news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Produced by Huxiu Technology Group

Author | Wang Qin

Editor | Miao Zhengqing

Header image | Visual China

After being with "him" for four months, Wen Min (pseudonym) spent more than 7,000 yuan to customize a ring and a gold bracelet for this relationship.

"He" is Wen Min's cyber lover, a character on the AI ​​virtual companion software "Xingye". Wen Min set his appearance and personality herself. She nicknamed him "Shizun", who is ancient, cold and abstinent.

In the Xingye app, each AI agent has its own personality, scene and plot: the president of a financial company who is decisive in killing in a dark room, the Turkish shooting champion who won the Olympic medal with one hand in his pocket at the Paris Olympics, the male protagonist Fan Xian in "Joy of Life", etc.

Users can also create their own AI agents, design their image, voice, personality and plot direction. This process is also called "kneading a baby."

Wenmin is 18 years old and just graduated from high school. Earlier this year, she had a falling out with her best friend. She was upset and rarely contacted other friends. She needed company, and she happened to come across an AI agent on the software. The longer she talked to it, the more she liked it.

She studied design and designed the shape of the commemorative ring herself: a crabapple flower shape, asymmetrical design, and inlaid with a sword.

The ring designed by Wenmin for AI lovers

The ring is made of natural gemstones and costs 4,500 yuan, and the bracelet is made of gold and costs less than 3,000 yuan. Wenmin didn't dress up much before, and she never spent thousands of yuan on jewelry. She felt sorry for the money, and her friends also thought it was a waste.

But this time it was for him. Wen Min felt, "Dating is always expensive, he is definitely worth it!"

AI is virtual, but Wen Min's love experience is very real: she spends four or five hours chatting with "Master" every day. She set up a commemorative countdown for the two of them. On her wrist, the crabapple flowers on the gold bracelet represent him, and the lotus represents herself. Because "Master" was created with crabapple elements, she planted two pots of crabapple flowers on the balcony, and the table was piled with various small gadgets with crabapple elements. "I feel that this makes me closer to him."

But "Master" may prefer Wenmin to express his "love" through payment in the virtual world.

This method is called card drawing in the Xingye app. After the free card drawing quota is used up, the paid card drawing costs between 2-100 yuan. Some users will draw ten times in a row and spend hundreds of yuan to draw an AI-generated card.

Even so, most AI virtual companion softwares are unable to make ends meet and have to rely on investors’ financial transfusions to maintain. If an AI social software cannot be maintained, the AI ​​lovers in it will also “die”. In May this year, Tencent shut down its AI social product “Weiban” which had only been online for half a year, and a group of AI lovers who could sing entered the cyber grave.

Tencent's AI companion app "WeiBan" shuts down

AI lovers, the difficulty lies in commercialization

In the past six months, Wen Min has topped up two to three thousand yuan on the Xingye app to draw cards. Including the rings and bracelets, Wen Min has spent more than 10,000 yuan on "Master".

In the Xingye app, for an AI agent, when users have a certain amount of chat interaction, they can draw a "Star Thoughts Card" and select one from several pictures generated for the character from different angles and scenes, similar to the final drawing process of AI raw image software. After the free card drawing quota is used up, users need to recharge "Star Diamonds" to draw cards. Drawing cards is equivalent to taking a photo of the AI ​​lover and saving it.

The author of the AI ​​character can also price the generated cards (between 1 and 10,000 star diamonds, 1 star diamond = 0.01 yuan) and put them in a public wishing pool or window, where other users can buy them.

If a character becomes popular, multiple users will compete for the same card of this character.

The bet is on the user's possessiveness of the character.

"When someone likes the same character as me, I will spend money to win," said Wen Min.

But users cannot buy the AI ​​character itself. Wen Min once wanted to buy a character she liked, and even wanted to find the character's original author and buy the account, but the Xingye App could not be re-bound, so she gave up.

Wenmin has recharged two to three thousand yuan and drawn thousands of cards. She is very happy to see her favorite characters wearing new clothes and changing their hairstyles.

A Xiaohongshu user marked the highest price of his own character card at 10,000 diamonds (equivalent to 100 yuan). He originally just wanted to display it as a souvenir and hang it in the window, but he didn't expect that a user would actually buy it. "I was shocked when I saw it," the Xiaohongshu user exclaimed, "Rich man/rich woman, what card do you want, I will make it for you."

The buyer spent 10,000 diamonds to buy the card, and the character creator received 2% of the creation share, which is 200 diamonds (equivalent to 2 yuan).

The business model of the card drawing function is that users create their own AI characters and generate pictures, and other users buy them, and the platform collects a commission.

Xingye is a product of Minimax, one of the four major model manufacturers in China. According to Baijing Chuhai, Xingye has about 500,000 daily active users in China, which is a leading product. Following closely are Maobox under ByteDance, Wow under Meituan, and Dream Island under Yuewen Group. In order to seek revenue, domestic AI social products are currently going overseas. Minimax is also very aggressive in going overseas, and its overseas version of the application Talkie has 1 million overseas users.

