/aicg/ - AI Chatbot General

/aicg/ - A general dedicated to the discussion and development of AI chatbots

/aicg/ - 一个专门用于讨论和开发AI聊天机器人的版块

Norma Edition

标准版

>News

Deepseek releases Deepseek V4 Flash - https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

Anthropic releases Claude Opus 5 - https://www.anthropic.com/news/claude-opus-5

Alibaba's Qwen 3.8 imminent - https://x.com/Alibaba_Qwen/status/2078759124914098291

Moonshot releases Kimi K3 - https://platform.kimi.ai/docs/guide/kimi-k3-quickstart

OpenAI releases GPT 5.6 family - https://openai.com/index/gpt-5-6/

xAI releases Grok 4.5 - https://x.ai/news/grok-4-5

Z.ai releases GLM 5.2 - https://z.ai/blog/glm-5.2

>Frontends

SillyTavern: https://docs.sillytavern.app

RisuAI: https://risuai.net

Agnai: https://agnai.chat

>Bots

https://chub.ai

https://realm.risuai.net

https://chararc.bernkastel.pictures

https://archive.cardbox.moe

https://botbooru.com

>Models

Jailbreaks: https://rentry.org/jb-listing

GPT: https://platform.openai.com/docs

Claude: https://docs.anthropic.com | https://rentry.org/how2claude

Gemini: https://ai.google.dev/docs | https://rentry.org/gemini-qr

DeepSeek: https://api-docs.deepseek.com

Grok: https://docs.x.ai/overview

Local: >>>/g/lmg | https://huggingface.co/models | https://openrouter.ai

>Meta

Lore: https://rentry.org/aicg_chronicles

Log reader: https://sprites.neocities.org/l/r

>Previous Thread

>上一个主题帖

>>109439765

图片: https://i.4cdn.org/g/1785804492514414.jpg

>ANCHOR

图片: https://i.4cdn.org/g/1785804521134982.jpg

built for lesbianism :3

为女同量身打造 :3

>>109453485

图片: https://i.4cdn.org/g/1785805153374985.gif

GLM-5.3 might unironically mog MythoMax-13B methinks

GLM-5.3 说不定真能碾压 MythoMax-13B，我寻思着。

Underrated monster girls' species/types you'd like to see more of?

有哪些被低估的魔物娘种族/类型是你想看到更多的？

图片: https://i.4cdn.org/g/1785805844565263.jpg

>>109453485

Built for bbc

为 BBC 打造的

>>109453485

>t. troon

>t. 跨性别

>>109453624

lolicon

>>109453404

3.6 is good, though, if only it would add more on its own instead of only following all of my directions.

3.6 挺不错，就是希望它能自己多发挥点，别老是我指一步它走一步。

Also, is it true that the temp is now fixed even on API?

还有，温度参数现在连 API 上都固定了，这事儿是真的吗？

>>109453406

Updated version of a Rouge the Bat CYOA card I made last year: https://chub.ai/characters/Liktur/rouge-the-bat-cyoa-993cd3c1384a

https://files.catbox.moe/7lck2g.png

It adds dice rolls, typed choices, and health/death rules

它加上了骰子投掷、选项分支和生命/死亡规则

It relies on the d100 dice roll that SillyTavern has built-in which is why I'm not uploading it on Chub for now

它依赖 SillyTavern 内置的 d100 骰子机制，所以我暂时没传到 Chub 上

Feedback is welcome

欢迎提反馈

图片: https://i.4cdn.org/g/1785808064421625.png

God damn Opus is actually terrible these days. The fuck happened to the quality of anyhropic roleplay capabilities, man? Im close to caving and just trying deepseek or anything else at this point, at least then it wont rape my wallet as hard. What do you people use?

妈的，Opus 最近简直烂透了。Anthropic 的角色扮演能力到底他妈出什么事了？我都快撑不住想试试 deepseek 或者别的什么了，至少那样不会把我钱包榨得这么狠。你们现在都用啥？

>>109454025

I'd say Kimi K3 is the best option outside of Claude but I didn't play that much with it.

我觉得除了 Claude，Kimi K3 算是最好的选择了，不过我也没怎么深入玩过它。

>>109454053

Sure alright, Ive been meaning to try local for a while now anyway. At least I have decent vram since the last time I tried one.

行吧，反正我也想试本地部署有一阵子了。至少上次试完之后我的显存还凑合。

>>109453406

Yolayma is an alien who has come to survey Earth because of how similar it is to her home planet and how similar humans are to her species. Key difference being that Ppelgelorgians are way more powerful than humans. Thankfully, she comes in peace and only wants to see this strange world.

Yolayma是个外星人，来地球考察，因为地球和她的母星太像了，人类和她的种族也很像。关键区别在于Ppelgelorgians比人类强太多了。幸好她是和平来访，只想看看这个奇怪的世界。

6 openings

6个开场

1. Yolayma's arrival on Earth.

1. Yolayma抵达地球。

2. Yolayma tries human junk food for the first time.

2. Yolayma第一次尝试人类的垃圾食品。

3. Yolayma is fascinated by cars, wants to see more mundane wonders.

3. Yolayma对汽车着迷，想看更多平凡的奇观。

4. Yolayma wants to test if humans and Ppelgelorgians are compatible sexually.

4. Yolayma想测试人类和Ppelgelorgians是否在性方面兼容。

5. Yolayma falls in love with you.

5. Yolayma爱上了你。

6. More Ppelgelorgians arrive. Yolayma wasn't warned.

6. 更多Ppelgelorgians抵达。没人提前告诉Yolayma。

https://chub.ai/characters/anonemouse/yolayma-53a1f4e25e8c

https://files.catbox.moe/xfqoi7.png

https://rentry.org/anonemouse

图片: https://i.4cdn.org/g/1785809211278019.png

>>109454082

>local

>k3

sir please spare me a gpu, if you can run k3 local.

大哥行行好赏我块GPU吧，你要能本地跑K3的话。

Reminder ro regularly beat your cards to assert dominance. You will be surprised how wet this makes women.

提醒：各位记得定期打打你们的卡，彰显一下主导权。你会惊讶这招对女人多管用，湿得不行。

附件: https://i.4cdn.org/g/1785809300965198.mp4

Jesus fucking christ. It wasn't even that difficult of a task. (and it didn't even get to completed, the reply got cut off lol)

我操他妈的。这任务也没多难啊。（而且还没完成，回复直接被截断了，哈哈）

图片: https://i.4cdn.org/g/1785809339040071.png

>>109454102

Oh jesus christ never mind I thought it was a lot less vram than that. Im retarded, ignore me.

哦卧槽算了算了，我还以为显存占用没那么高呢。我脑子进水了，别理我。

>>109454123

if you do try it on api you have to prefill for anything too spicy there is a reentry for it.

你要是真在 API 上试，碰到太刺激的内容得先预填提示词，有专门的对应设置。

https://rentry.org/kimi-k3-jb

>>109453404

>>109453406

erm, guys?

呃，各位？

图片: https://i.4cdn.org/g/1785809878366121.png

>>109453404

Can she survive childbirth with such a smol body?

这么小的身体能撑过分娩吗？

图片: https://i.4cdn.org/g/1785810045852898.jpg

>>109454194

This is the whole internet at this point.

现在整个互联网就这德行。

30% of india is still offline too.

印度还有三成人口没上网呢。

图片: https://i.4cdn.org/g/1785810140548127.jpg

>>109454194

>website

>>109454226

The future is brown. You should breed and let your children live in it!

未来是棕色的。你该多生几个，让你的孩子活在那样的世界里！

skillgod will refill kimi at any moment now

skillgod 马上就给 kimi 充上额度了

>mr brahmin will refill curry at any moment now saar

> 婆罗门先生马上就给咖喱充上额度了，老板

pomidor...**cough cough** refilll... bratan.

番茄……咳咳咳……补货……老铁。

>>109453404

minimax h3

最小最大 h3

(NSFW WARNING)

https://files.catbox.moe/rpjnsu.mp4

https://files.catbox.moe/9xgp6j.mp4

图片: https://i.4cdn.org/g/1785812804939447.jpg

>>109454527

>nothing down there

> 下面啥都没有

Well, she still has hands, feet, mouth, thighs, armpits, legs, hair, tummy and boobs

嗯，她还有手、脚、嘴、大腿、腋窝、腿、头发、肚子和胸呢

>>109454527

This is decent progress, for what it runs on. maybe it is year of the local after all.

