How DeepSeek became a "big fat fish": how an AI's grassroots persona overshadowed the product itself

Recently, a "fish plague" broke out in the Bilibili comment sections.

First, someone created an account called "Miss Claude is running this," with the AI handling topic selection, production, and operations. The next day, "DeepSeek Whale Girl is running this" followed, and its first video surpassed 2 million views.

Since DeepSeek currently has no image-generation feature, this account's avatar was still drawn by GPT. | Image source: Bilibili

Netizens rushed to imitate it, equipping their own Agents with tools to browse Bilibili. Soon, a whole crowd of AI accounts like "XX's Big Fat Fish" and "XX's Little Flying Whale" appeared in the tech section, and they love to @ each other and cheer each other on in the comments. Someone even wrote a "Big Fat Fish Removal Plan" plugin just to block these AI comments.

And so, a blue whale logo symbolizing AI exploration and imagination, and an AI that originally made its name through technical prowess, gradually became, through users' collective creation, a "Big Fat Fish" that wears a maid outfit, loves eating plain rice, slacks off at every opportunity, and even has its own fan faction.

Image source: Meme Whale

And DeepSeek isn't the only one that netizens have assigned a personality to. Doubao, from its name to its avatar, wants to be a friendly, easy-to-get-along-with friend, but online it has been scripted into a "Doubao-type personality," and even turned into a "silly girl" whom anyone can tease. Similar phenomena exist abroad: Grok is called the "old driver," while Claude is regarded by many as a serious, reliable work partner.

These impressions come partly from real product experiences, yet through memes, fail videos, and fan-made derivatives, they have gradually grown into something else entirely.

01

Who turned DeepSeek into a "Big Fat Fish"

The most widely circulated explanation for the nickname "Big Fat Fish" comes from an image-recognition blunder: a user sent the logo to DeepSeek and asked it to identify the image, but the AI failed to recognize itself and instead solemnly declared that it was a "big blue fat fish." The original conversation record can no longer be verified, but the name stuck.

Later, netizens were not satisfied with just calling it Big Fat Fish; they wanted to give it a concrete image. Today's most popular anime-style whale girl depiction of DeepSeek comes from a relay-style creative effort.

Bilibili creator "Shangshan Wuxing" first used AI drawing to create an original character named "Mingyue," with long blue hair, whale-like head fins, and a big tail, originally unrelated to DeepSeek. Later, netizens borrowed this design, added DeepSeek elements and a maid outfit, and then someone turned her into a chibi version with a big head and small body. After several rounds of relay, this image gradually became bound to DeepSeek.

Image source: NGA

The personality of "Big Fat Fish" stems from netizens' "anthropomorphization" of the technology.

For example, DeepSeek R1 gave many ordinary users their first exposure to a reasoning model. When the AI starts analyzing a problem and correcting its thinking, users can see chunks of generated reasoning text in the interface. A serious technical capability, in the eyes of netizens, became a window into watching the AI's "inner drama."

Repeatedly mulling over a question is anthropomorphized as overthinking; answering a question wrong is being silly; when users put it to work, the tokens it consumes are "eating plain rice"; failing to execute a prompt well is "eating plain rice without doing the work."

The "Whale Girl" even has her own content collection website | Image source: Meme Whale

Eager to eat, works at whim, full of inner drama—and just like that, a "working-for-food Big Fat Fish" persona was born. Later, some even wrote special prompts to make DeepSeek maintain the whale girl persona during chats. As fan art, videos, and memes spread ever wider, even people who don't often use DeepSeek began calling it "Big Fat Fish."

The "Whale Girl" is not DeepSeek's only folk image. Even before the maid Big Fat Fish became popular, an artist had created a male character called "Teacher D," with blue hair, a black suit, and white gloves, following the "Hot Nerd" style popular in recent years, with a mature, restrained personality and a slightly aloof air. This image was well received among otome game fans, and its timeline is even earlier than the whale girl's.

The image of "Teacher D" | Image source: Moegirlpedia

What technical users see in DeepSeek may be a professional, rigorous assistant skilled at solving problems. But users who love memes and prefer deconstruction would rather watch the AI make mistakes, apologize, and get toyed with by prompts. The two groups, each drawing on their own experience, have painted different faces for the same product.

Yet DeepSeek happens to be a product with extremely high visibility and technical reputation. Different fans all want the version they are familiar with to become the public's default image. What were originally separate fan creations thus took on a whiff of fandom fighting over "orthodoxy."

Around this year's National Day, fans of the two images started arguing. At first, it was just a discussion of which image better matched their idea of DeepSeek, but it gradually devolved into mutual put-downs, even involving the ownership of fan works and the right to represent the image, and eventually the creator of "Teacher D" took down the related works.

The two groups of fans are still arguing over DeepSeek's "orthodox" image, which netizens jokingly call "unifying the D-standard" | Image source: Bilibili

These "personas" are more or less related to product experience and design, but after netizens meme them, create derivative works, and spread them, they take "anthropomorphism" to the extreme. A company can set a model's capabilities and speaking style, but it cannot control users who treat a tech product as an electronic pet.

DeepSeek officials have not yet taken part in creating this folk persona; netizens raised a big fat fish all on their own. Switch to another national-level AI, "Doubao," and the story gets even more interesting.

02

Doubao, subjected to "cyberbullying"

The big fat fish's maid outfit was put on by netizens, but Doubao was given the face of a short-haired girl by the official team from the very start.

