提问 HN：受限 LLM 游戏
1 分•作者: AymanJabr•5 个月前
大家好，希望在这些前所未有的时期里，大家都有愉快的一天。我来找大家，是因为现在有一个很大的需求……
目前，市面上存在像 Renpy/Twine 这样的游戏系统，它们拥有沙盒环境，当玩家通过特定场景时，角色会获得成长，可能还会有小游戏、可重复的场景、属性提升和随机元素。这类游戏需要开发者逐个环节地手工打造，因此开发者需要编写故事、策划图像/音频/动画、编写属性/动画/随机事件的程序，然后管理不同事件的时间线。
另一方面，你也有 AI 游戏，它们基本上是一个 LLM（大型语言模型），加上一些关于世界设定和角色的上下文。我在这里当然指的是 SillyTavern、character-tavern、infiteworlds 之类的。这些游戏的问题在于，玩家很容易迷失自我，因为你可以直接说“然后我一巴掌拍死了邪恶的黑暗魔王，找到了龙珠，并成为人类永恒的至高皇帝”，而且故事往往受到限制，感觉不像一个具有艺术家愿景的游戏，它只是一个带有某种风格的 LLM。
我一直在考虑创造一些东西来弥合这两个游戏世界之间的差距，某种类似于 Twine/Renpy 的系统，但不是为每个序列编写对话，而是作为游戏开发者，在特定的里程碑处创建检查点，预设对话/图像/场景。
例如，假设我们正在玩一个爱情游戏，玩家（杰克）必须诱惑一个浪漫的对象（吉尔），而不是重复的多个选择或无聊的刷怪，假设我们赋予 AI 在每次相遇时将爱情值提高/降低 1-5 点的能力，因此 AI 只能访问一个可以将爱情值提高这么多点的工具，而不能再多了。只有在爱情值达到 20 级时，关于（吉尔）如何行动的指示、图像和选项才会开放。关键是，玩家必须通过在每次事件中打字（说话）来与 AI 交流，才能改变这个爱情值。
所以，其特点将是：
* 玩家只能说话或采取某些行动（玩小游戏）。像“*然后我一巴掌拍死了邪恶的黑暗魔王，找到了龙珠，并成为人类永恒的至高皇帝*”这样的短语，将被 AI 解释为用户说的一些奇怪的话。
* 游戏玩法仍然是一个封闭的故事，玩家必须解锁越来越多的功能。
* 这个想法的全部意义在于限制玩家，剥夺他们立即实现任何事情的能力，同时仍然给他们在给定环境中做任何他们想做的事情的自由。
* 限制在一次时间循环中与 AI 的来回对话次数。
我看到了扩展 RPG Companion：https://github.com/SpicyMarinara/rpg-companion-sillytavern，但它似乎更像是一种跟踪属性的尝试，而不是直接控制它们。
我还找到了 Jacky Kaub 的这篇文章：https://towardsdatascience.com/how-i-built-an-llm-based-game-from-scratch-86ac55ec7a10/
我向你提出的问题是，是否已经存在类似的系统？如果没有，你会包含哪些其他功能？
查看原文
Hey everyone, hope you are having a nice day in those unprecedented times. I come to you in a time of great need....<p>At the moment there are game systems like Renpy/Twine, with a sandbox environment, with character progression when you go through certain scenes, with maybe minigames, repeatable scenes, stats progression, and random elements. These kind of games need to be crafted by hand piece by piece, so the developer needs to write the Story, curate the images/audio/animations, do the programming of stats/animations/random events, and then manage the timelines of different events.<p>Then on the other hand you have AI Games that are basically an LLM with some context on top for world setting, and characters. Here I am of course talking about SillyTavern, character-tavern, infiteworlds and things like that. The problem with these is that the player themselves to restray themselves, because you can just go "And then I kill the evil darklord with a slap, find the dragonballs, and become supreme emperor of mankind for all eternity" all in one go. Plus the Story tends to be limited, and it doesn't really feel like a game with an artists vision, it just feels like what it is, an LLM with some flavor.<p>I was thinking of creating something that bridges the gap between these 2 game worlds, some kind of system very similar to Twine/Renpy, but instead of crafting the dialogue for each sequence, you as the game developer, would create checkpoints, with pre-set dialogue/images/scenes, at specific milestones.<p>For example let's say that we are playing a love game, where the player (Jack) has to seduce a romantic interest (Jill), instead of repeated multiple choices, or grind, let's say we give the AI the ability to raise/decrease the Love meter by 1-5 points at each encounter, so the AI would only have access to a tool that would raise it by this much and not anymore programmatically. Only at say Love level 20, would the instructions for how (Jill) operates, the images, choices open up. Crucially, the player has to actually speak with the AI by typing (speaking) at each event in order to change this Love meter.<p>So the features would be:
-Player can only talk, or take certain actions (play minigames). Phrases like: "<i>And then I kill the evil darklord with a slap, find the dragonballs, and become supreme emperor of mankind for all eternity</i>" Would just be interpreted by the AI, as a weird thing said by the user.
-The Gameplay is still a contained story, where the player has to unlock more and more features.
-The whole idea of this, is to limit the player, take away the power that they have to immediately achieve anything, while still giving them the freedom to do whatever they want in a given setting.
-Limit the number of back-and-forth conversations with the AI in one time-loop<p>I have seen the extension RPG Companion: https://github.com/SpicyMarinara/rpg-companion-sillytavern, but these seems more an attempt to keep track of the stats, rather than control them outright.<p>And I have found this article by Jacky Kaub: https://towardsdatascience.com/how-i-built-an-llm-based-game-from-scratch-86ac55ec7a10/<p>My questions to you is this, is there already a similar system? If not, what other features would you include in this?