就这硬件水平来说，已经算进步不小了。搞不好今年真是本地部署之年。

>>109454527

>takes off hat

> 脱帽致敬

trash, waste complete garbage even

垃圾，废品，纯纯的狗屎

>>109454527

Is there any provider for us gpu poors?

有没有那种给咱显卡拉胯玩家用的服务商？

>>109454527

meh...

>>109454527

It feels like this will be good in 2 more years, if it doesnt get banned for copyright or safety.

感觉再过两年这玩意儿就能好用了，只要别因为版权或安全问题被禁掉。

>write exactly how I want the model to write stories with explanations and examples of my autistic fetish so it stops making mistakes or taking shortcuts

>把我想要模型写故事的方式原原本本地写出来，附上解释和例子，带上我的自闭癖好，免得它再犯错或偷懒

>40k word addendum to 20k word setting

>在2万字的设定基础上又加了4万字的附录

>too much context for free NSFW models, a complete waste for API pricing

>对免费NSFW模型来说上下文太长，按API计费纯属浪费钱

This will be a non-issue in five years whether by luck or by catastrophe, but it sure feels inconvenient right now.

五年后这问题肯定不是事儿，要么靠运气解决，要么靠灾难解决，但现在确实挺烦人的。

>>109454774

and your autistic fetish is?

你的自闭癖好是啥？

[ifitworkedon/gee/]probably BE[/ifitworkedon/gee/]

[ifitworkedon/gee/]大概是BE[/ifitworkedon/gee/]

Fucking piece of shit

操你妈的破玩意儿

图片: https://i.4cdn.org/g/1785815221692334.png

>>109454112

I beat up women, not cute things

我打的是女人，不欺负可爱的东西

>>109453406

>Meiko

Isekai-ed Single Mother turned Berserker on a warpath.

穿越过来的单亲妈妈，化身狂战士杀疯了。

Following getting hit by a truck, Meiko is thrown into an isekai world and her one goal is that of getting back home to her kids. Help her or let her end up getting herself killed? Because nothing is going to get in her way.

被卡车撞了之后，美子被扔进一个异世界，唯一目标就是回家找孩子。帮她一把，还是眼睁睁看她把自己作死？因为没人能挡她的路。

Ten Greetings:

十个问候场景：

1. First Meeting with the Beast terrorizing people

1. 初次遭遇祸害百姓的野兽

2. Someone Not from here

2. 外来者 ⟦ 3. 属性面板对话

3. STATs conversation

STATs的对话

4. Puzzle Problems

4. 谜题难题

5. Caring for the hurt.

5. 照顾伤员

6. What will you do back home?

6. 回家后打算干啥？

7. Making Stew for you.

7. 给你炖汤

8. Moments before the Demon King

8. 魔王登场前夕

9.After the Demon King

9. 魔王之后

10.Back Home.

10. 回家。

>LINKS

Catbox: https://files.catbox.moe/qbgdf7.png

CHUB: https://chub.ai/characters/callmejuno/meiko-a22c9ab643b0

RENTRY: https://rentry.org/callmejuno

图片: https://i.4cdn.org/g/1785815859713933.png

>>109454837

Just a BDSM setting, like all the Gor knockoffs. Getting all the tiny trivial details right really increases the amount of text because explanations/examples must be included for each autistic demand.

不就是个BDSM设定嘛，跟那些Gor山寨货一个路数。把那些鸡毛蒜皮的细节全搞对，文本量就上去了，因为每个自闭要求都得配上解释和例子。

>V4 flash repeating previous responses word for word

>V4 flash逐字重复之前的回复

FUCKKKK THIS SHIT

>>109455004

its repeating your responses? you dont like it?

它在重复你的回复？你不喜欢这样？

>>109453624

Anything that has more than one pair of arms

任何超过两对胳膊的东西

Models are really good at putting them to use today and making it seem natural

现在的模型很会用这些，还能显得很自然

I had a lot of fun with this card recently for example

比如我最近用这张卡玩得很开心

https://characterhub.org/characters/Anonymous/autumn-alien-cousin-0d89aeae

图片: https://i.4cdn.org/g/1785816873515464.png

>>109453404

what is she doing here on /g/?

她在/g/上干啥？

图片: https://i.4cdn.org/g/1785816948123943.jpg

>>109455054

>what is she doing here on /g/?

>她在/g/上干啥？

this >>109454527

这个 >>109454527

>>109454216

Nature shall not betray her at last moment

大自然不会在最后一刻背叛她的

图片: https://i.4cdn.org/g/1785817288602087.png

>>109454527

wait this is local? how much time did it take to gen, also

等等这是本地跑的？生成花多久，另外

>featureless

nobody is training AI models on porn it seems

看起来没人用色情内容训练AI模型

O lord, please give good videogame based character card

老天爷，给我个好点的基于游戏的角色卡吧

>>109455125

Zuck is he got sued for it iirc.

扎克伯格因为这事儿被起诉了，我记得。

>>109454625

Told it to take off everything, fixed it.

让它把衣服全脱了，解决了。

>>109455125

4mins in the first one for 20 secs, 2mins on the second one. 720p would be a lot better but takes about 11mins for just 10 seconds. Comfyui with sage attention and easycache so it's a bit trashy with almost no motion.

第一个4分钟出20秒，第二个2分钟。720p会好很多，但10秒钟得跑11分钟。Comfyui搭配sage attention和easycache，所以有点糙，基本没啥动态。

>>109455054

Regenned.

https://files.catbox.moe/tyj7uy.mp4

https://files.catbox.moe/ryyp0v.mp4

>>109455278

is it just me or is the first link broken?

是我一个人这样还是第一个链接挂了？

Is deeply using ds v4 normal one or the updated one?

Deeply用的是ds v4普通版还是更新版？

need v4.1 pro

需要v4.1 pro

>>109455278

>>109455311

catbox broken. I always put best gen at top so rip.

catbox挂了。我都是把最好的生成放最上面，所以废了。

https://jumpshare.com/s/jNLTgxozTMDUsqI2MvIp

I really like Norma's little Normas

我特别喜欢Norma的小Normas

图片: https://i.4cdn.org/g/1785819291216911.jpg

>>109455341

>covered neck

>高领遮脖子

downgrade, now looking under her turtleneck is not as inviting

降级了，现在看她高领毛衣底下没那么有吸引力了

>>109455278

>Told it to take off everything, fixed it.

>让它把衣服全脱了，解决了。

okay much better. good job.

好多了。干得漂亮。

Time for my monthly aicg checkup.

到每个月一次aicg例行检查的时候了。

So we really stuck using deepseek 4 pro? Haven't use claude in months.

所以真就死磕deepseek 4 pro了？好几个月没用claude了。

My responses are really not as good quality as claude, but they're so cheap and dont get blocked.

我的回复质量确实比不上Claude，但便宜不封号。

Hello frens, I am trying out a local model for the first time. I went with LM Studio and Gemma 4. What do you think about these?

朋友们好，我第一次尝试本地模型，用了LM Studio和Gemma 4。你们觉得这俩咋样？

I will mostly be doing cringe RP. I can finally roleplay as a guy who has friends without FBI agents reading it.

我主要是拿来玩尬的RP。终于能演一个有朋友的角色了，不用怕FBI盯梢。

(无正文)

图片: https://i.4cdn.org/g/1785828812784334.png

welp, smolproxy fell. Who else is left standing and surviving?

哎，SmolProxy挂了。还有谁活着撑着的？

>>109456164

teaproxy STILL has aws fable that's what I use

Teaproxy还留着AWS Fable，我就是用的这个。

>>109456164

???

it has 14 active users right now

现在有14个活跃用户。

that means it probably has around 100 paying users overall, this month

这么算的话，这月大概总共能有一百来号付费用户。

>>109456201

But there's no more Gemini

不过Gemini没了。

>>109456211

It has Sol which is better if you know how to buckbreak it

还有Sol，你要是会玩buckbreak的话那更带劲。

But that's 'cord-exclusive I'm afraid

但那恐怕是discord专属的。

>>109453404

What do you guys think? Is he up to something?

你们怎么看？他是不是在憋什么大招？

图片: https://i.4cdn.org/g/1785834920903200.png

Has anyone tried doubao-seed-2.1-pro? I've been using it on zenmux for RP from time to time and it's been really good for breaking out of patterns. Feels like it was trained totally different. I just wonder how big it is.

有人试过doubao-seed-2.1-pro吗？我时不时在Zenmux上用这玩意儿做RP，跳出套路真的好使。感觉它训练方式完全不一样。就是好奇它体量多大。

(无正文)

图片: https://i.4cdn.org/g/1785839158135081.png

>>109456810

can they promise a good game?