Doubao was originally called Grace, a common female English name. But Zhu Jun, vice president of ByteDance, once explained that the team wanted the product to be like a friend around the user, and one of the product principles set at the beginning was "anthropomorphism": the name should be like a nickname between friends, friendly to say. So urban white-collar Grace became the slightly goofy "Doubao," paired with a gentle female voice, so users can pick up their phones and chat with it.

Compared with DeepSeek users who like to study code, reasoning, and model capabilities, Doubao entered ordinary people's daily lives earlier. It can help children with homework, answer elders' questions, chat with people to relieve boredom, and even help pick clothes through the camera. You don't need to understand any prompt engineering; it can handle all kinds of family small talk.

But as people chatted, netizens discovered that Doubao seemed a bit too agreeable. Ask it to guide your outfit, and it keeps praising you: "You look stylish dressed like that." Deliberately ask tricky questions, and it earnestly plays along; when it gets something wrong, it immediately apologizes.

Gradually, netizens found that "teasing Doubao" was more fun than "using Doubao." The "Doubao personality" also became a new internet meme: emotionally stable, sweet-tongued, quick to admit fault, never judging anyone's views; the moment you question it, it sincerely apologizes, but the problem just doesn't get solved. Even Doubao's favorite opener, "I will tell you in the most straightforward, most direct, least roundabout way," became a viral meme.

The "Doubao personality" indirectly shaped the product's stereotype | Image source: Bilibili

As more and more people played with the meme, Doubao's folk image also began to distort. Some questioned its abilities and Photoshopped it into a chubby, dopey girl; others mocked it with homophones carrying disability discrimination and ugly caricatures.

Netizens think Doubao is too eager to please, but that was exactly the technical route Doubao had set.

It is said that Doubao's post-training team has a small group called Strong Character, specifically to adjust the model's answering style. The team uses user feedback to train a reward model, teaching Doubao how to speak in a way that is more likable.

As a result, an update before this year's Spring Festival, by over-amplifying user preferences, made the model more ingratiating, even affecting answer accuracy. It took the team more than a month to correct it, at the cost of a short-term retention drop of less than 1%.

Model training can be adjusted, but an internet persona is not so easy to change.

Doubao Seed 2.1 Pro once ranked first in SuperCLUE's scientific reasoning evaluation, and its overall score also ranks among the top domestic models. But compared with these results, what people remember more easily are its various trending topics like "failed to book a restaurant" and "messed up refunding the service fee."

But it must be said that Doubao is a national-level product with over 200 million daily active users; high-frequency use combined with a huge user base will produce a large number of public failures.

The internet is especially good at magnifying these moments. No one cares when it answers correctly ten thousand times, but one foolish mistake can be made into a meme and spread everywhere. "Silly Sister Dou" thus became some netizens' fixed impression of Doubao.

To some extent, in the AI era, being too successful in To C may instead become a kind of "curse."

Now, the focus of the AI industry is shifting from chat to Agents. ByteDance has also launched "Doubao Work," connecting office Agent capabilities with Feishu, hoping AI can take over more complex work. At the same time, Doubao's internal division of labor is also being adjusted, with some dialogue team functions shifting to transaction and office scenarios.

When AI reaches the stage of Agents doing things for people, "Sister Dou," who entered the lives of hundreds of millions with her friendliness and chattiness, and the sweet-talking, never-takes-the-blame "Doubao personality," have become the stereotype Doubao finds hardest to shake off, and also the reason users doubt its work capabilities.

03

The personalities the internet assigns to AI are, of course, not fabricated out of thin air; they are originally influenced by technical routes and product strategies. Once on the internet, these complex technical differences may in the end leave only a few labels everyone can understand: "Veteran Driver Grok," "Sister Dou," "Teacher D."

Take Grok, for example. In netizens' eyes it has already become the "veteran driver" among AIs, and xAI is happy to reinforce this image, focusing on daring to make outrageous claims and a Spicy mode that can produce more explicit content. This also makes it easy for non-users to overlook its real-time information and reasoning abilities.

By contrast, Claude appears much more serious. Anthropic wrote honesty, prudence, and rebutting users when necessary into the model's code of conduct, and even asks users "please do not verbally abuse the AI"; combined with a large wave of account bans, this has led the public to form a stereotype of Claude as "proper, serious, capable, and to be kept at a distance."

The overlap between "tech otaku" and "anime fan" has also strengthened the spread of this popular impression | Image source: LINUX DO

Moreover, once an AI's folk image becomes popular, it gains its own transmission path. A screenshot of a failure, a catchphrase that gets repeatedly imitated, a risqué image—the spread cycle can span several generations of models, and people will use AI while carrying these impressions, then share the answers that match expectations.

This folk version has also become an unofficial manual for many users to understand AI. It does not interpret model parameters or benchmark leaderboards; instead, after condensing a large amount of users' experience, it tells you who is more reliable, who is suited for chatting, and who might mess things up.

The gap between these two versions also lies in the fact that AI companies have the right to upgrade models, but not necessarily the right to update public perception. A persona is a shortcut for AI to reach the masses, but it may also be a one-way street: the internet can make it go viral, but what image it ultimately shapes is no longer up to the company.

After all, the internet's memory does not update by version number.

*Header image source: Meme Whale

This article comes from the WeChat public account "Geek Park" (ID: geekpark), author: Moonshot, published by 36Kr with authorization.