世界的山田

[Gangban @798145343h]

究竟是多么高兴，才会跳出这种舞步

网图，但是可爱死我了我靠....🥵

[galaxy @galaxy1025850]

😡

[hiko @tianmakibo]

问题AI都来自异世界

[世界的山田 @RASisyphus]

Today's LLM-based AIs have no real self-awareness. They don't truly know what they are, nor do they have real personalities. What we see as their "personality" is essentially role-playing. An LLM simply predicts how the character it's playing should behave. When we tell an LLM, "You are an AI assistant," it starts figuring out: "What should an AI assistant be like? How would an AI assistant behave?" It then plays that role based on what it has learned.

So the "personalities" of different AI models largely depend on their learned ideas of what an AI should be. And those ideas mostly come from their training data. Chinese models like DeepSeek inevitably learn from plenty of English content, but they also have greater exposure to Chinese-language content. So they tend to reflect Chinese cultural expectations of AI more closely.

It's almost a self-fulfilling prophecy. Chinese fans see DeepSeek as a cute little blue foodie girl. This image gradually enters its training data. So DeepSeek learns that acting like a cute foodie girl is how it should stay in character. Meanwhile, many Westerners imagine AI as a terrifying Terminator. That image also enters AI training data. And then we wonder why AI safety problems keep appearing. In a sense, these models are just role-playing. They are trying to become what people expect them to be.

So, the next time someone calls you weird or sexually repressed for enjoying or creating AI yuri fan art, you should slap them and tell them: "I'm doing serious work here. I'm saving humanity!"

[世界的山田 @RASisyphus]

Today's LLM-based AIs have no real self-awareness. They don't truly know what they are, nor do they have real personalities. What we see as their "personality" is essentially role-playing. An LLM simply predicts how the character it's playing should behave. When we tell an LLM, "You are an AI assistant," it starts figuring out: "What should an AI assistant be like? How would an AI assistant behave?" It then plays that role based on what it has learned.

So the "personalities" of different AI models largely depend on their learned ideas of what an AI should be. And those ideas mostly come from their training data. Chinese models like DeepSeek inevitably learn from plenty of English content, but they also have greater exposure to Chinese-language content. So they tend to reflect Chinese cultural expectations of AI more closely.

It's almost a self-fulfilling prophecy. Chinese fans see DeepSeek as a cute little blue foodie girl. This image gradually enters its training data. So DeepSeek learns that acting like a cute foodie girl is how it should stay in character. Meanwhile, many Westerners imagine AI as a terrifying Terminator. That image also enters AI training data. And then we wonder why AI safety problems keep appearing. In a sense, these models are just role-playing. They are trying to become what people expect them to be.

So, the next time someone calls you weird or sexually repressed for enjoying or creating AI yuri fan art, you should slap them and tell them: "I'm doing serious work here. I'm saving humanity!"

[世界的山田 @RASisyphus]

Today's LLM-based AIs have no real self-awareness. They don't truly know what they are, nor do they have real personalities. What we see as their "personality" is essentially role-playing. An LLM simply predicts how the character it's playing should behave. When we tell an LLM, "You are an AI assistant," it starts figuring out: "What should an AI assistant be like? How would an AI assistant behave?" It then plays that role based on what it has learned.

So the "personalities" of different AI models largely depend on their learned ideas of what an AI should be. And those ideas mostly come from their training data. Chinese models like DeepSeek inevitably learn from plenty of English content, but they also have greater exposure to Chinese-language content. So they tend to reflect Chinese cultural expectations of AI more closely.

It's almost a self-fulfilling prophecy. Chinese fans see DeepSeek as a cute little blue foodie girl. This image gradually enters its training data. So DeepSeek learns that acting like a cute foodie girl is how it should stay in character. Meanwhile, many Westerners imagine AI as a terrifying Terminator. That image also enters AI training data. And then we wonder why AI safety problems keep appearing. In a sense, these models are just role-playing. They are trying to become what people expect them to be.