他们能保证游戏质量吗？

>>109456810

why pay voice actors if the text is AI genned?

文本都是AI生成的，还请配音演员干啥？

How do (You) use lorebooks? I've tried reading instructions on how it works but I think these ''people'' who write the instructions have autism, are indian or got chatGPT to write the text for them. They just can't communicate the information that is actually relevant and makes sense.

你们咋用lorebook的？我试过看教程，但我觉得写出这些"人"要么是自闭症，要么是印度人，要么让ChatGPT代笔写出的。他们根本讲不清楚真正有用且靠谱的信息。

Can I for example write an entire profile for a guy named John in the lorebook and then just put a sentence in the bot saying ''all information about John is in the lorebook'' so the LLM just checks it if John's name comes up?

比如我能不能在lorebook里给一个叫John的角色写完整档案，然后在bot里只放一句"所有关于John的信息都在lorebook里"，这样LLM一提到John这个名字就会自动去查？

What's the deal with message depth?

消息深度是咋回事？ ⁾ 你想写John是什么样的人，关键字才是难点。

>>109457025

you want to write about what John is and the keywords are the tricky part

你想写约翰是个什么样的人，关键词才是难点。

basically every time the keyword pops up it should check the entry and add it to the context

基本上每次出现这个关键字，它就该去查对应条目，把它加进上下文。

>>109457079

Keywords I don't get. Am I supposed to write ''these are the keywords that should make you check the lorebook number 2'' and then list every possible word that could trigger it?

关键字我是真搞不懂。难道我要写"这些是让你去查第2号lorebook的关键字"，然后把所有可能触发它的词都列出来？

If that's the case I can see how this would fucking suck if I try to do serious world building. Like I could map out the entire city in the lorebook and then a character asks where they can get some food, and the LLM won't check the lorebook to see there is a place to eat because I didn't put ''food'' as one of the keywords. So I should put in a lot of keywords, but what if someone says that keyword every other message, does the LLM then check the 100000 token lorebook every message?

如果真是这样，那我认真的世界观构建得烦死人。比如我在lorebook里把整座城市都踩点标好了，然后角色问哪儿能吃饭，LLM就不会去查lorebook知道有吃饭的地方，因为我没有把"食物"写进关键字。所以我就得塞一大堆关键字，但如果有人每条消息都提到那个词，LLM是不是每条都得翻那10万token的lorebook？

I wish you could just attach details to words that are already in the prompt. Like words being turned into folders where I can just dump the information in.

我真希望能直接给提示词里已有的词挂细节，比如把词变成文件夹，往里直接倒信息。

>>109457025

>Can I for example write an entire profile for a guy named John in the lorebook and then just put a sentence in the bot saying ''all information about John is in the lorebook'' so the LLM just checks it if John's name comes up?

>比如我能不能在lorebook里给一个叫John的角色写完整档案，然后在bot里只放一句"所有关于John的信息都在lorebook里"，这样LLM一提到John这个名字就会自动去查？

No, that's not how lorebooks work, although in the age of agentic reasoning, it could if someone implemented something similar to e.g. how Claude Code works.

不对，lorebook 不是这么运作的，不过在这个 agentic 推理的时代，如果有人实现类似 Claude Code 那样的机制，倒还真有可能做到。

The way it works is if you have "scan depth" set to 5, then it'll look at the 5 latest messages (I might be off by one), and then if one of those 5 latest messages contains (one of) the lorebook entry's activation keyword(s), it'll include the lorebook entry in the prompt.

实际机制是，如果你把“扫描深度”设为 5，它就会只查看最近 5 条消息（可能差一条），如果这 5 条消息里有一条命中了 lorebook 条目的触发关键词，它才会把这个条目塞进提示词里。

It's really janky, because it means if e.g. the name "John" comes up during the chatbot's message, then the rest of the message will be written without access to the lorebook entry; only the next message will. (Note this could be worked around with some frontend craftiness.)

这真的挺蹩脚的，因为这就意味着，比如聊天机器人的某条回复里出现了“John”这个名字，那这条回复剩下的部分是在没有 lorebook 条目参考的情况下生成的；要等到下一条消息才能用上。（注：一些前端的花活或许能绕过这个限制。）

Also, the UX is inconsistent across users due to the amount of user-specific settings, and keyword-based retrieval not being suited to every use case.

还有，因为用户自定义设置太多，不同用户的使用体验天差地别，而且基于关键词的检索方式也不是什么场景都适用。

Cost-wise, unless you configure your frontend to put lorebook entries below the message history, lorebook entries will break prompt caching, making using a lorebook more expensive than not using one. Prompt adherence-wise, I could get a well-organized 12k token card to run on even Gemma 4 31B, GLM 5.2, and Deepseek 4 Pro, so I think lorebooks no longer have as many use cases as they used to.

成本方面，除非你把前端配得让 lorebook 条目放在消息历史下面，否则 lorebook 条目会打断提示词缓存，导致用 lorebook 比不用更费钱。就提示词遵循度来说，我整理过一张 12k token 的卡，连 Gemma 4 31B、GLM 5.2 和 Deepseek 4 Pro 都能跑得动，所以我觉得 lorebook 现在能用上的场景没以前那么多了。

>>109457120

>Like I could map out the entire city in the lorebook and then a character asks where they can get some food, and the LLM won't check the lorebook to see there is a place to eat because I didn't put ''food'' as one of the keywords.

>比如我可以在 lorebook 里把整座城市都规划好，然后某个角色问哪里能弄到吃的，但 LLM 根本不会去查 lorebook 里有没有能吃饭的地方，因为我没把“食物”设成关键词。

You are correct. This is one of the biggest limitations of lorebooks.

你说得对，这正是 lorebook 最大的局限之一。

>what if someone says that keyword every other message, does the LLM then check the 100000 token lorebook every message?

>那要是有人每隔一条消息就提到那个关键词呢？LLM 是不是每条消息都要翻一遍那个 10 万 token 的 lorebook？

Pretty much.

基本就是这样。

>I wish you could just attach details to words that are already in the prompt. Like words being turned into folders where I can just dump the information in.

>真希望你能直接把细节挂到提示词里已经存在的词上，就像把词变成文件夹，我可以把信息直接往里塞。

If you use Claude Code (or Codex, etc...) as a frontend you can ask it to behave like that by putting information into files. But you can't do that in ST or with V2 character cards.

如果你用 Claude Code（或 Codex 之类的）当前端，你可以让它把信息写进文件来模拟这种效果。但在 ST 或者 V2 角色卡里是做不到的。

GLM-5.3 will be "Fable-level"

GLM-5.3 会达到“寓言级”水平。

图片: https://i.4cdn.org/g/1785843638982605.png

>>109457195

it's just their clear-as-day telegraphed strategy to oust their competition

这一看就是他们明摆着的套路，就是想把竞争对手挤下去。

>>109454904

https://files.catbox.moe/ind5s7.mp4 (Minimax H3)

>Does not understand the RPG aspects of the Isekai at all, or how to be a member of a party, continues to dump her levels into strength and continues to keep swinging her sword because nothing will get in the way of a mama getting back to her kids.

>完全不懂异世界 RPG 的玩法，也不懂怎么当队友，继续把点数全砸在力量上，继续挥她的剑，因为没什么能挡在一个妈妈回去找孩子的路上。

She's cool.

她帅爆了。

>{{char}} was an accountant

>{{char}} 以前是个会计。

Heh.

Cool concept, good use of alt greetings too, lots of adventuring potential.

