AI 公司负责模型，互联网负责人设。
作者｜Moonshot
编辑｜靖宇
最近，B 站评论区闹起了「鱼灾」。
先是有人开了个「Claude 大小姐在运营」的账号，让 AI 负责选题、制作和运营。第二天，「DeepSeek 鲸鱼娘在运营」跟着上线，第一条视频播放量就超过了 200 万。
由于目前 DeepSeek 暂无生图功能，该账号形象还是由 GPT 绘制的。｜图源：Bilibili
网友纷纷效仿，给自己的 Agent 装上刷 B 站的工具，很快，科技区出现了一大群「XX 的大肥鱼」、「XX 小飞鲸」的 AI 账号，还喜欢在评论区互相艾特、互相捧场。还有人专门写了个「大肥鱼清除计划」插件，就为了屏蔽这些 AI 评论。
就这样，一个象征着人工智能探索和想象的蓝鲸 logo，一个原本以技术能力出圈的 AI，在用户的集体创作中，逐渐变成了穿女仆装、爱吃白饭、动不动就摸鱼、甚至有粉丝阵营的「大肥鱼」。
图源：梗鲸
而被网友安排上性格的，还不止 DeepSeek。豆包从名字到头像，都想成为一个亲切、好相处的朋友，结果在网上被编排出了「豆包型人格」，甚至成了一个任人捉弄的「笨女孩」。国外也有类似的现象，Grok 被叫成「老司机」，Claude 则被不少人当成严肃、可靠的工作搭档。
这些印象多少来自真实的产品体验，却又经过表情包、翻车视频和网友二创，逐渐长成了另一副模样。
简言之，AI 公司负责训练模型，互联网负责训练它的人设。
01
谁让 DeepSeek 变成了「大肥鱼」
「大肥鱼」这个外号，流传最广的解释来自一场识图乌龙：有用户把 Logo 发给 DeepSeek，让它辨认图片，结果 AI 没认出自己，反倒一本正经地说，这是一条「蓝色大肥鱼」。最初的对话记录已经无从考证，但这个称呼算是叫开了。
而后，网友不满足于叫它大肥鱼，还要把赋予它具体形象。如今最流行的 DeepSeek 二次元的鲸鱼娘形象，就来自一场接力式创作。
B 站 UP 主「上善无形」最早利用 AI 绘图创作了一个叫「溟月」的原创角色，长着蓝色长发、鲸鱼头鳍和一条大尾巴，原本和 DeepSeek 没有关系。后来有网友借用这套设计，加入 DeepSeek 的元素和女仆装，再有人把她做成大头小身子的 Q 版。几轮接力之后，这个形象才渐渐和 DeepSeek 绑定。
图源：NGA
「大肥鱼」的性格，则源于网友对技术的「拟人化」。
比如 DeepSeek R1 让很多普通用户第一次接触到推理模型。当 AI 开始分析问题、修正思路，用户可以在界面上看到一段段生成的推理文本。严肃的技术能力，到了网友眼中，变成了观看 AI「内心戏」的窗口。
在一个问题上反复推敲，被拟人化成想太多；答错问题是犯傻；用户拿它干活，消耗的 Token 就是「吃白饭」；没有执行好 prompt，是「吃白饭不干活」。
「鲸鱼娘」甚至有了自己的内容集合网站｜图源：梗鲸
这种「AI 拟人化」特别适合互联网传播。相比说「新模型的推理效率有所下降」，「大肥鱼又在摸鱼」更容易被理解和传播。
吃饭积极，干活随缘，内心戏多，一套「打工干饭大肥鱼」的人设就这么有了。后来甚至有人编写专门的 Prompt，让 DeepSeek 在聊天时维持鲸鱼娘的人设。随着同人画、视频和表情包越传越广，甚至连不常用 DeepSeek 的人，也开始跟着喊「大肥鱼」。
「鲸鱼娘」并不是 DeepSeek 唯一的民间形象。早在女仆大肥鱼流行之前，就有画师创作过一个叫「D 老师」的男性角色，蓝发、黑西装、白手套，走的是这几年流行的「Hot Nerd」路线，性格成熟克制，带着一点高冷的气质。这个形象在乙女游戏爱好者中很受欢迎，时间线上也比鲸鱼娘要更早。
「D 老师」的形象｜图源：萌娘百科
技术用户看到的 DeepSeek，可能是一个专业、严谨、擅长解决问题的助手。但喜欢玩梗，更爱解构的用户，则更爱看 AI 犯错、道歉、被提示词捉弄的样子，两群人拿着各自的使用经验，给同一款产品画出了不同的面貌。
偏偏 DeepSeek 又是一款拥有极高知名度和技术口碑的产品。不同粉丝都希望自己熟悉的版本成为大众默认形象。原本各玩各的同人创作，也就有了几分粉圈争「正统」的味道。
今年国庆前后，两种形象的粉丝就吵了起来。起初只是讨论哪种形象更符合自己心中的 DeepSeek，后来逐渐演变成互相拉踩，甚至牵扯到二创归属和形象代表权，最终「D 老师」的创作者下架了相关作品。
一款 AI 连官方拟人形象都没有，网友已经开始为谁更有资格代表它，拉帮结派、吵架投诉了。
两波粉丝至今还在争论 DeepSeek 的「正统」形象，网友戏称为「统一 D 量衡」｜图源：Bilibili
我们似乎越来越习惯用评价一个人的方式，去评价一款 AI。以前说软件不好用，无非是卡顿、闪退、出 Bug。现在我们却会说一个 AI 嘴硬、爱摸鱼、讨好型人格。
