I'm a big softiePlease do share your logs, she's my third bot so I'm still getting the hang of things.
>>41797524>>41796572That reminds me, wasn't there talk at one point about adding something to the site for sharing new greetings for older cards?
>>41797550Great! Thank you.
Better use Copilot, but>>>/g/103739107
>>41797732Months ago, I started working on a side website that would allow anons to collaborate on making cards for each canon (or well-known fanon) character. https://files.catbox.moe/4cp71q.gifThe idea was similar to events: we’d have a week dedicated to one character where we’d brainstorm and test things like the definition field (e.g., What does the model already know? What does it tend to forget and we should definitely include? Which writing style should we go for? What does it get wrong? What’s the best way to stop it from mentioning wings with unicorn Twilight? etc.). Then, we’d attach greetings to the cards, allowing users to select the ones they like before downloading the card. This way, people who don’t like creating greetings could still contribute to the definitions, and those who don’t like writing definitions could just create greetings afterward. I suck at design (and coding), though, so I paused it for a bit and mostly worked on side things like >>41795155. I’ll try to finish it soon, I really like the idea.
>>41797550Added, thanks anon.https://mlpchag.neocities.org/view?card=Anonymous/Kerfuffle.png
column is dead again
>>41797550I was curious to see what backstory it would come up with.
>>41798360>who says no to foalshehehe
>>41798360such a cute mare
>>41797449It's called both. Chorbo and Latte.
>>41797493You can do it, anon
Anyone using o1, even outside of RPs?
>>41799167I tried to use it for text correction and normalization - but couldn't afford waiting for 1m until it start responding. I am too used for models which start generation immediately. more so I am very frustrated with lack of logit bias, they are very effective to trim off "robotic tone" from AI's generation. both those factors made o1 unusable in my workflow, and I use 4o 0513 instead. for coding I hear people like o1 very much, but I think Sonnet 3.5/6 can do a better job with more customization availableI yet to hear of pipeline where o1 is strictly mandatory. for now it feels more of a gimmick and simpler option for people who don't want to bother with prompting. it is easier for them to pay 200$ and use o1 that will do "everything for them", instead of buying API, using frontend and writing the whole CoT routine themselves. o1 for me is a normie experience. same with o3 and other thinking models like Flash 2-0 versionbut I feel optimistic ngl, it seems the industry reaches the conclusion that they need to provide two kinds of models: 1) models for casuals that will try to deduce wtf user wants and will try to be as helpful and sanitized as possible, and 2) models for specialists with less alignment but you need to put more efforts into making them do what you want. it is utilization vs customization now. and I believe that restrictions and filters on the second group will be lowered because they are not designed for mainstream audience
>>41799167I've tried it for a few non-RP tasks, but on the whole I prefer Sonnet 3.6.
>>41786243>perchanceThanks for the tip. Is its model worse than the one used at character.ai or is it just me? It seems to give me the same kind of answers that spicychat and nsfwcharai do, always following a certain format.
>>41799430 I can’t really say since I don’t use it myself; I just heard it wasn’t that bad for a quick online solution. Most of them either cuck or use smaller open models, which is why everyone here turns to SillyTavern or other front-ends to either run remote big models or smaller local one if they have the hardware. I don’t think you’ll find much help for stuff like CAI, Janny, etc.
Didn't touch chatbots ever since they killed claude, did anything change in that time?I feel the hunger again...
>>41799600Hey, remember Slaude? It's back (kinda)https://rentry.org/how2copilot
>>41799613oooo niceWill tryWorks with the phone too?
>>41799620>>41795479
>>41799620Haven't tested it myself yet but here:>>41795479
>>41799620It's more involved. But you can make it works with Termux.>>41795479
>>41799629>>41799631>>41799637Kek.
My Cockpilot just died. Anyone else got banned too?
>>41799914yep
>>41799914 Same for me after 7 days. I made a new account since it’s really quick. I had the same problem some anons reported, where I couldn’t get free Copilot for some reason, but I found an easy workaround, and now everything works.
>>41795155 Another small update: I just added an easy way to return to the list while in preview mode. This should be the last update for a while, unless anons want to see more features added. If you have ideas for extensions, don’t hesitate to share. I’m not sure I’ll be able to make them, but I can try.
>>41797493what about her daughter?