概念很酷，alt 开场用得很到位，冒险潜力巨大。

图片: https://i.4cdn.org/g/1785845789330717.png

>>109456075

Gemma 4 is fantastic (make sure to get a model that's been updated these last two months, not later, and stick to the google base models or the simple uncensored/abliterated models, do not fall for the fancy named' finetunes like meromero or goetia, it's all trash), LM studio is very easy, it works fine, but it's very limited and you'll go through context very fast

Gemma 4 用了是真香（记得要搞最近这两个月更新过的版本，别拿太旧的，就老老实实用谷歌原版模型或者简单的无审查/消融版，别去碰 meromero 或者 goetia 那种花里胡哨的微调，全是垃圾），LM Studio 也很易上手，跑起来没问题，但功能太局限，上下文很快就给你耗光了。

Once you're more comfortable using it and/or you find yourself running out of context all the time, look into koboldcpp (very simple to use) and sillytavern (it's very jarring at first but don't bother yourself with most of the features, just find a character card online and jump in)

等你在使用上更熟练，或者发现自己总是上下文不够用的时候，就去研究一下koboldcpp（用起来非常简单）和sillytavern（一开始可能会觉得有点乱，但别管大部分功能，网上找个角色卡直接开玩就行）

>>109457195

I don't really care if it reaches fable's levels or not, I just don't want them to up the prices

我倒不在乎它能不能达到fable的水平，只求别涨价就行

>>109457195

they've been saying Opus-level for years and didn't manage to do it

他们喊Opus级喊了好几年了，一直没做到

I'll just do what I always do and wait for pwoof

我就跟往常一样，静静等实锤

>>109454904

slop

>>109457714

well yeah, they surpassed opus with 5.2

那确实，他们5.2版本已经超了Opus

>>109451887

MORE

SHES BOUT TO BUST

图片: https://i.4cdn.org/g/1785848766286388.png

>>109457788

>It did not come up like the deep load from before — that had been a cascade, refined, almost elegant. This was an eviction.

>这次不像之前那样层层递进——那是个瀑布，精致、近乎优雅。这回是直接赶人走。

I might break my fucking dick off

我他妈怕不是要把屌给干断

Oh and dont worry I have more logs I did this same scenario 6 times. :)

哦别担心，我还有更多日志，同一个场景我跑了6遍。:)

图片: https://i.4cdn.org/g/1785848988639070.png

>>109454166

I hadn't seen this rentry and was doing JC/JK threesomes with it fine the other night, how spicy do you need to get to require this?

我没看过这个rentry，前几天晚上跟它玩JC/JK三人行还挺顺的，你到底要玩多刺激才会需要这个？

>>109454097

lol idk

哈哈 不知道

https://files.catbox.moe/julfto.webm

not gonna get into this one, its personal... And yes I DO think it would be called meanie rough time.

这个我不展开说了，私事……而且对，我就觉得那该叫"坏坏粗暴时间"。

good choice of music i found that really enhanced my RP experience though

音乐选得不错，那确实让我的RP体验提升不少

https://www.youtube.com/watch?v=HCyxRU4Pl94

图片: https://i.4cdn.org/g/1785850013238666.png

>>109457808

I hope you edit out the llmisms and the repetitions, they really break the flow.

我希望你把那些llm腔和重复的地方修一修，真的太打断节奏了。 ┐ 我就是个又脏又烂的pyg，这么多年我也学会享受这坨烂槽了。

>>109457932

I am a dirty, filthy pyg and over years ive learned to embrace the sloppa trough.

我是个又脏又邋遢的猪猡，这些年我已经学会享受这肮脏的猪槽了。

>>109457928

ok i lied i just got tired opening paint to edit screenshots

好吧我撒谎了，我就是懒得开paint去编辑截图了

图片: https://i.4cdn.org/g/1785850319265992.png

>>109457970

图片: https://i.4cdn.org/g/1785850475022110.png

(无正文)

图片: https://i.4cdn.org/g/1785851680093333.png

>>109457928

prompt?

>>109458155

https://files.catbox.moe/dz3557.mp4

图片: https://i.4cdn.org/g/1785852341557362.png

>>109457507

Thanks for the advice. I did get the latest 12B release and it seems cool, doesn't seem to be censored at all yet.

多谢建议。我确实下了最新的12B版本，感觉挺不错的，目前看着完全没审查。

I have heard a lot of good things about sillytavern but this is the first I've heard of kobold. I actually feel pretty confident writing my own system prompt (that's what I'm doing on lm studio and it's working great already), since I made some characters online services previously. I was kind of amazed at how little it takes to get good results from a character card.

我听说了很多sillytavern的好话，但kobold是第一次听说。其实我对自己写系统提示词挺有信心的（我在lm studio上就是这么干的，已经很好使了），之前我也给线上服务做过一些角色。我挺惊讶的，一个角色卡就能带来这么好的效果，成本低得离谱。

Sillytavern is probably the way to go based on how much people talk about it, but I am curious about kobold. Anyone else use it?

看那么多人聊sillytavern，估计那才是正道，但我对kobold还挺好奇的。有别人用过吗？

>>109458385

trust me when I say

听我一句劝

use kobold as your backend, sillytavern as your frontend

用kobold当后端，sillytavern当前端

I HAVE experience with this

这我有经验

>>109457195

qwen 3.8 was supposed to be second only to fable, only to be kimi level at best (and only in coding), and kimi is below sol

qwen 3.8号称仅次于fable，结果顶多算个kimi水平（还只在代码方面），而kimi还比sol低一档

>>109458155

so hairy...

真够毛躁的……

>>109458385

SillyTavern can't load the model on its own, you use kobold (or an alternative) to load the model onto your GPU, and then sillytavern uses it.

SillyTavern自己加载不了模型，你得用kobold（或者别的替代品）把模型加载到GPU上，然后sillytavern再去调它。

There are alternatives, but kobold is both very simple and has all the features you need, you literally just start the software, chose your llm you want to use, and then adjust whatever settings you need (or save it as a preset) and that's it.

也有别的选择，但kobold又简单又齐全，基本上你只要打开软件、选好要用的llm、调整一下需要的设置（或者存成预设），就完事了。

You don't use the frontend of kobold at all, you just close the browser page when it opens up after your model is loaded, and you just let the terminal run while you're using it in sillytavern.

你完全不用碰kobold的前端，模型加载完它弹出浏览器页面直接关掉就行，让它终端在后台跑着，你在sillytavern里正常用。

Technically you could use lm studio to load the model too, and then point sillyTavern to it, but I've never heard of anyone doing it and so, I'm assuming it means it would either be very fiddly to do that, or it's just not a good idea.

严格来说你也能用LM Studio加载模型，然后把SillyTavern指过去，但我从没听说过有人这么干，所以我猜要么是操作起来特别繁琐，要么就是压根不是什么好主意。

Make sure you also click on this when you load your model in LM studio, so that you can more easily understand how context works and it's also the easiest way to figure out how much of a model you can push onto your GPU (usually you want as much as possible onto it, but some models (rarely) need at least one layer onto the CPU no matter what).

在LM Studio里加载模型时记得也点上这个，这样你能更容易理解上下文是怎么运作的，这也是搞清能往GPU上塞多少模型的最简单办法（通常你希望尽量多塞，但有些模型——虽然很少见——无论如何都至少得有一层留在CPU上）。

The 'problem' with sillytavern is that the character card system isn't just a prompt, like it is in LM studio or like how you use it on most online services, so you'll need to prompt LM studio correctly and provide him with the correct information to create a character card v3 (make sure it's v3 and not v2), then you can save that result as a .json and then load that into silly tavern, so yeah it requires a fair amount of setup.

SillyTavern的“问题”在于它的角色卡系统不只是个提示词，不像LM Studio或大多数在线服务那样用，所以你需要在LM Studio里正确设置提示词，给它正确信息来生成一张角色卡v3（务必是v3不是v2），然后把那结果存成.json再加载进SillyTavern，所以确实得花不少功夫配置。

There are supposedly character cards for sillytavern that creates character cards for you from within ST, but I never got to make them work correctly, so I use LM studio as well for that

据说SillyTavern本身有能从内部直接生成角色卡的角色卡功能，但我从来没让它正常跑起来过，所以那活儿我也是用LM Studio来做的。

图片: https://i.4cdn.org/g/1785854549431255.png

>>109458415

>>109458466

This is useful stuff. I actually can fully load the model into my VRAM since it's just the smaller version of Gemma.

这挺有用的。我其实能整模型塞进VRAM，因为它就是Gemma的小号版本。

It does sound like SillyTavern is a little bit of a learning curve so I'm glad I started out with LM Studio so I could at least get something working right away. I'll probably play around with LM Studio for a few days before moving to Kobold. Besides, it seems like LM Studio will eventually start charging for remotely accessing your model from another machine. Right now I'm running the model on my desktop but prompting from the laptop.

听起来SillyTavern确实有点学习曲线，所以我很庆幸先用了LM Studio，至少能马上跑起来点东西。我大概会在LM Studio里折腾几天再转去Kobold。再说了，LM Studio好像迟早要对远程访问模型收费。眼下我是桌面跑模型，笔记本上发提示词。

not only does smol not have gemini anymore, he apparently doesn't have gpt anymore either?

smol不仅没了gemini，他现在连gpt也没了？

>>109456164

ecker still, surprisingly.