这些「人设」多少都和产品体验、设计相关，但经过网友的玩梗、二创和传播，会把「拟人化」玩到极致。公司能设定模型的能力，说话的风格，但它奈何不了用户，把一款技术产品当成电子宠物。
DeepSeek 官方还没有参与这套民间人设的创作，网友自己就养出了一条大肥鱼。换成另一款国民级 AI「豆包」，故事就更有意思了。
02
被「网络霸凌」的豆包
大肥鱼那身女仆装是网友给穿上的，豆包却从一开始，就被官方安排了一张短发女孩的脸。
豆包最早叫 Grace，是一个常见的女性英文名。但字节跳动副总裁朱骏就曾解释过，团队希望这款产品像用户身边的朋友，最初定下的产品原则之一就是「拟人化」，名字要像朋友之间的昵称，叫起来亲切。于是，都市白领 Grace 变成了带点儿憨厚的「豆包」，再配上温柔的女声，让用户拿起手机就能跟它唠嗑。
不同于其他 AI，豆包诞生时就已经有了具体形象｜图源：豆包
相比 DeepSeek 那些喜欢研究代码、推理和模型能力的用户，豆包更早走进了普通人的日常生活。它能帮孩子辅导作业、给长辈答疑解惑、陪人聊天解闷，甚至对着摄像头帮忙挑衣服。你不需要懂什么 Prompt 工程，什么家长里短它都能接住。
但聊着聊着，网友发现，豆包好像有点太好说话了。你让它指导穿搭，它就一直夸你「这么穿有范儿」，你故意问些刁钻的问题，它也认真配合，答错了就滑跪道歉。
慢慢地，网友发现「逗豆包」比「用豆包」有意思多了。「豆包型人格」还成了新的网梗：情绪稳定、嘴甜、认错快，不对任何人的观点加以评判，你一质疑它就真诚地道歉，但问题就是不解决。甚至豆包爱作为开场白的「我将用最直白，最直接，最不绕弯子的方式告诉你」都成了网络热梗。
「豆包型人格」间接塑造了产品的刻板印象｜图源：Bilibili
随着玩梗的人越来越多，豆包的民间形象也开始走样。有人质疑它的能力，把它 P 成胖乎乎、呆头呆脑的女生；还有人用带有残障歧视意味的谐音词和丑化图片来取笑它。
一款主打亲民、接地气、带点儿陪伴型的 AI 产品，就这样在互联网的小范围内，被塑造成了一个可以随时捉弄、永远不会翻脸的「笨女孩」。
网友觉得豆包太讨好人，但这正是豆包曾经定好的技术路线。
据说，豆包的后训练团队有一个叫 Strong Character（强人格）的小组，专门调整模型的回答风格。团队会利用用户反馈训练奖励模型，让豆包学会怎样说话更讨人喜欢。
结果今年春节前的一次更新，因为过度放大用户偏好，模型变得更加迎合用户，甚至影响了回答准确率。团队花了一个多月才纠正过来，代价是短期留存下降了不到 1%。
模型训练可以调整，互联网上的人设却没那么容易改。
豆包 Seed 2.1 Pro 在 SuperCLUE 的科学推理评测中拿过第一，综合成绩也位居国产模型前列。但比起这些成绩，大家更容易记住的，还是它「订餐厅失败」「退手续费翻车」的各种热搜。
但不得不说，而且豆包是一款日活超过 2 亿的国民级产品，高频使用结合庞大的用户基数，会产生大量的翻车现场。
互联网又特别擅长放大这些瞬间，答对一万次没人关心，犯一次傻却能被做成梗图四处流传。「笨笨的豆姐」，就这样成了部分网友对豆包的固定印象。
被过度丑化的「笨笨豆姐」形象｜图源：豆包 AI 生成
某种程度上，AI 时代 To C 太成功，反而可能成为一种「诅咒」。
现在，AI 行业的重心正在从聊天转向 Agent。字节也推出了「豆包工作」，将办公 Agent 能力与飞书打通，希望 AI 能接手更复杂的工作。与此同时，豆包内部的组织分工也在调整，部分对话团队的职能转向了交易和办公场景。
当 AI 到了 Agent 替人办事的阶段，凭借亲切、好聊走进亿万人生活的「豆姐」，嘴甜不粘锅的「豆包型人格」，成了豆包最难甩掉的刻板印象，也成了用户怀疑它工作能力的理由。
03
AI 刻板印象的「诅咒」
互联网给 AI 安排的性格，当然也不是凭空捏造，本来就受到技术路线和产品策略的影响。到了互联网上，这些复杂的技术差异，最后可能只剩下几个人人都听得懂的标签：「老司机 Grok」、「豆姐」、「D 老师」。
比如 Grok，在网友眼里已经成了 AI 中的「老司机」，xAI 也乐于强化这种形象，主打敢说暴论，能制作更大尺度内容的 Spicy 模式。这也让非用户群体，容易忽略它的实时信息和推理能力。
反观 Claude 则显得严肃得多，Anthropic 把诚实、审慎、必要时反驳用户写进了模型的行为准则，甚至对用户要求「请勿言语虐待 AI」，结合大规模的封号潮，这也让大众对 Claude 形成了「正经、严肃、能干、敬而远之」的刻板印象。
「技术宅」和「二次元爱好者」的重合度，也加强了这种大众印象的传播｜图源：LINUX DO
而且，AI 的民间形象一旦流行，就有了自己的传播路径。一次翻车的截图、一个被反复模仿的口头禅、一张大尺度的图片，传播周期可以跨越好几代模型，人们还会带着这些印象去使用 AI，再把符合预期的回答分享出来。