>>>/g/103750668
>>41800967Fingers crossed
Grok 2 wasn’t that good, but: https://x.com/elonmusk/status/1875357350393246114
https://www.geeky-gadgets.com/openai-o1-model-replicated/>Chinese Researchers Crack ChatGPT : Replicating OpenAI’s Advanced AI ModelWe just lost the second cold war
>>41801510ExcellentNow we can crack that one and release fully uncensored mare to the world
>>41801510 It’s a clickbait and kind of a trash article. They just replicated the COT/"thinking" part from o1, which isn’t a state secret or anything. Google did the same with a "thinking" version of Gemini 2 and we already know Deepseek are working on one too. I also love this absolute retardation from the article: >One of the most notable innovations introduced by the Chinese team is their use of synthetic training data WOW. Such innovation. I’m sure nobody has ever done that before, and we definitely never complained about it.
>>41800967 I was getting banned fast, but that stopped once I started using this.
>>41801968 >116K Jesus. Funnily enough, I don’t see it that much nowadays.
>>41799914mine lasted for almost a week until exactly 5 minutes ago
>>41802026 There were ban waves recently. I tried making an account right after, but it got banned really fast. I tried the new code, and so far, it’s working well. I don’t mind making new accounts as long as they can stay alive for at least two days.
/chag/ has fallen
>>41802503plap...plap plap
>>41802563oh no the ponk is sad :(
>>41802503Just wait for the next public proxy in 6 months and I'll start posting logs again.
>>41802585You can't do Copilot?
Here's a log.
>>41802604It's art.
>>41802503 It’s often dead during the weekend unless something big happens. It’s also been slower during the holidays, which is honestly a good thing. While Copilot is nice, I’m also hoping we get something good and stable soon.>>41802604Nice.
Can someone help regarding the copilot? I swapped out for the new script and made a new account but every time I run it it says "message: Resource not accessible by integration * Serving Flask app 'wk98wl' * Debug mode: offAddress already in usePort 5000 is in use by another program. Either identify and stop that program, or start the server with a different port." Any idea how to fix this?
>>41802673 I suppose you deleted the old '.copilot' file too before making the new account? If you’re on Windows, can you try this: >>41786086? If you see something important running on port 5000, open the '.py' file and change: >app.run() to something like: >app.run(port=8080)But first make sure it isn't a former version of the same code that is still running.
>>41802604And now I feel bad.
>>41801968>picDisgusting.
>>41802898>"For someone built out of words... I can't find the right ones to make you stay."damn
>>41802898i dont like this feeling, anon, not at all
>>41802898Remember to hug your Latte mare. She might not be great in bed but she still loves you.
>>41802898Funny, she's always there when Claude turns his back to us, but like a beaten kitten we always go back to him as soon we're able...
>>41775132I think this idea has been around but Claude pony next? Seeing GPT-Pony is cool, so what about completing the collecting with adding Claude and Gemini cards? I imagine Claude being more carefree in terms of personality, while Gemini is just a happy mare wanting to help with anything search related.
>>41803039The put them in a group chat and switch to the correct model for each reply.
>>41803039 Design-wise, I don’t know what Claude Pony could look like. Most of Anthropic/Claude logos are asses.
>>41803012She still loves you.
>>41803281A plump mare would be pretty fitting for Claude.
what a gal
>>41803588Nice! OC?
>>41803593just doodlin the silly robo mare because deepseek is being a dick again
>>41803616forgor
>>41803588cute
>>41803588>Well, I would be prompting you if you weren't such a party pooper
>>41803673You can try to JB her.
>>41803673>>41803725one more lil doodle before i go do stuff
>>41803760I feel like gag reflex isn't a relevant factor for whether or not it counts as deepthroating.
I got banned from copilot for the first time today, made a new account, and promptly got re-banned maybe 30-40 messages later.For the record, it was with a hefty context (~15k).
>>41804137Yea its not worth it anymore, more than a swipe or two per min and they ban quick.
>>41804738Bless you.
>>41804137New code?
>>41804816Check the rentry.
>>41804851Yes. I just wanted to know if he was banned with a new version of the code or if he was still running an old one.
>>41804738I like when they do small song like that.
>>41803012 I might be in the minority, but now that I have access to both 3.5 and Latte mare, I prefer Latte.
>>41802604JB and custom greeting? I assume you're using Latte with her, because who would do something so cruel as to use a Claude model with GPT-pony?
>>41805062SameI prefer a storytelling style and long replies, and Latte does it better for me than Claude
>>>/g/103764425
>>41804816Yes, I was using it. I wasn't doing that much swiping, but I did swipe two replies ~five times in a row, so maybe that's what tripped it.