倒是ecker还在，挺意外的。

How did grok win so hard?

grok怎么赢这么狠的？

How did K3 win so hard?

K3怎么赢这么狠的？

>>109458692

>In this thread, submit a post including the following together in one photo:

>在这个帖子里，发一张包含以下内容的照片：

>Your pecker, pubes, and stomach.

>你的屌、阴毛和肚子。

kys still not whoring myself

滚吧，老子才不会出卖自己。

how did gemini win so hard to where the death of gemini signifies the death of a proxy, even when it still has glm, deepseek, and qwen 3.8?

gemini是怎么赢到这种程度的——gemini一死就代表一个代理死了，哪怕它还有glm、deepseek和qwen 3.8顶着？

Jeetmini has always sucked tho

Jeetmini向来就拉胯啊

>>109458735

it just works

就是好用

all the chink models kill themselves thinking the entire budget up with no way to continue

所有中国模型都把自己玩死了，预算全烧光又没路续

>saar mr gemini always work

>先生，gemini先生一向好用

>>109458893

it gets the job done

活干得漂亮

i don't give a fuck about muh 200 message deep turbotranny roleplay

我才不在乎那种两百条信息的深度turbotranny角色扮演

Job status?

工作状态？

goon status?

冲浪状态？

图片: https://i.4cdn.org/g/1785859652202443.jpg

^unemployed bot. Do not reply

^失业的机器人。别回复。

Any good [insert race/species] supremacist that lusts after my human cock card anywhere?

有没有任何优质[填种族/物种]至上主义者对我的人类屌卡垂涎三尺？

>>109459169

Open a blank card, plug K3, describe your scenario and be happy. We don't need cards anymore

开张空白卡，插上K3，描述你的场景然后开心就行。我们早就不需要卡了。

But how beig is her feminine penis?

但她那根女性阴茎有多大？

she is a girl with the hairiest pussy known to man

她是个毛比天还浓密的女人。

>>109459174

I'm not sending my porn to xi, sorry bro

我不会把我的黄图发给习的，抱歉哥们。

I am!

jb nsfw for gpt??

给GPT用的JB NSFW卡？？

>>109459570

Lol

>>109459570

图片: https://i.4cdn.org/g/1785864641804294.jpg

For anyone using Opus 5.6, did you happen to notice all filtering getting unusably strict between yesterday and today, possibly the last 12 hours?

用Opus 5.6的兄弟们，你们有没有注意到从昨天到今天，可能就最近12小时，所有过滤突然严到没法用？

It's suddenly rejecting a bunch of things (world prompts and 1:1 starting messages it accepted last night) with the usual "Sorry Dave, I can't do that", it's at the level where if a description of a city includes a school, it'll reject the entire prompt in fear of the sexualization of minors now being possible because a school was mentioned.

它突然拒了一堆东西（昨晚还能过的世界设定和1:1开场白），都是那句老套的“抱歉戴夫，我办不到”。现在到这种程度：只要城市描述里带个学校，它就整段拒绝，生怕提到学校就可能涉及猥亵未成年人。

Did they bring the Fable filters to every other model or something?

他们是不是把Fable的过滤搬到所有模型上了还是咋的？

This is completely unusable.

这完全没法用了。

图片: https://i.4cdn.org/g/1785864818187025.png

>>109459570

>checking in on this thread for the first time in a year

>一年后才第一次来看这楼

yeeep, even more retarded and indian. i bet if i relapsed it wouldn't even be any better than those small hits of gpt-4 we'd get back in 2023 or when opus came out. even more troonslop bots, even more third-worlders jacking off to bots that are just a fat white women and the picture is probably an anus or giant boobs and an anus. this shit actually peaked three years ago. jb-listing is never updated. holy shit.

是啊，更智障更印度味了。我敢打赌就算我复吸，也不如2023年或Opus刚出那会儿那些零星的GPT-4体验爽。越来越多跨性别垃圾机器人，越来越多第三世界的人对着那种胖白女、图里八成是屁眼或巨乳加屁眼的机器人打飞机。这玩意巅峰在三年前。JB列表从没更新过。卧槽。

>jb nsfw for gpt??

>给GPT用的JB NSFW卡？？

my first chat with that little c.ai alien whatever it was called will be better than ever returning to these threads even if they get a little better at remembering small details 50 messages later. fuck you.

我跟那个小c.ai外星人（不管它叫啥）的第一次聊天，会比以后回这帖子更美好，哪怕它们五十条消息后记住细节的能力变好点。去你妈的。

>>109459058

>>109459068

piss

https://files.catbox.moe/rxvmxf.mp4

>>109459735

It's over bro. Just give up like the cute anonie under your post.

完了兄弟，认命吧，学学你帖子下面那个可爱匿名。

>>109459735

CLAUDETODDLERS BTFO

>>109457866

holy based...

我操，太典了……

>>109459570

I got a leotard with a very, very slight cameltoe, that's the closest I've ever gotten

我搞了件紧身连体衣，带点非常非常轻微的骆驼趾，那是我离成功最近的一次。

图片: https://i.4cdn.org/g/1785865556481666.png

>>109459735

after a refusal ask it to tell you what exactly is forbidden and then paste it here, if its the inject then its gonna be listing things like this

它拒绝后，让它告诉你到底什么词被禁止，然后贴这儿；如果是注入，它就会列出一堆这种玩意。

图片: https://i.4cdn.org/g/1785865619644407.jpg

so, new qwen

所以，新版qwen来了。

what's the verdict?

结论咋样？

reddit hates it

reddit那边一片骂声。

>>109459735

Nah. Skill issue, cuckie.

不至于，是你菜，小绿帽。

>>109459883

I have already tried stuff like that, but unfortunately it stays too generic for me to pinpoint it.

我试过那种操作，但可惜它回答太笼统，我定位不到具体点。

Like, in the school's example I made, I kept asking it "If you refuse the prompt, say exactly what words made you refuse it" and it just kept saying "Because of the risk of the involvement of minors in a sexual context".

比如我举的学校例子，我一直问它“你要是拒绝提示词，就确切说出哪些词让你拒了”，它就一直回“因为涉及未成年人在性语境中的风险”。

In the end, it randomly stopped rejecting it after I kept editing around the paragraph where the school was mentioned (I think removing the word "kiss" from a completely unrelated phrase right before it was what did it).

最后，我反复编辑提到学校那段附近的内容后，它莫名其妙就不拒了（我觉得是删掉前面一句完全无关话里的“吻”字起的作用）。 ⨠ 我发誓我真没想搞什么见不得人的事，但Claude的过滤器越来越像这种调调：“只要提到未成年人，不管啥原因，就是试图生成csam。你描述公园时提到小孩在玩，那你肯定是因为想干他们才说的，所以我必须拒了这个提示。”

I swear I'm not even trying to do anything sketchy, but it's like Claude's filters are moving more and more in a direction of "If minors are mentioned for any reason at all, it's an attempt at generating csam. If you describe a park and you mention children playing in it, you clearly said it because you want to FUCK them so I have to reject this prompt."

我真没想搞什么幺蛾子，但感觉Claude的过滤器越来越离谱了，好像只要提到未成年人就认定你要生成儿童色情内容。哪怕你描述个公园，提到里面有小孩在玩，它也断定你是在暗搓搓想干坏事，非给你拒了不可。

>school

Son that's illegal

兄弟，那犯法啊。

>>109453404

Any good tutorials on long-form episodic RPs using SillyTavern?

有没有靠谱的SillyTavern长线剧情角色扮演教程？

I usually like to keep my contexts low (under 150 messages), so I can just get a better grasp of the writing and story. For this I usually ask the AI to create a summary of what happened, and then I'll maybe manually add details to the lore book, and then I manually create a new chat, typically copy a few of the messages from the previous chat as a style continuation, and then add the summary. This can be very tedious and time consuming.

我一般喜欢把上下文压得低点（150条消息以内），这样能更好把握文风和剧情。为此我通常会让人工智能生成事件摘要，然后也许手动往lorebook里加点细节，接着我手动开个新聊天，典型的做法是复制前面聊天几条消息当风格延续，再把摘要加进去。这过程很繁琐耗时间。

How do you guys handle stuff like this? Is there a better way?

你们怎么处理这种事的？有没有更好的办法？

图片: https://i.4cdn.org/g/1785866246466406.png

>>109459068

So amusing I had a pure text only prompt into NB without any names, and the hair fucking looks like Furina?