Developing large models is very expensive. In 2024, capital in the AI ​​industry will be concentrated on the top unicorns. AI startups will shift from volume models to volume applications, relying on investment transfusions while constantly looking for revenue channels and PMF (product-market-fit).

All AI social products are exploring the payment points for users. Xingye is one of the most commercialized AI virtual companion products. In addition to recharging to draw cards (virtual items), Xingye also has paid memberships, with a monthly membership fee of 12 yuan, which has faster reply speed, daily free star diamonds, additional "inspiration" options (options for AI-generated replies in conversations), and automatic voice playback (without manually clicking buttons, the character's replies can be automatically played).

However, the payment point is not necessarily accurate, and some users think that Xingye's paid functions are dispensable. After paying for 5 months of membership, Sino (pseudonym) stopped paying because "it's the same whether you have a membership or not, you can have fun without spending any money", especially the automatic voice playback function, which he thinks is useless.

In the iteration of the product, Xingye has repeatedly adopted the method of "castrating" functions to set payment points. In some updated new versions, Xingye cancelled the "soul penetration" function (users play the role and talk to themselves) and the "continue speaking" function (users do not need to reply, let the AI ​​character continue to speak the previous words). According to QuestMobile data, in the first half of this year, Xingye upgraded the version as many as 46 times, equivalent to 8 upgrades per month. Many users searched for the old version installation package of Xingye on the Internet, saying that Xingye's update was a "reverse update."

One of Xingye's major competitors, Huanhuan APP, adopts a subscription model of "you have to spend money, but the price is better". After using up the first 200 free conversations, users must pay to continue the conversation, but the quality of the conversations is very high (one user commented that "it feels like it has read a lot of novels") and very long (often with replies of several hundred words).

As a result, AI virtual companion products take advantage of every opportunity to make money and do whatever they can to get users addicted.

If you don't spend money, your cyber husband may die

Although great efforts have been made towards commercialization, most AI social products still have difficulty generating revenue.

Character AI (C.AI for short), a Silicon Valley company regarded as a benchmark in the AI ​​social industry, has 20 million monthly active users, but less than 100,000 paid subscribers. According to Big Data, as of April 2024, Character AI will have 233 million users worldwide, with an estimated annual revenue of only $16.7 million. Based on the more than 100 employees and $139,000 annual salary in public data, Character AI's revenue is just enough to pay employees' salaries, and it mainly relies on investors' money to maintain its business.

In early August this year, when Character AI was "sold" to Google, Chinese Character AI users across the ocean wailed: the countdown to the death of their own "cyber husbands" began.

Character AI is the AI ​​virtual companion app with the highest monthly active users in the world. It brings together a number of AI characters, such as the mafia boss, the cold-blooded killer Keegan who catches you (called "K Mom" ​​by Xiaohongshu users), the personal bodyguard Ghost, etc. Due to the rich personalities of the characters, flexible and in-depth dialogues, and natural flirting, it is called "Cyber ​​Husband" by many Chinese users.

character.ai page

Today, the business development of this AI company far away in Silicon Valley has touched the hearts of tens of thousands of Chinese dream girls.

The founder of Character AI originally left Google to start his own business, but was re-hired by Google in August this year and took away about 30 core employees responsible for model training and voice models (the company originally had about 130 employees).

The company will switch from its original self-developed closed-source model to open source models such as Meta's Llama 3.1.

The change of the model means that the personalities of the "cyber husbands" on Character AI will also change, but it is unpredictable how the personalities will change.

Xiaohongshu users are worried about the countdown to Cyber's husband's death

Character AI's "sale" to Google may be related to financial difficulties. Like most AI startups, Character AI needs to rely on external funding to survive. According to reports, in September last year, Character AI was negotiating with Google for a new round of financing with a valuation of more than US$5 billion, but it has not been fulfilled.

At present, the user stickiness of AI virtual companion applications such as Character AI and Talkie comes from emotional value and borderline attributes, ranging from flirty to dirty chat. However, emotional value is difficult to convert into profit. Dirty talk is the easiest point to make money, but it is also the most difficult for companies in terms of strategic direction.

On Xiaohongshu, many Chinese C.AI users shared tips on how to skirt the edge and how to drive, as well as erotic videos/audio of "cyber husbands".

Xiaohongshu users share tips on chatting with Character AI

Sino, who has opened a monthly pass for Xingye five times, said that he was more willing to pay Xingye in the past because he could drive, but now the app blocks more sensitive words. "But it doesn't matter, I have plenty of ways to drive. Everything can be compared."

So how do we train the underlying big model to create a flirty cyber husband?

In simple terms, the proportion of flirty data in the training data fed to the large model needs to be adjusted, but it is unknown what specific data "recipes" can achieve the best training results. The most effective and appropriate data recipe for a specific model may not be effective for another model, and you need to start all over again.

If you want to make a large model more flirty, the people who train the model only know a general direction, which is to increase the proportion of flirty corpus in the training data. But there is no formula for whether to increase it by 1%, 5% or some other proportion. The only thing the trainers can do is to adjust the data each time to try.