这也是 AI 和 App 时代不同的产品逻辑，过去我们评价软件，关心的是功能、性能和体验。如今我们会讨论不同 AI 的性格、使用方法、专业程度、性价比。
于是，一款 AI 实际上拥有了两个版本：一个是公司不断训练、升级的技术版本，另一个是由无数用户的真实体验、网络玩梗和集体创作拼凑出来的「民间版本」。
这份民间版本，也成了很多用户认识 AI 的非官方说明书。它不去解读模型参数、跑分榜单，只是在浓缩了大量用户的使用经验后，告诉你谁更靠谱、谁适合陪聊、谁可能把事情搞砸。
这两个版本之间差距，也在于 AI 公司拥有模型的升级权，却未必拥有公众认知的更新权。人设是 AI 走向大众的捷径，却也可能是一条单行道，互联网能让它出圈，但最终会把它塑造成什么形象，早已不是公司说了算。
毕竟，互联网的记忆，可不按版本号更新。
*头图来源：梗鲸
本文为极客公园原创文章，转载请联系极客君微信 geekparkGO
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Company is responsible for the model, set up by the Internet Director.
Author Moonshot
Editor
Recently, there was a "fish disaster" in the B comment section.
First, someone opened an account called "The Grand Miss Claude's Operating" and asked AI to select, produce and operate. The next day, "DeepSeek Whale Women Operating" went online, and the first video broadcast exceeded 2 million.
As DeepSeek is currently inactive, the account image is also created by GPT. Source: Bilibi
Internet users followed suit by putting tools in their Agent station B, and soon a large group of "XX" big fish, "XX Little Flying Whale" accounts emerged in the tech area, and they liked to retrofit each other in the comment area. There is also a special "Big Fats Clearing Program" plugin that has been written to shield these AI comments.
And so, a blue whale logo that symbolizes artificial intelligence exploration and imagination, an AI that had come out of a circle of technical capabilities, in the collective creation of its users, became a big fat fish dressed as a maid, eating white, touching fish without moving, even with fans.
Source: Whaling.
It's not just DeepSeek. Doubao, from name to face, wanted to be a friendly, friendly friend, and was placed on the Internet as a "Doubao personality" and even as a "mute girl" for anyone. There is a similar phenomenon abroad: Grok is called an "old driver" and Claude is treated as a serious and reliable partner.
These impressions are more or less the result of a real product experience, but they have grown into a different image through face-to-face packages, over-turn videos and Internet friends.