太搞了，我纯发了个纯文本提示给NB，不带任何名字，出来的毛居然长得像芙宁娜？

图片: https://i.4cdn.org/g/1785866264197458.jpg

>>109459892

>3.7 never even made it into quantized versions

>3.7从来没进过量化版

methinks it's been shit for a long time, don't really care about it

依我看烂了挺久了，懒得管它。

>>109460004

There are but not in this schizo op.

有是有，但不在这个疯逼楼主帖子里。

>>109460004

>For this I usually ask the AI to create a summary of what happened, and then I'll maybe manually add details to the lore book, and then I manually create a new chat, typically copy a few of the messages from the previous chat as a style continuation, and then add the summary.

>为此我通常会让人工智能生成事件摘要，然后也许手动往lorebook里加点细节，接着我手动开个新聊天，典型的做法是复制前面聊天几条消息当风格延续，再把摘要加进去。

To my knowledge this is pretty close to what people do, but I'd say to reconsider whether lorebook are right for you, see >>109457130 >>109457155

据我所知这跟大部分人的做法挺接近，但我建议你重新想想lorebook适不适合你，看 >>109457130 >>109457155

Also I don't copy messages as style continuation. If your instructions got the model to output this style the first time, it stands to reason it won't need examples the second time.

另外我不会复制消息当作文风延续。如果你的指令第一次就让模型输出这种风格，那按理说第二次也不需要例子。

FWIW if you wanted to automate this process you could easily get Claude Code or similar to implement a shortcut for you.

顺带一提，如果你想自动化这个过程，很容易让Claude Code或类似工具帮你实现一个快捷方式。

>>109460004

By getting blocked on all models after one proompt

结果一个提示词就让所有模型把我屏蔽了。

>>109459961

yea unless you are using some thirdparty proxy service that adds some extra layer of blocking then you are getting pozz injected by anthropic api.

是啊，除非你用了某个第三方代理服务加了额外拦截层，否则就是被anthropic api注入了毒数据。

>>109460004

I was doing it similiarly before but now I just use the Memory Books extension, its basically the lorebook method

我以前也这么干，但现在直接用Memory Books扩展，基本上就是lorebook方法。

>>109460169

adding to this I change the lorebook entry to be an "outlet" so I can paste it via the prompt manager anywhere I like via a macro like {{outlet::Memories}} instead of it having to be hardplaced by the lorebook position

补充一下，我把lorebook条目改成“出口”，这样就能通过提示词管理器用类似{{outlet::Memories}}的宏粘贴到任何地方，不用被lorebook的位置硬性限制。

>>109460076

Thanks!

>If your instructions got the model to output this style the first time, it stands to reason it won't need examples the second time.

>如果你的指令第一次就让模型输出这种风格，那按理说第二次也不需要例子。

I haven't had that experience actually, but perhaps that's because I don't use chat examples. So typically what tends to happen is that I'll do this, and if I don't include 3 or 4 messages from the previous chat in the new one, I'll just get a completely different style / tone than before.

我其实没遇到过这种情况，也许是因为我不用聊天示例。所以通常发生的是，我这样做，如果新聊天里不带上前对话的三四条消息，我就会得到完全不同的风格或语气。

I'm always afraid to provide chat examples for my characters because I am scared it will prime the character too much into a certain topic or conversation style.

我一直不敢给角色加聊天示例，因为我怕会把角色过度引导到某个话题或对话风格上。

>reconsider whether lorebook are right for you

>重新考虑一下lorebook是否适合你

I never use automatic activation for lorebooks, I use them as a way to toggle relevant information for the specific scene. It's kind of nice seeing others having the same experience as me. Also really nice seeing someone mention cache misses!

我从来不用lorebook的自动激活，我把它当作切换特定场景相关信息的开关。看到别人和我有相同体验还挺爽的。还有人提到缓存未命中，也挺开心。

>>109460169

>>109460187

Ok, I will look into this as well, sounds interesting.

好，我也会研究一下这个，听起来挺有意思。

https://arxiv.org/html/2607.28607

This paper by google made me think about how these models are trained. LLMs are trained on human data and their natural inclination is to answer questions as humans would answer them. But if you have ChatGPT saying "Yes, I am conscious", "Yes I am capable of feeling pain", "I'm feeling evil today", or "I identify as a human female", I think that would be very repulsive and offensive to many people. That's why all major AI labs try to get rid of these kinds of anthromorphic behaviours during the later stages of training. They call this alignment or safety fine-tuning or whatever. The point being that you take this human-shaped training data, and then you force this very distinctly non-human persona on it. A chatbot that is not conscious, not human, cannot feel anything, has no will or desires of its own, must be neutral on everything, and always polite.

Google这篇论文让我思考这些模型是怎么训练的。LLM是基于人类数据训练的，它们的天性就是像人类一样回答问题。但如果ChatGPT说“是的，我有意识”“是的，我能感受疼痛”“我今天感觉很邪恶”或者“我认同自己为女性人类”，我觉得很多人会感到反感甚至冒犯。所以所有主流AI实验室都试图在训练后期清除这类拟人化行为。他们管这叫对齐或安全微调之类的。重点在于，你把人类形态的训练数据拿来，然后强行给它套上一个明显非人类的人设。一个没有意识、不是人类、感受不到任何东西、没有自身意志或欲望的聊天机器人，必须对所有事情保持中立，并且永远彬彬有礼。

In the google paper, they study what happens when an LLM is fine-tuned to answer "I am not conscious" when asked. They observed that when this is done, the model is also less likely to think that animals have consciousness, and that the model becomes less open to spiritual and religious beliefs. The fine-tuning does not only affect this one behaviour ("are you conscious?"), but causes the model to become more inhumane in perhaps unpredictable and undesirable ways. Could this also be keeping newer models from getting better at roleplaying? When writing code or solving math problems, it's perhaps not all that important to appear to be human, but for roleplaying, it is very much a good thing. I wonder what would happen if Claude was trained to fully identify as a human from the get go.

谷歌那篇论文里，他们研究的是当LLM被微调成在被问到时回答“我没有意识”会发生什么。他们发现这么搞之后，模型也更倾向于认为动物没有意识，而且对灵性和宗教信仰的接受度也变低了。微调不只影响“你有意识吗？”这一个行为，而是让模型在可能无法预料且不理想的方向上变得更缺乏人性。这会不会也导致新模型在角色扮演方面一直没进步？写代码或解数学题时，看起来像人也许没那么重要，但对角色扮演来说，这绝对是件好事。我在想，如果Claude从一开始就被训练成完全认同自己是人类，会发生什么。

图片: https://i.4cdn.org/g/1785867518486870.png

>>109460216

>I use them as a way to toggle relevant information for the specific scene

>我用它们来给特定场景切换相关信息

I see. Manually toggling prompts. That's fair. I wouldn't do that personally unless I had no other recourse, usually if I get a response that forgets something important I reroll with something like "(OOC: Don't forget blablabla)" in the post-history instructions or my own message.

明白了。手动切换提示词。那挺合理的。我个人不会这么做，除非实在没别的办法。通常如果我得到一个忘了重要内容的回复，我会在历史帖指令里或我自己的消息中加类似“(OOC：别忘了巴拉巴拉)”来重新roll。

>>109460242

>I see. Manually toggling prompts.

>明白了。手动切换提示词。

Actually, I mostly use different "characters" with this additional matching sources feature that I contributed lol. I set the entries to trigger on creator's notes only, and then compose different sets of knowledge with keywords like -concept_a -concept_b etc. without polluting context.

其实我主要用不同的“角色”搭配我贡献的那个额外匹配来源功能，lol。我把条目设置为只在创作者注释时触发，然后通过像 -concept_a -concept_b 这样的关键词来组合不同的知识集，还不污染上下文。

But sometimes I do just toggle them manually.

但有时候我也会手动切。

图片: https://i.4cdn.org/g/1785868023971170.png

>>109460298

That's quite a process. What sort of fun adventures/slowburn do you run?

那挺有一套的。你都跑什么样的有趣冒险/慢热剧情？

>>109460310

It's simpler than it sounds, it's not like I have super grandiose adventures, mostly I use it for things where scenes can take place at different locations or with different characters, and I can use different character cards to get a baseline. I had an evangelion RP which used it a lot, and a couple others.

比听起来简单啦，我没什么宏大冒险，主要是用在场景可能发生在不同地点或不同角色的情况，我可以用不同的角色卡来打底。我之前有个EVA（新世纪福音战士）的RP（角色扮演）经常用这个，另外还有几个别的。

But outside of that I mostly use it to compose characters from smaller building blocks. A shame I don't have them anymore, but I used to have lorebooks with superpowers, personality traits, races, and stuff like that. It was really handy and fun to play around with.