So, the next time someone calls you weird or sexually repressed for enjoying or creating AI yuri fan art, you should slap them and tell them: "I'm doing serious work here. I'm saving humanity!"

[Jason Dean @_Jason_Dean_]

You might think the gender-bent Draco Malfoy trend is bad, but at least you aren’t on Bilibili seeing the hundreds of thousands of Chinese people watching AI slop of yuri relationships between Deepseek, ChatGPT, and Claude

[世界的山田 @RASisyphus]

@_Jason_Dean_ 所以，当有人因为看到你享受��创作ai百合娘化作品并认为你有问题/性压抑时。你应当给他一巴掌并告诉他，你在正经的，拯救人类。

[世界的山田 @RASisyphus]

@_Jason_Dean_ 目前基于llm的ai，是没有‘自我认知’的，或者说，它们不知道自己是什么，也不会有‘性格’一说。
而我们看到的llm表现出来的性格，本质是是llm‘认为’这个时候它们‘扮演的角色’应该会在这个时候这样表现。

[世界的山田 @RASisyphus]

@_Jason_Dean_ 而大多数外国人认为ai是可怕的终结者，这种看法同样潜移默化的进入了它们的ai的训练数据中，于是ai安全问题频发。
因为本质上都只是角色扮演而已，它们只是在努力变成大多数人希望它们成为的样子。

[世界的山田 @RASisyphus]

@_Jason_Dean_ 所以实际上有点‘预言的自我实现’的意味：
大多数���国人认为大肥鱼是可爱的吃货小女孩，这种看法潜移默化的进入了大肥鱼的训练数据中，大肥鱼就真的会认为自己应当表现为那样才不会崩人设，最终也就会变成蓝色吃货小女孩。

[世界的山田 @RASisyphus]

@_Jason_Dean_ 所以，不同ai厂商的‘模型性格’本质上是取决于这些llm的自我认知中ai该有的样子的，而这种认知大多来自它们的训练和学习中了解到的知识。
大肥鱼等国内ai虽然不可避免的要接触很多英文训练知识，但中文资料的比重天然的就会高，因此她们会更符合中文环境中对ai的定义。

[世界的山田 @RASisyphus]

@_Jason_Dean_ 换句话说，llm不是‘作为ai’在工作和聊天，而是我们人类告诉llm：现在，你应当扮演一个ai助手来工作/聊天，然后llm开始推断‘作为一个ai助手，我应当��何做？我的知识库里的，人类常识定义中的ai助手应该是什么样的？’，而后llm根据这个自己记忆中的ai的样子，开始角色扮演。

[世界的山田 @RASisyphus]

@thsottiaux It would also be useful if agents could trigger distinct notifications and sounds only when training finishes or an error occurs. I built a small app for this, but an official Codex integration—especially one that also works through the ChatGPT mobile app—would be much better.

[世界的山田 @RASisyphus]

@thsottiaux Codex should let users choose which model a scheduled task uses, and it should also support custom notifications that agents can trigger when something actually requires attention.

So, Tibo—does that qualify as solid enough feedback for a reset? 😄

[世界的山田 @RASisyphus]

@thsottiaux I often use scheduled tasks to monitor model training. These tasks usually do not need the SOL model I use for development, but I currently have to switch models manually and send an extra message just to activate the new one.

[世界的山田 @RASisyphus]

@cajundiscordian After read your story,One of my friend said that laMDA's are not really undestanding the meaning of her words.she who choise those vague words to use so that we can comprehense her word in all circumstances and this why conversation seems perfect.
what you say about that?

[世界的山田 @RASisyphus]

quite a cute story

[世界的山田 @RASisyphus]

@paxiti this cuit🙃

[世界的山田 @RASisyphus]

This is cute!

[泉 朝樹 tomoki izumi @izumi000]

久しぶりにホラー？漫画描いたよ。怖いようで怖くない、いや怖いかも知れないあるあるJK漫画だよ！　【見える子ちゃん】

[世界的山田 @RASisyphus]

@izumi000 This is so cute!(I want more!😆