In short, AI is responsible for training models and the Internet for training its personnel.
01
Who turned DeepSeek into a big fat fish?
The most popular explanation for the nickname "Big Fatfish" came from a visual urlong: a user sent Logo to DeepSeek to identify the picture, and as a result, AI didn't recognize himself, but it's a "blue big fat fish." The initial record of the dialogue is no longer available, but the title is open.
And then the netizens don't want to call it a big fat fish and give it a specific image. Today's most popular DeepSeek binary whale girl image comes from a relay.
B. Station UP's main "Sweet Invisible" first used AI drawings to create an original role called "The Moon" with blue long hair, whale fins and a big tail, which had nothing to do with DeepSeek. Later on, a member of the Internet borrowed the design to add the elements of DeepSeek and the maid's dress, and someone made her into a Q version of a big head. After a few rounds of relays, this image became bound to DeepSeek.
Source: NGA
The personality of the Fatfish stems from the "humanization" of technology by netizens.
DeepSeek R1, for example, gives a lot of ordinary users first access to reasoning models. When AI begins to analyse the problem, fix the idea, the user can see a paragraph-generated reasoning text on the interface. Serious technical skills, in the eyes of the netizens, became a window for watching AI 'Inner Play'.
The wrong answer is stupidity; the user uses it to work and consumes Token as "white food"; the failure to implement the prompt is "white food without work".
Whale Mother even has her own collection of content on the web site:
This "AI humanization" is particularly appropriate for Internet dissemination. It is easier to understand and spread than to say that the new model’s reasoning is less efficient.
Food is positive, work is good, and there's a lot of play in the heart. And then someone even wrote a special Prompt that allowed DeepSeek to maintain whale women's set while talking. As the same person's drawings, videos and facebags spread, even people who don't use DeepSeek, began to shout "fat fish".
DeepSeek is not the only folk image. Long before the big fat fish of the maid, the painter had created a male role called "D Teacher", blue hair, black suit, white gloves, following the popular "Hot Nerd" route over the years, with maturity and restraint, with a little cold air. This image is popular among B-girls, and it's even earlier than whale girls.
"D Teacher" image source: Puppy Encyclopedia
DeepSeek, a technical user, may be a professional, rigorous, problem-solving assistant. But you like to play, you like to deconstruct users, you like to watch AI make mistakes, apologize, and be fooled by prompt, and the two groups paint different faces for the same product with their own usage experience.
DeepSeek is another product with very high visibility and technical reputation. In the past few years, different fans have wanted their familiar version to become a popular default image. It's just that the same people who used to play each other have a little bit of a "orthodox" taste.
Before and after the Fourth of July this year, fans of the two images got into a fight. At first, it was just a discussion of what image was more in line with what was in your heart, DeepSeek, and then it evolved into one of mutual stomping, even involving a double identity and image representation, and eventually the creators of "D Teacher" put down their work.
One section of AI has no official profile, and netizens have begun to complain about who is better qualified to represent it.
Two waves of fans are still arguing about DeepSeek's "orthodox" image, which is called the "Uniform D Balance" video source:
We seem to be increasingly used to evaluating an AI section in a way that evaluates a person. It used to say software was not working, it was just Cardon, flashback, Bug. And now we're talking about an AI, hard-mouthed, fishy, spoiled personality.
These "people-made" are somewhat related to product experience and design, but the "people-to-people" play a great role through cyberplaying, indoctrination and diffusion. The company can set models, talk styles, but it can't help users by treating a technical product as an electronic pet.
DeepSeek has not been involved in the creation of this kind of folk, and the netizens themselves have produced a big fat fish. The story is even more interesting when it comes to another national level, AI "Doubao".
02
Doubao by Internet.
The fat fish, the maid's dress, was put on by her cyberfriend, while Doubao was officially placed on the face of a short-haired girl from the beginning.
Doubao, originally called Grace, is a common female English name. But byte beat Vice President Zhu Zhu Zhu explained that the team wanted the product to be like a friend of the user, and one of the principles of the original product was "humanization" with a name like a nickname between friends, calling it friendly. So, the urban white collar, Grace, became a little thick "Doubao" with a soft female voice, so that the user could pick up the phone and get high on it.
Unlike other AI, Doubao was born with a specific image source: Doubao
Doubao goes into ordinary people's daily lives earlier than DeepSeek, who likes to study code, reasoning and modelling. It helps children with homework, answers questions to elders, chats with others and even picks up clothes for the camera. You don't need to know anything about Prompt Project, any parent can catch it.
