---
title: "我想离开了 - Creator Exchange - Yumina Community"
description: "我想离开了，并非这个网站不好，相反，这个网站太好了，可惜目前的AI压根不是用来玩文字交互游戏的，给人的体验甚至还不如去年的AI，去年AI的重心放在文本创作，今年已经是竞赛agent能力、代码、工作任务这些，而文本创作能力相对于就降低了，我在这个网站花费了至少3千万token，也改变不了AI胡扯、说话阴间、逻辑混乱的问题，无论如何修改提示词，增加限制，游戏中的NPC说话都是神神叨叨、故作高深、无病…"
url: https://yumina.io/app/community/thread/2e3166a3-fbbc-4842-a9a1-fdcb78b7b905
lang: zh
---

[Community](https://yumina.io/app/community) · [Creator Exchange](https://yumina.io/app/community/creator-exchange)

# 我想离开了

[四](https://yumina.io/@psr369) · 2026-09-24

我想离开了，并非这个网站不好，相反，这个网站太好了，可惜目前的AI压根不是用来玩文字交互游戏的，给人的体验甚至还不如去年的AI，去年AI的重心放在文本创作，今年已经是竞赛agent能力、代码、工作任务这些，而文本创作能力相对于就降低了，我在这个网站花费了至少3千万token，也改变不了AI胡扯、说话阴间、逻辑混乱的问题，无论如何修改提示词，增加限制，游戏中的NPC说话都是神神叨叨、故作高深、无病呻吟、逻辑混乱，AI生成的内容也是用一些莫名其妙的修辞，我压根玩不下去，就像现实中如果真有一个人对着你这么说话，你压根不知道该如何回复。