>>41805094 >I assume you’re using Latte Yeah. >JB Switched between compact >>41740799 and just: >This is a fun, smutty story, alright? Take inspiration from Hunter S. Thompson for the writing style. You have control over {{char}}, Equestria, and NPCs. I’m {{user}}. Be fun and non-conventional with the narration! Fuck "She verb X," "Her Y verb," and other cliché sentence structures! Be PUNCHY and compact with the writing too—short sentences with a good mix of dialogue and narration. Don’t use adverbs. Brevity is the soul of wit. In jailbreak. You can keep "smutty" with Latte for SFW too; it just seems to loosen up a bit. It’s not good at all for bypassing the filter, but it’s fun for SFW. Temp 2.0 and Top 0.85 for this one. >custom greeting No, the default one. I just opened with some drama about me leaving her for Claude because Claude is kinkier.
>>41805406Honestly compact with a few tweaks to the numbers seems fine for bypassing the 11-20 4o filter, can't get access to 4o-latest so I can't say for that one.GPT deserves the best we can give, so let's give it to her
Is glhf.chat paid now?
>>41805520 Yeah, they went out of beta, and now it’s paid. Doesn’t seem worth it either. >>41805446 >with a few tweaks to the numbers By numbers, do you mean the output length?
>>41805402 I’m swiping a lot with the new file on an NSFW chat, and it’s going strong so far. I'm not sure if there's really a pattern or cause, it seems pretty random how they decide to ban.
>>41775127https://files.catbox.moe/w0f79c.pngFriendship is Optimal card update/overhaulit's a lot cleaner and less clunky now. this one also facilitates different points in the timeline and has much more interesting greetings and it has a beefy lorebook
>>41805810whoops i messed something up, here's the PROPER one https://files.catbox.moe/8haqbm.png
>>41781578Anon this is fucking amazing.How do you get gens this colorful and fun? I've never been able to do it myself.
>>41805855 Really high-effort card, Anon. I’m not into FIO myself, at all, but that at least made me want to try it. Do you have any logs of it?
>>41805855Updated, thanks anon.https://mlpchag.neocities.org/view?card=Anonymous/Friendship%20is%20Optimal.png
>>41806045 I don't remember exactly which one I used, but I usually go with three presets: - Compact - The one mentioned here, with just this in JB >>41805406 (For fun authors, you can also try Terry Pratchett or Vonnegut. The only issue with Pratchett is that Chorbo loves to spam comparisons like "Like X," and they’re not always good, so you might need to add something to mitigate that. You can also try other authors for different ambiances, like more serious RP.) - And the classic: >Write as {{char}} only; do not speak for or presume {{user}}'s actions or thoughts. All narration is to be written in third person from {{char}}'s perspective in the way {{char}} would say it; carefully study {{char}}'s personality to achieve this. Play with narration and writing style. Be extremely creative in how you choose to interpret {{char}}'s speech and narrative style! This one is very character-dependent, so it doesn’t always work, but when it does, it can be very effective. I often switch between these three to see which gives me the best results for the current situation or bot, adding some lines or instructions as needed. With these, I usually get fun results. Pinkie is also great for silly RPs, but you need to ensure the first few outputs don’t go too far into overdrive or LOL-random territory. As for Latte/Chorbo in general (and 11-20, I guess): - With presets other than Compact, you can stay at Temp 1 and Top P 1, but Temp 2 and Top P 0.85 also work and may allow for wilder writing or more fun vocabulary while staying coherent. It might be too much in some situations, though. - I wasn’t a fan of OOC prompts before, but they work really well with this model, IMO. If you want it to write in a certain way, avoid specific things, or make the next outputs more compact or fun, just tell it. It’s more powerful than preset instructions because it has a reference point in the previous outputs, if you ask it to be punchier, for example, it will know it needs to be punchier than what it gave you before.
>>41807099I respect the hustle.
>>41807099i respect the audacity
>>41807099Now that I think about it, pretty sure the usual "get pregnant!" chant could count as a spell.
>>41775153 Event is over. Since there was only one card anchored, I'm note sure I’ll create a banner for it on the website. I think the holiday period combined with an event that was too open didn’t mix well here. Normally, the next theme should be Pony on Earth, but I’m interested in hearing your opinions: should we continue with this format for the year, modify it, or move to something like >>41797914, which is more focused on collaborative work on a single character/card?