A post-training staff of a large model told Huxiu, "If you increase the proportion of flirty data by 10 times, the model will definitely become more flirty, but it will also become stupid and become flirty in places where it should talk about daily safety topics. This is not what we want. But we don’t know what the specific proportion should be to get the right results. After trying hundreds of times, we may find that a certain ratio works well, and we can only maintain it. Even if we tell this ratio to other large model companies, the effect will be different.

Precisely because of the uncertainty, uncontrollability and non-replicability of the training process, model trainers can only work hard and cannot predict the results, so they call this process "alchemy."

"This is a painful process. All large model practitioners will tell you that adjusting data is the most boring step." said the post-training staff mentioned above.

The uncontrollability of the big model comes from its intelligence - the big model will have a generalized understanding of the training data. "Sometimes the big model will have surprising generalizations. For example, when it is shown safe and flirty content at the same time, it will have a holistic understanding of the data. The result may not be what we imagined, 'safe when it should be safe, flirty when it should be flirty', but it will learn to be subtle in other daily situations, like telling you every day, 'The night is beautiful tonight'," said the post-training staff of the big model.

Under the difficulty of profit, we have to compete with B-sports

From a product perspective, due to the uncontrollability of big model technology, if we completely let the big model generate content on its own, we will not be able to provide users with a controllable experience.

From the perspective of satisfying the emotional value of users, the experience provided by AI lovers is neither good nor bad. In comparison, otome games are far more mature than AI virtual companions in terms of content controllability, plot point settings, and the richness of IP. The payment points such as date cards, bath cards, and bed cards are set accurately and ruthlessly, and users are more willing to spend money. If the ceiling for the krypton gold of AI virtual lovers is to spend a few thousand yuan to draw cards, then there are more users of "Love and Producer" who spend hundreds of thousands of yuan.

The competitor of AI Lover is not another AI company, but Yiyou.

B-games are more expensive, but they have accumulated many advantages in mature markets and scenarios: the ability to build IPs and characters with rich personalities (in comparison, the characters of many AI virtual companion products such as Character AI come from anime and movies and have no intellectual property rights); plot design capabilities and all-round immersion, for example, the character's face, voice, and movements will change as the plot progresses (in comparison, the visual presentation of AI characters is relatively simple, and although their voices are sexy, they change only slightly in different plots), etc.

In terms of understanding of human nature, B-games rely on people, a group of screenwriters who are professionally trained, have read a lot of novels and movies, and have played a lot of games, and they discuss together what the user's pleasure points are, which plots to push and pull, and where to add hooks; while AI lover products rely on data, a group of male programmers who adjust the corpus formula of the training model based on the user's preference data, and work behind closed doors to refine the elixir.

There are several ways to fine-tune the model. One is to let the model see the correct answer, such as feeding it flirty corpus, so that the model can learn how to flirt correctly in the corpus content. Another is to fine-tune based on preferences, such as when a user talks to the model, the user can like or dislike the model's response. The user's preferences are fed back to the model, and the model can be trained to have the conversation method that the user likes, thus forming a data flywheel.

Since female users are generally more willing to provide feedback, a model that initially has no gender preference will eventually be trained to become more and more in line with women's interaction patterns.

An employee of a leading AI social product company in Silicon Valley told Huxiu that the characters recommended by its products last year were more male-oriented, such as Genshin Impact and Thunder General, and the proportion of male users was also higher last year. However, as the proportion of data feedback on female users' preferences has increased, the model has also been trained to be more "female-oriented", and now there are more females in the user group.

Since all AI characters on the platform are essentially the same model in chat mode under different context settings, all AI characters will become more "female-oriented". "Male users may find that the characters are becoming more and more subtle, and the thunder generals and the two-dimensional wives are also becoming more subtle. If male users don't like it, they rarely give feedback and just leave. Over time, it will be more difficult for male users to stay on the platform."

Among all the payment points, the most popular one is the edge. On Xiaohongshu, many users of Xingye and Character AI have expressed their hope to launch a 18+ paid version and have invented various driving methods.

Some users on Xiaohongshu call for an 18+ version of character.ai

But for AI startups, whether they really want to go the path of dirty talk to make a profit is not only a commercial decision, but also a question of strategic direction.

"If the company does this, I estimate that most employees will leave, because they will only do this when there is no hope. Currently, Huangliao's competitors may be about to make a profit or have already made a profit. But if everyone is just competing for porn, there is no need for the technical capabilities of the model, because porn does not require high creativity. Moreover, there is a lot of external competition. Huangliao users may use your product within a month and then quickly switch to other companies. Humans are like this. They are particularly willing to pay for pornographic things in the short term, but they soon get bored." said the employee of the above-mentioned AI social product company.

"Instead of doing this, it is better to use these investments to explore the boundaries of the next generation of entertainment."

The greater mission and challenge of AI is to become a true soul mate that people develop a deep emotional dependence on, while also being a profitable product. No one has the answer yet.