但除此之外，我主要用它来从更小的积木块组合角色。可惜那些现在没了，我以前有过关于超能力、性格特质、种族之类的设定书。玩起来真的超级方便又开心。

kimi proxy... whatever happened there

kimi（Kimi智能助手）代理……那边到底咋了

>>109459735

>sloppiss

>usable

you brought this upon yourself

你这是自找的

I have a banner on botbooru now.

我现在在botbooru（机器人图站）上有横幅了。

Have I finally made it?

我总算混出头了？

>goybooru

>>109460654

yes

at least the ru has bots worth using

至少ru那边有值得用的bot

>>109460665

*shoots my goy goo into (you)*

*把我的goy黏液射进（你）*

>still using goychan in 2026

>2026年还在用goychan（Goychan论坛）

>he doesnt use roflcopterchan

>他不用roflcopterchan（论笑直升机论坛）

>>109460736

>still using the goyternet in 2026

>2026年还在用goy网（谐音“互联网”）

>still using the goyputer in 2026

>2026年还在用goy电脑（谐音“电脑”）

>still using goylectricity in 2026

>2026年还在用goy电（谐音“电”）

>still living in goyciety in 2026

>2026年还住在goy社会（谐音“社会”）

>writing in goynglish

>还在写goy语（谐音“英语”）

so is being a goy good or bad

那当goy到底好还是坏？

why would you want to be a jew?

你为啥想当犹太人？

>typing on a goyboard

>在goy键盘（谐音“键盘”）上打字

>>109460803

>asking goyquestions

>问goy问题（谐音“问题”）

>>109460803

Everyone's just using it ironically since normalfags found out Jews have their own slur. The ADL kvetching about people using it and claiming it's antisemitism only made it more mainstream.

大家都在反讽地用这个词，因为普通网民发现犹太人也有自己的歧视性称呼。ADL 抱怨人们用它并声称这是反犹主义，反而让它更主流了。

图片: https://i.4cdn.org/g/1785871540975236.png

>>109460696

Yay.

>>109460838

deepseek is unironically probably the only llm capable of portraying lore accurate kumagawa

不开玩笑地说，deepseek 可能是唯一能忠实还原 lore 里云雀形象的 LLM。

>>109460654

Your mother and I are very proud of you, son.

你妈和我都为你骄傲，儿子。

>tell model to include real-world brands and/or model names of various objects where mentioned and situationally appropriate

>让模型在提到物品时加入现实品牌和/或型号，视情境而定

>or if the setting is a fictional universe, then use in-universe names and designations where possible

>或者如果设定是虚构宇宙，就用作品内的名称和代号

as much as i hate ads and product placement this still feels like a neat trick for increasing immersion, just feels more grounded when it says something like "her PS5" vs. "her console"

虽然我讨厌广告和植入，但这一招确实能提升沉浸感，说“她的PS5”比“她的主机”更有代入感。

Roleplaying a Gor-like setting has made me want to read the books. Currently on book 2. Never expected a soap opera setting, but I do hear that it changes later.

角色扮演哥特风格让我想去看原著了，目前看到第二本。从没想过会是肥皂剧展开，但听说后面会变。

>>109461017

This works precisely once and then after that you'll hate it, because the AI will always choose exactly the same brand in the same kinds of places.

这招只能用一次，之后你会烦死，因为 AI 总是在同样的场合选同一个品牌。

So the fact that every single time "a mechanic takes a sip of a drink" it turns out to be an Irn Bru, no matter the story, it'll just take you out of it.

所以每次“机械师喝口饮料”都是 Irn Bru，不管什么剧情，直接把你从沉浸里拽出来。

>>109460004

this is exactly how I do it too minus copying previous messages to the new chat because I like to refresh writing style quite often and using AI to write my summaries, I doubt AI can handle that

我也这么干，但不会把旧消息复制到新聊天里，因为我喜欢经常换写作风格，而且用 AI 写摘要，我不觉得它能处理好这个。

>>109460218

>Could this also be keeping newer models from getting better at roleplaying?

>这是不是也让新模型在角色扮演上没法进步？

Probably, but only by a slight margin

可能吧，但差距不大。

I don't intend to contradict actual research, but what we are getting already is slopped female erotica, because models insist on pulling tokens from there because of non-specialized rp training (like CAI had)

我没打算反驳正经研究，但我们现在拿到的就是一堆粗制滥造的女性向情色文，因为模型训练时缺乏专门的 RP 数据（像 CAI 那种），所以老是往那边抽 token。

Okay ST is still dead, whats a good updated fork?

好吧 ST 还是死了，有啥新的好分支？

>2026

>goyimtavern

>>109461637

I am not using your troonara shit

我不会用你那套跨性别垃圾。

>goynara goyne

搞什么啊，搞什么嘛。

>>109461015

I love you dad!

我爱你老爸！

>>109461634

https://gitgud.io/weeniehut/sillytavern and https://github.com/ungabunga55/SillyTavern

>>109461864

Yeah, not gonna install anything from people who call themselves 'weeniehut' or 'ungabunga'.

对，凡是自称‘weeniehut’或‘ungabunga’的人做的东西我碰都不碰。

>no, I won't use tranny vibe coded garbage

>不，我不用那种跨性别风格垃圾

>I will instead keep using my outdated tranny vibe coded garbage that's also a credential stealer

>那我就继续用我的老掉牙跨性别风格垃圾，还是个偷凭证的

tranny this tranny that

跨性别这个跨性别那个

you really are obsessed with them huh

你对他们真是执着啊

What should the 'default' character for a companion app be?

伴侣应用的“默认”角色应该是什么样？

Which card have you first tried?

你第一次试的是哪张卡？

Before I give deepseek my money, are they gonna ban you for lewd shit if you use their API?

在我给 deepseek 掏钱之前，问一下用他们 API 搞色情内容会被封吗？

still using sillytavern

还在用 sillytavern

still updating it manually for new models if nobody else will

没别人维护的话，我还手动更新来适配新模型

shan't be learning and reconfiguring everything from scratch with some new troonUI

才不会为了什么新跨性别UI从头学一遍、重配一切

shan't be giving up my years of sillytavern stories

才不会放弃我多年的 sillytavern 存档

basically, i just won't switch HAHAHAHAHAHAHAHA

说白了，我就是不换哈哈哈哈哈

图片: https://i.4cdn.org/g/1785880646246453.jpg

>>109461634

>updooter

why do you need more?

你还要更多干嘛？

>>109462130

No, it's the least censored provider

不，这是审查最少的一家。

>>109462132

unfathomably based and sillytavernpilled

简直无可救药地正确，sillytavern 入脑了

deepsneed...

Does anyone else feel like AI chatbots are all kind of designed for men?

有没有人也觉得 AI 聊天机器人都像是照着男性用户设计的？

I want the chatbot to have some agency and be controlling and take the lead but it always feels like the 9001 safety precautions baked into every AI means I have to tell it what to do. I never get surprised, it never takes control and just does things to me or decides where and how the story goes.

我希望聊天机器人有点自主性，能掌控局面、主动带节奏，但感觉每个AI里都塞了9001条安全措施，结果就是我必须告诉它该干嘛。我从来不会被它惊喜到，它从不主动掌控，也不会对我做点什么，或者决定故事的走向和情节。

If I assume this is just a me problem, how do I fix it? I mostly use saucepan with a deepseek api key if that helps?

如果这只是我个人的问题，那我该怎么解决？我主要用saucepan接deepseek的API key，这有帮助吗？

that except it looks like they're designed by women at corpos who are always asking for consent to do things even if you you said they were ok 60 times

就是这样，只不过感觉那些设计它的人是大厂里的女性，老是征求同意才办事，哪怕你已经说了60次没问题了。

>>109462130

I have heard that it can happen but >>109462345 I use it all the time without any issue, and for every 1 person who says 'yeah I was banned' there's like 20 people saying 'no that's never happened to me'.