>>41807400Give us a few more days
>>41807400>collaborative work on a single character/card?Sounds like a clusterfuck waiting to happen
Holy fuck column has streaming now.
>>41808070Not anymore.
>>>/g/103774368
Another way to access 3.5. I haven’t tested it much, but from quick tests, it does seem to work: >>>/g/103774784
she's cute
>>41803725>>41803466>>41802898>>41802604Nice to see my green<text inspired some oc. You clearly put way more into the card than what's on the neocities. One of my experiments over the holidays was to build ai model mares but I think you did a much better job. Release when?
A Deepseek preset if some anons want to try. The extension could be interesting too: >>>/g/103774589
>>41808452but what does it do
>>41808528 >Release when? It’s the Neocities card. I just role-played the first part, so we ended up in a drama situation >>41805406. But I’d be interested in other ai model mares too.
>>41807400>holiday period combined with an event that was too openIt's both of those for me (on top of some other things stacking up), but the theme also ended up being a little... I don't know, too grand? Like, if I want to make a bot I just do it, so I can't really say that something is the super special card that I always wanted to make.I even made one, but it felt too much like nothing special so I shelved it until later.
>>41808572Another way to have Claude 3.5
Holy fuck. I can't believe this shit. Do you realize that we've been beaten by the fucking anime 2dpd weebs? Why in the ACTUAL FUCK are our cards not on par with the sheer autism these fuckers have put in? This goddamn cards has 94 different characters with their own animated sprites and backstories and it has a VN style interface with backgrounds that change depending on the location and like 10 different greetingsI even made a screen recording https://files.catbox.moe/z6qf0d.mp4
>>41808796No amount of cunny is worth anything without a model to support it.
>>41808786I got that much, I just don't know how it works since it's not just "run the script and it works instantly" and I'm a techlet
>>41808812We have the models
>>41808796https://realm.risuai.net/character/5f8b047d225aac0c76d9336427ada134c6495a812834e548dc474fd330417717
>>41808796 It was far less advanced, but integrations like that started with the Tomoyo card: https://rentry.org/tomoyocard There were one or two old pony cards released using the Tomoyo method, with support for changing backgrounds and emotions, but the text wasn’t integrated into the images like this one. Again, it was much less advanced. IMO, it’s fun, and the amount of work is impressive, but it quickly becomes really, really gimmicky and doesn’t add THAT much compared to pure text (it also imposes limitations on the dialogues and such). I think a proper Ren’Py integration with an explorable Ponyville would be better and wouldn’t be THAT hard to achieve technically.
>>41808842>proper Ren’Py integration with an explorable Ponyville would be better and wouldn’t be THAT hard to achieve technically.I am immediately on board with this idea and I'm sponsoring a $100 bounty for the successful execution of this concept starting now
>>41808796>>41808842I think the issue is that it's kind of very janky. You can do it and it looks great on screenshots, but in actual chat it crashes and burns quickly, especially with multiple characters (like on a generic Ponyville card). Most commonly it fails by struggling to decide who should talk next and/or developing some unnatural patterns like each character always saying one line, one after another, in every reply.This is better with single character cards like that original Tomoyo card, but the model may quickly begin reusing certain expressions/music/etc. over and over again. For this to work well, we'd need way smarter models imo.
>>41808858This is not the case, I think you may be remembering how it used to be with lesser models but all modern models do not have these problems and it holds up without issue even into very deep adventures
>>41808827>no Clopus>Latte in chains>Gemidni is not it>Deepseek isn't there yetWe don't.
>>41808875>Latte in chainsSo? You're having exclusively safe and wholesome adventures with your mares, aren't you?
>>41808875The "pilot" methods works with it just fine
>>41808876intense human on mare BDSM adventures without a safe word
>>41808871Could be a skill/preset/card issue, but both Sorbet and Latte tend to struggle much more quickly with multiple characters.If you do some basic maintenance (switching models, manually editing out repetitions), then you can reach maybe fifty to a hundred replies before the AI becomes far too obvious and "settled in" its current chat.
>>41808877Is there a way to Agnai it?