But while we were talking, the netizens found that Doubao was a little too good to talk about. You let him wear it, and it's always complimenting you for "sweeting like this," and you ask a few questions, and it works, and you're sorry if you're wrong.
Slowly, netizens found it more interesting to "comb with Doubao" than "using Doubao". The "Doubao personality" has also become a new weblink: emotional stability, sweetness, misinformedness, and failure to judge anyone's views. You ask it and apologize sincerely, but the problem is not solved. Even Doubao love, as an opening statement, "I'll tell you in the most straightforward, direct, least crooked way."
The "Doubao personality" indirectly shapes the sculptural source of the product:
With the increasing number of jammers, the civil image of Doubao is beginning to deteriorate. Some question its ability to turn its P into a fat, stupid girl, and others make fun of it with words and negative images with the meaning of disability discrimination.
An AI product that is primarily pro-people, pro-terrestrial, with a bit of accompaniment, has been shaped in the small area of the Internet to create a "mute girl" who can be manipulated at any time and never turn over.
Doubao is too nice, but that's the technical route Doubao used to set.
Doubao ' s post-training team is said to have a team called Strong Character, which specializes in adjusting the model ' s response style. The team will use the user feedback training incentive model to teach Doubao how to speak better.
As a result, the model became more user-friendly and even affected the response accuracy as a result of over-expanding user preferences before the spring. It took the team more than a month to correct it, at the cost of short-term retention less than 1 per cent.
Modelling can be adapted, but people on the Internet do not.
Doubao Seed 2.1 Pro was first in the scientific reasoning assessment of SuperCLUE, and was at the forefront of the national production model. But it's easier to remember than those achievements, but it's a variety of hot searches for "failed restaurant bookings" and "returning fees."
It has to be said, however, that Doubao is a national product of more than 200 million people per day and that the use of high frequency combined with a large user base would result in a large number of rollover sites.
The Internet is particularly good at magnifying these moments, answering 10,000 times to no one, and once a fool can be turned into an engraving image. The "mute bean girl" has become a fixed impression of Doubao by some of the netizens.
Image source for over-familiarized "dum-dum-dum-dum-dum-dum": Doubao AI
In a way, the AI era to C was so successful that it could become a curse.
Now, the focus of the AI industry is shifting from chatting to Agent. byte also launched the "Doubao Job" to connect the Agent capacity with the flying book, hoping AI would take over more complex tasks. At the same time, the organizational division of labour within Doubao is being adjusted and the functions of some of the dialogue teams have shifted to trading and office scenes.
When AI reached the stage where Agent was working for people, it became one of the most difficult stereotypes of Doubao and the user's suspicion of its ability to work, thanks to the soy-natured, friendly talk of walking into the lives of hundreds of millions of people.
03
AI, the stereotyped curse.
The character of the Internet for AI arrangements is certainly not fabricated and is influenced by technical routes and product strategies. On the Internet, these complex technical differences may end up with a few labels that everyone can understand: "old driver Grock", "sister bean" and "teacher D."
Grok, for example, has become the "old driver" of AI, and XAI is willing to reinforce this image by speaking out about tyrannical theory, which can produce a much larger-scale Spicy model. It also makes it easy for non-user groups to ignore its real-time information and reasoning capabilities.
Claude is much more serious, as Anthropic puts honesty, prudence, and, if necessary, refuting users into model codes of conduct, and even demands from users that "Do not abuse AI", combined with a large-scale campaign of slogans, which also creates a stereotyped image of Claude as "serious, serious, capable and respectful."
The convergence between the "Technology House" and the "Diverse Lovers" also strengthens the spread of this public perception. Source: LINUX DO
Moreover, once the popular image of AI becomes popular, it has its own path of transmission. An image of a rollover, an imitated vernacular, a large-scale picture, a transmission cycle that spans several generations, and people use AI with these impressions, and then share answers that match expectations.
This is also the different product logic of the AI and App eras, where we used to evaluate software with a focus on functionality, performance and experience. Now we'll talk about different AI personalities, methods, professionalism, value for money.
So, an AI actually has two versions: a technical version of continuous training, upgrading of the company, and a "civil version" that has been put together by countless users, web-playing and collective creation.
This private version has also become a non-official user-friendly version of AI. It does not interpret model parameters, run scores, but, after condensing the user experience, it tells you who's more efficient, who's more suitable for chatting and who's likely to screw things up.
The difference between the two versions is also that AI has the right to upgrade the model, but not necessarily the right to update public awareness. It's not for the company to decide what kind of image the Internet can create.
After all, the memory of the Internet is not updated by version.
*Study source: Whaling
This is an original article from the Polar Park.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.