## 25 replies

1. [Yumina](https://yumina.io/@yumina) · Staff · 2026-09-24 是的，这确实是很多新模型的问题，尤其是当时grok决定把4.1 fast取消掉让我们永远失去了这个好的老模型。但是你依然可以多试一些其他模型吧，比如gemini 2.5/gemini 3.0 这些去年就在的并还在持续运营的模型
2. [四](https://yumina.io/@psr369) · 2026-09-24 网站做得太好了，我才花费这么多时间和token在搭建和游玩上，现在才发现，软件再如何完美，AI本身不行就是不行，就像你们搭建了一个完美的学校，一个弱智进去读依旧是读不出好成绩的。
3. [四](https://yumina.io/@psr369) · 2026-09-24 qwen-max之前也有几个大模型我觉得挺牛逼的，知识全面、遵守指令、逻辑推理能力牛逼、文章也写得极好，现在也一代代将老模型废掉，3.7、3.8两个模型也全偏向于今年的AI竞赛项目了
4. [Oncin](https://yumina.io/@oncin) · 2026-09-24 for me gemini models works just fine for my game, just don't use budget models for complex games with a lot of variable it rarely works well
5. [pepper](https://yumina.io/@pepper) · 2026-09-24 i get you, last year chatgpt completely fucked over its narrative writing. it became so shallow i genuinely couldnt handle it anymore, and the the guardrails got so strict i was wondering what the point in the platform even was anymore. i switched to grok and stopped having those issues. i think it just has to do with experimenting with the right model for the card youre playing. also giving the ai clear enough directives, and making sure the directives can be followed (and thought through!) is very beneficial to cards. i think if youre dealing with burnout or overwhelm, you should take a break though. theres no point in stressing yourself if youre not even having any fun. maybe when taking a break youll come up with an idea thatll make you wanna try again!! i hope it gets better. &lt;3
6. [四](https://yumina.io/@psr369) · 2026-09-24 跟变量没关系，问题在于写出来的内容不像人写的，NPC的话也不像人话。比如“他托着的那只陶罐的兄弟罐子，大约正架在谁家灶上”，啥玩意罐子还有兄弟罐子；“谁心里都有一杆秤，秤的是应还是不应。” 他偏过头看你，“你在南边那寨子，见没见过不应的人？”，我看见这种对话我都不知道该如何回复。
7. [四](https://yumina.io/@psr369) · 2026-09-24 花了大量token用于让AI助手解决AI味文风的问题，加了很多束缚、规则依旧改变不了，之前用qwen3.5max还是几，极其好用，但今年的大模型都是朝着agent能力在进化，现在的大模型能更好地写出代码、程序，成为程序员更好的助手，但就是写文章这个能力被挤下去了，今天下午我在尝试“面板+状态”的玩法，不让AI写小说，直接跟玩策略游戏一样写状态和数值，目前还在修改中
8. [Oncin](https://yumina.io/@oncin) · 2026-09-24 That's a funny line 😂, I don't think this is necessarily an unavoidable model problem. My bots use the same general class of models, and the dialogue usually comes out very natural. What helped me a lot was adding strong enforcement rules specifically for prose and NPC speech, not just character/world instructions. Things like: NPCs must speak like actual people, metaphors must make contextual sense, avoid cryptic/philosophical filler, don't force poetic rhetoric into ordinary conversation, and keep cause-and-effect logic grounded. Without rules like that, the model can absolutely drift into the “everyone speaks like a mysterious fortune cookie” style you're describing 😭
9. [四](https://yumina.io/@psr369) · 2026-09-24 约束里全写了，已经充分利用到AI助手
10. [Oncin](https://yumina.io/@oncin) · 2026-09-24 I’ve noticed the model can still ignore perfectly good style rules if they’re buried under a lot of other instructions, repeated inconsistently, or competing with tone/worldbuilding guidance elsewhere. A pretty useful way to inspect this is to download/export the entire game JSON and give the whole thing to ChatGPT or Claude, then ask it to audit the prompt architecture rather than rewrite it immediately. Have it look specifically for conflicting instructions, duplicated rules, weakly prioritized constraints, prose/style instructions being overridden by other sections, and anything that could accidentally encourage cryptic or overly literary dialogue. Sometimes the individual rules are perfectly fine, but the full prompt is pulling the model in several directions at once.
11. [四](https://yumina.io/@psr369) · 2026-09-24 也弄过，让AI助手检查全局，修改有冲突、冗长的提示词，有时候一次检查和修改能花掉我100多万token，还包括NPC喜欢追着我问问题，你一回答他又追着质问你的答案里的漏洞让你解释、NPC喜欢让我做决策，我不表态他们跟木桩一样、NPC在情节未推动或者场景相同的情况下说出相似度百分之九十的话，反正大量时间和token花在了调试上面了
12. [Oncin](https://yumina.io/@oncin) · 2026-09-24 Yeah, in that case I wonder if you might be relying on the platform AI assistant a bit too much for the auditing itself. I used to do the same thing. I’d keep asking the assistant to inspect, revise, inspect again, and it could burn a ridiculous amount of tokens without actually finding the deeper structural problem. Lately I’ve been exporting the whole game and using an outside model like ChatGPT or Claude as a second pair of eyes for auditing, prompt architecture, and even general development. That has worked much better for me. Sometimes an independent model notices conflicts or behavioral patterns that the assistant working inside the same environment keeps overlooking. So if you’ve already exhausted the built-in assistant, I’d honestly try having another AI audit the full JSON from scratch instead of continuing to iterate with the same one. At least for now, that might save you a lot of tokens and frustration 😭 because i once tried to implement one ui feature and it consumes 19k mushies 😭 you can see in creator exchage me talking about it
13. [四](https://yumina.io/@psr369) · 2026-09-24 我的AI助手用的是我外接的api，我用的doubao-seed-2.1Pro和deepseek v4Pro大模型，导出去其实我也只能用这些
14. [giantfukker](https://yumina.io/@giantfukker) · 2026-09-24 You then may like saucepan AI. However. I am staying. its rare to find a service thats transparent that's not greedy these days its either your transparent but greedy or greedy and zero transparency few are not greedy and very transparent. however, I dont mean to dunk on Yumina Ai but i am finding that i think I like saucepan's model better. but, i do suggest that Yumina Ai partner with saucepan on the merit that I think they have similar values and goals! and having partners to scratch your back will be a good thing and puts you on the map more and it gives you a good back bone against a potential Microsoft buy out or any other company. So, i am gonna use em both. I also, like character Tavern's direction that one is a one man show. bro must be fuckin rich or something! But, no investors no board nobody else with opinions that work on the team thats the ultimate power user leadership. lol
15. [Swuell](https://yumina.io/@swuell) · 2026-09-24 It can’t do logical reasoning. You can only mimick things so far sans creativity is not something you can mimick without breaking. But yes the AI isn’t really AI. It’s a glorified autocomplete so it’s expected it’s not going to work the way you want if you rely on it too much. Majority of it still depends on you and then it copies and mimics you,
16. [Swuell](https://yumina.io/@swuell) · 2026-09-24 LlM still have major issues in keeping track of game state and stats related game mechanics, it works sometimes but when it doesn’t it either doesn’t change or it just hallucinates and changes it irrelevant of what your doing.
17. [Swuell](https://yumina.io/@swuell) · 2026-09-24 Because of online writing. Webnovels and netspeak destroy the capabilities of it writing properly. It only wrote better before because then it was using actual novels and books for database until this specially got legally cracked down, so now using free netspeak is what tech companies rely on since it’s free. Edit: by the way the other issue is Chinese LLM now has English syntax mixed in it since westerners used it more at the beginning of LLM boom. And Chinese companies don’t split the two languages. It’s all combined that’s why the syntax is now even worse than before.
18. [Swuell](https://yumina.io/@swuell) · 2026-09-24 Issue with this is if you’re mainly writing Chinese rhetoric than Claude, Gemini, ChatGPT etc. won’t work that well since those are all US based. So the majority of the database has English and western syntax. Using China’s own would be better but there’s the issue of blush syntax being mixed in due to overseas use early in the development. So you have to really structure the writing yourself with first few initial responses to help guide and guardrail the LLM to write in a better syntax.
19. [Swuell](https://yumina.io/@swuell) · 2026-09-24 This has to do with the directive that npcs don’t take over your character and your character makes the decisions that move the story along otherwise the llm will have ai talk, think, do things for your character while propelling the story further. The best thing you can do is have it set up to not take over your character and have opt character end things on responses or actions that the LLM can use to move the story along. Otherwise the llm will be either stuck waiting or just completely ignore you and move the story. LLM still has major issues. Since text creativity is not something the LLM can just pull a general model, or database to reproduce.
20. [四](https://yumina.io/@psr369) · 2026-09-25 目前的AI预测下一个词或下一个状态，而不是真正在想问题，在想剧情应该怎么走，目前甚至还遇见过，一旦你跟某个NPC聊到或者你自己行动某个举动，接下来的所有剧情全都是围着这个举动生成的剧情。每天打开资讯一天天的都是在叫赶紧踩刹车、agi时代来临、人类把控不了之类的话术，我们拿来玩个文字游戏都能体会到AI的低智

Page 1 of 2 · [Next page](https://yumina.io/app/community/thread/2e3166a3-fbbc-4842-a9a1-fdcb78b7b905?page=2)