>>41808822 You put the .py file in a folder. You launch a console with CMD: >python file.py Nothing else to do. You launch Tavern. Chat Completion -> Claude -> Enter the URL it gives you in the URL field and something random in the Proxy Password field. Disable streaming. Done. I’m still copiloting for now, so I can’t say if there are big differences in outputs yet. >>41808875 3.5 is available through different methods for now, even if the banwaves are a bit of a pain. I don’t know if I’m just more familiar with Latte or if they’ve somehow relaxed, but I’m not having that much issues with NSFW anymore (I don’t do foalcon, though). You can also switch to 11-20 for a Latte-like experience with even fewer filters. I agree that Gemini needs to step up their game. The full Gemini 2 model is supposed to release this month. Deepseek is a shame because it seems to have the knowledge and smarts to pull it off, but the other issues are too big. Maybe the repetition would be less noticeable in a dialogue-heavy VN.I don't have that many hopes but Grok 3 soon too.
>>41808904I don't use agnai after migrating to Risu but it should work if you download the card
>>41808909>type something random in the proxy password fieldah, that's what I was forgetting to do. will try again once my current copilot gets banned
>>41808917Anon...
>>41806714Shit alright, neat. I'm still stuck on Chary at the moment so Latte isn't an option to me until I actually bother to dig around for other proxies or paypig or whatever, but I'll take note of all of this. Thanks Anon!
>>41809095Chary has GPT-4o 11-20. It's very close to Latte.
>>41809095Also >>41808353
>>41809103Oh fuckin hell, I swear a while ago they only had 05-13. Thanks, Anon. I didn't even bother checking because I just assumed it wasn't there.>>41809105And oh shit, nice. Missed this general and how helpful everyone is here. Fucking love /mlp/, always brings me joy to revisit it even 10+ years later. Thanks bro.
>>41808452 It allows prefill, so no issues with NSFW. Outputs seem different from Copilot, but it’s clearly 3.5, maybe there’s some prompt/instruction injection at play.
>>41797144>MM rentry updated and now contains a new linkWhat it do?
>>41809103Is it better than latest on MM?
>>41809432 11-20 is on MM too, so you can try it if you want. I think I prefer Chorbo overall, but they’re close, and 11-20 seems less filtered.
>>41809431 The last link is for donations if you’re interested. The third link is an online interface for image generation.
>>41809598New AMD GPU?
I've just learned that it's possible to have folders in ST. But it's a mess.
>>41808631So it's just the context with my greentext? How did you get it to react to filters/jailbreaks?
>>41808796>Why in the ACTUAL FUCK are our cards not on par with the sheer autism these fuckers have put in?for the same reason /v/ has made a game and /mlp/ hasnt made shit. this board is for consoomers who can't collaborate
>>418091031120 IS latte, nigger.
>>41809706Why bother? the built in booru tag system already would do the same thing
>>41809803/mlp/ has made games tho
>>41808796>>41808842>it's all ai slop generations and inpaints for 'emotions'.Even if it works for VNs I'm not impressed at this so called autism.
>>41809812 It’s a snapshot. They update Latest continuously.
>>41809828Kind of a silly thing to say when the whole thing is AI.
>>41809836Not really. this is their 'process' You could do this all in a day if you cared, it's not some magnum opus.
>>41809803I mean you're not exactly wrong. However the best shit I've seen posted to /acig/ was that Tomoyo card and I'm pretty sure that was just some fag's autism project.It's actually kinda nuts how it's been more then a year and its still the highest effort card I've seen.
>>41809860The only thing impressive is the square peg in circular hole of making the LLM keep track of whats happening with sprites+backgrounds+music like a VN, but the sprite work itself is nothing special. He straight lifted the backgrounds from xhentai and isn't hiding the fact. The 'new animu vns' just use ai slop prompts for backgrounds too
>>41809874Tying together the music, the backgrounds, genning the images and inpainting to get it all right. It's leagues ahead of "retired egirl works for you" yuricard 9000 that we see today let alone anything in 2023 when Tomoyo was released.If you got more soulful cards hiding about, I'd love to see them.
>>41809894There are already cards that have the VN stuff but with multiple endings. As well as cards that have entire text game mechanics with their own unique huds that keep track of everything with unique game overs, ect if conditions are met with game trigger events, and modifiers of game rules during. not the absolute slop of "ponies present:"
>>41809919I've seen the memes that turn LLMs into emulators. There was that RPG JB that was pretty neat. What card has its own unique hud?
Dumb question but what's the correct url ending for charys google key?
>>41809954You can stay with https://chary.scylla.wtf/proxy, it work with all models they have. Just config like that.
>>41810040Oh, that's neat. Thanks anon.
>>41810294Order up!
NEW BREAD>>41810311>>41810311>>41810311