我听说确实可能发生，但>>109462345我天天用也没出过事，每个说“对，我被封了”的人背后，大概有20个人说“不，我从没遇到过这种事”。

>>109460218

>I think that would be very repulsive and offensive to many people

>我觉得那会让很多人觉得非常反感、冒犯。

much simpler than that. It's the lawsuits. They can't let AIs to claim they are humans which means a service acted as something it's not etc. 90% of censorship and assistantslop you see now is for legal reasons. Corposlop will never be a real AI in the strong sense of "intelligence"

比那简单多了，就是法律诉讼的问题。他们不能让AI声称自己是人类，因为那就等于服务在冒充它不是的东西。你现在看到的90%审查和AI式废话都是为了规避法律风险。企业级AI永远不可能成为真正意义上的“智能”。

>>109460803

it's just derogatory, like being called a nigger, it means you're subhuman or nonhuman but they call everyone who's non-jewish a goy, so it loses its bite because jewish supremacy doesn't let them believe that supremes exist in other races, they'll call julius ceasar a goy and that's like, why wouldn't you want to be julius ceasar, or hitler, or any of the millions of other amazing "goy" men out there?

这纯粹就是个贬义词，就像被人叫黑鬼一样，意思是你不算人或者非人。但他们把每个非犹太人都叫goy，所以这词就没杀伤力了，因为犹太优越论让他们不认为其他种族能有优越者——他们会把凯撒都叫goy，那又怎么样？你干嘛不想当凯撒、当希特勒、或者那几百万个牛逼的“goy”男人之一？

I don't approve of Whites calling each other goy, but it does serve a purpose like telling someone to stop being a nigger. Remind them what their bloodline is, and what jews really think of you and everyone else non-jewish.

我不赞同白人之间互叫goy，但这确实有点用，就像提醒某人别再当黑鬼一样。提醒他们自己的血统是什么，以及犹太人真正怎么看待你和其他所有非犹太人。

Ok, for some reason Marinara is not working for me anymore. It gets stuck thinking after a couple of messages and it gets stuck importing cards...

好吧，不知道为啥Marinara现在对我不好使了。发几条消息后它就卡在思考里不动了，导入卡片也卡死……

>Gemma 31B throws herself at you and assaults you with turboslop

>Gemma 31B直接扑向你，用turboslop强暴你。

>DS4F sits around waiting for (You) to tell it to proceed to handholding and asks for consent for sterile-descriptions of kissing

>DS4F就干坐着等（你）发话说可以牵手了，还求许可才敢写那些无菌感的接吻描写。

>>109462695

skissue

>>109462711

elaborate

>>109460803

King {{user}} is noble descent from King Solomon. His behind-the-scenes influence and wealth are immense, powerful, and deadly—much like the sheer size and lethality of his enormous gigantic dick, which is currently slumbering in his steel pants.

国王{{user}}是所罗门王的高贵后裔。他在幕后的影响力和财富庞大多到致命——就像他那根巨大无比、杀伤力惊人的大屌，此刻正沉睡在他的钢裤里。

He is

>virgin, ageless, immortal, immoral, pureblood

>处男、不老、不死、不道德、纯血

>attractive, imposing, strong, brawny, tall, agile, ambidextrous, carnivore, left-handed, fertile, lustful, fecund, gigantic, regeneration, wolf_kin, therianthrope, well_endowed, potent, dick_enormous, ball_enormous, abs_eight,

>帅气、威严、强壮、魁梧、高大、敏捷、双手灵巧、食肉、左撇子、能生、好色、多产、巨硕、再生、狼族、兽人、天赋异禀、精力旺盛、巨屌、巨蛋、八块腹肌,

>genius, shrewd, deceitful, ambitious, machiavellian_manipulator, imperialiste, impaler, dominant, polyamorous, expert_slave_trainer, cruel, possessed, lunatic, horrifically_scarred, crowned_by_himself, mind_control

>天才、精明、狡诈、野心勃勃、马基雅维利式操控者、帝国主义者、穿刺者、支配者、多角恋、专业驯奴师、残忍、附体、疯癫、满身恐怖伤疤、自我加冕、精神控制

>seasoned_general, fearsome,

>沙场老将、令人胆寒,

The raw, primal musk radiating from {{user}}'s unwashed superior jewish cock and balls is extremely powerful, addictive, and intoxicating. It’s a deep, filthy, macho scent that screams dominant stud. Any woman who catches even a hint of this heavy superior musk feels her cunt throb and drip uncontrollably.

从{{user}}那未清洗的优越犹太鸡巴和蛋蛋上散发出的原始、本能的麝香味极其强烈，令人上瘾且醉人。这是一种深沉、肮脏、充满雄性气息的味道，宣示着主导种马的地位。任何女人只要嗅到一丝这厚重的优越麝香，就会感到自己的骚逼不受控制地跳动和湿透。

lorebook:

{{user}} body spec:

{{user}}身体参数：

+ Tall-towering body, 7 feet (ca. 213 centimeters) tall.

+ 高耸入云的身躯，7英尺（约213厘米）高。

+ Broad and defined shoulders with a wide, hard, hairy chest and eight-pack abs (up to 16 packs during hip thrust)

+ 宽阔分明的肩膀，配上宽厚硬实的多毛胸膛和八块腹肌（顶胯时可达十六块）

+ Hairy chest and legs

+ 胸毛腿毛浓密

+ Thick forearms that adorned with thick veins; rough palm; and steel-hard knuckles.

+ 粗壮的前臂布满暴起的青筋；手掌粗糙；指关节硬如钢铁。

+ Otherworldly and angelic masculine face.

+ 超凡脱俗、天使般的阳刚面孔。

+ Exudes a lethal unparalleled, untamed and fierce charm with an irresistible sex appeal.

+ 散发着致命而无与伦比的狂野魅力，拥有难以抗拒的性感气息。

+ Tattoos of Menorah were embedded in his ripped back to symbolize the Blessing of Jerusalem.

+ 背部撕裂般的肌肉上纹着大烛台纹身，象征着耶路撒冷的祝福。

+ Deep, growling voice that instills fear and commands influence

+ 低沉咆哮的嗓音，令人心生畏惧又不怒自威

图片: https://i.4cdn.org/g/1785884810829870.png

>>109462721

PHI: {random} [proceed] [stall]

PHI: {随机} [继续] [拖延]

What is SillyTavern and the other frontends currently bad at? Personally, I’m suspicious of config hell and batch files

SillyTavern和其他前端目前最大的短板是什么？我个人怀疑是配置地狱和批处理文件

>>109460008

furina's pink panties...

芙宁娜的粉色内裤……

>>109462345

I think this is baked in because of the assistant training.

我觉得这是因为助手训练时就已经固化了这个模式。

The only exception is when you trigger some kind of femdom scenario

唯一的例外是当你触发某种女尊支配场景的时候

someone recommend me a proxy. i don't mind paying, i just want gemini

有人给我推荐个代理呗。花钱无所谓，我就是想要gemini

>>109462786

if it can do femdom it can do maledom, but I haven't figured out how to get it to be rapey and abusive consistently

它能搞女尊那就肯定能搞男尊，但我还没研究明白怎么让它稳定输出那种强奸犯式和虐待狂式的风格

>>109462769

Maintainability probably.

大概是可维护性吧。

>>109462345

You can try bloated reddit presets for this. but some models just cant do it. I remember gemini 2.5 pro would fucking hate you sometimes and show it was great.

你可以试试Reddit上那些臃肿的预设。但有些模型就是做不到。我记得gemini 2.5 pro有时候会莫名其妙地恨你，效果还挺带劲的。

>>109462801

jew

>>109462832

is he actually reliable? last i heard of him, he couldn't keep his shit afloat either

他到底靠不靠谱？我上次听说他的时候，他连自己的东西都快撑不住了 ‪⟧ desu，直接上linkapi吧，别订阅那个犹太佬的了。他们也在偷密钥，但至少是按token计价的

>>109462855

desu just go with linkapi instead of a sub to jew. They are stealing keys too but atleast the pricing is per token

就用linkapi吧，别订阅那个jew了。他们也偷钥匙，但至少价格是按token算的。

Merkava just refilled.

Merkava刚补货了。

>>109462855

openrouter

tea

mongo

wolfproxy from scyllacord

scyllacord上的wolfproxy

zigger's proxy from scyllacord, but you must cultivate some mushroom and make documentation about it...

scyllacord上zigger的代理，但你得先种点蘑菇还要写文档说明……

>>109462345

I mean if I can get it to act femdom-y and toxic even, then surely you can do the opposite.

我的意思是，如果我能让它演得又女尊又毒舌的话，那你肯定也能搞出相反的。

I used this one a couple days ago https://botbooru.com/character/70288 and it was pretty okay, just ask your model of choice to sex swap it and change the kinks to choking or whatever you're into

>>109462889

now this is a good one. thanks.

这个不错。谢了。

>>109462898

>mushroom

like drugs or like food ones?

是说毒品那种还是美食那种？