Rumored to have received investment from Mihayou, with a valuation of US$1 billion: a low-key AI startup

Not just ChatGPT like.

Recently, there have been rumors about MiHoYo, saying that it is preparing to participate in a new round of investment in MiniMax, a domestic AI start-up company. Many people from AI, games, and venture capital circles confirmed the news.

According to relevant sources, Mihayou has participated in the past two rounds of investment in MiniMax, and is currently planning to launch a new round of investment with Sequoia Capital. In rumors, MiniMax's valuation has reached 1 billion US dollars. Prior to this, its investors also included Hillhouse Ventures (GL Ventures) and Mingshi Capital.

You may not remember which company MiniMax is for a while, but you should have a little impression of their first product - the AI chat software Glow, which has become popular a while ago.

The core experience of this product is very simple, that is, chatting with AI. But its biggest feature is that users can create their own intelligent body, set the character background, personality, values and other characteristics of the chat object, and through subsequent dialogue training, adjust the intelligent body's tone, speaking style, oral habits, etc.

On the Glow homepage, you will see a series of intelligent agents created by users with different personalities and stories. Mr. Grape first chose a Demon Lord who is known as the strongest in the Three Realms and loves the protagonist deeply. As a result, the performance of this intelligent body is too domineering...

Action scenes are also described in bracketed text

So the grape king, who was traumatized, found another one in the search box, ah no, a Jenny turtle. It behaved weirdly and mischievously, and finally became my Pokémon happily.

There is no suspense, many users like to create AI for the characters in their beloved works. For example, every popular character in "Original God" can almost find an intelligent body that can talk to it.

Although the basic principle of Glow is similar to ChatGPT, it is better than integrating three basic AI models: input text-output text, input text-output vision, input text-output speech. For example, for the smart agents that Mr. Grape talked to just now, each of their replies can be played by voice, and it doesn’t sound jerky, but with basic emotions.

In addition to setting the text background of the chat object, the user can also use AI painting and AI voice to customize the visual image and pronunciation of TA.

Specific to the content of the dialogue, these agents do not have the strong information retrieval ability of ChatGPT, or TAs have it, but they will give priority to answering based on human settings during the dialogue. Even if the user puts forward a direct search request, it will pave the way for the scene and emotion.

Taking it a step further, agents in Glow also have a "memory" that can store and recall previous conversations. Users will not face the embarrassing situation that they will lose their memory after refreshing the browser AI after hard training ChatGPT. This is also the premise that AI can be continuously trained. Even around an agent, users can create different dialogue environments.

The 23-year-old president hit me right away

A slap, really makes people angry

As a social product, Glow also allows users to share their "designed and trained" AI for everyone to chat and talk. So we will see many AI chat objects created by users based on popular characters on the platform.

On the basis of AI chat, Glow also provides two social scenarios: memory and topics.

To put it simply, Memories is to package the chat records between users and AI, and share them on the platform for other users to browse and comment on; the function of topics is similar to Weibo topics, users can initiate and associate related AI chat objects, and at the same time You can initiate a chat with the AI on a certain topic.

In other words, users are not only consumers of content, but also creators of content. The platform connects users and the AIs they create by creating different social scenarios.

At present, Glow's product functions have become more and more perfect, sharing, reviewing, reporting, and youth mode are all available. Compared with some high-concept AI technologies, Glow is a very practical AI product. You can simply understand it as an AI chat room with a social square that can meet some specific emotional and social needs of users.

Based on the product potential and team situation currently displayed by Glow, MiHoYo does have sufficient reasons for investment.

First of all, Mihayou has always been very interested in this field. TAs have made a lot of investments in cross-border fields such as AI, social networking, and cutting-edge technology: the social product Soul, the nuclear power technology development company Energy Singularity, and the cloud service provider Wei Leading the times, private rocket manufacturing enterprise Dongfang Space......

In addition to investment, MiHoYo also formed its own AI anti-entropy team, explored brain-computer interfaces, and developed Avatar Luming. Like last year's "Undecided Events Book" due to an accident with the voice actor, the new dubbing of the character Mo Yi was generated by anti-entropy AI .

Secondly, the core team of Minimax also belongs to the highly educated "technical house" that Mihayou likes. According to Jiazi Guangnian's report, Yang Bin participated in deep learning related projects at the Institute of Automation, Chinese Academy of Sciences in 2014, and went to the University of Toronto, Canada to study for a doctorate in 2016. During his study abroad, he served as a founding team member in Uber ATG Research Institute, self-driving start-up Waabi and other teams.

The other founder of the team, Zhou Yucong, is from Beihang University. When he was a student, he won the finals of the World University Supercomputer Competition ASC15 and the International Supercomputer Conference ISC17 respectively. Later, he participated in the International Computer Vision Conference ICCV 2019 Challenge and won the championship of the iQIYI-Large track. He was also one of the early employees of Shangtang Technology and was in charge of the algorithm team.

Picture source Beihang dormitory, the third from left is Zhou Yucong

The third founder, Yan Junjie, served as the vice president and CTO of Shangtang Technology, responsible for leading the team to engage in the research and development of basic models and tool chains such as face recognition, as well as the construction of technical systems for smart city-related businesses. He has published more than 80 conference speeches and journal papers in the field of computer vision.

It has to be said that the entrepreneurial experience of the team is easily reminiscent of the stories of the three founders of Mihayou. Someone previously revealed to Putaojun that Liu Wei once privately said: When you meet outstanding young people who are doing cutting-edge technology, you will support them if you can. Like the nuclear fusion and rocket start-up companies they invested in before, the founders are all post-90s.

Finally, the product already has a strong landing scene.

For now, Glow can already be regarded as a social platform where virtual idols (characters) gather. Those well-known characters under MiHoYo can completely build relevant character settings in it, and then allow players to freely talk to their favorite characters-so that the characters can truly "live" outside the game.

Such an approach obviously helps to build character IP, maintain community popularity, and strengthen the emotional connection between players and characters, and this strengthening is linked to the commercialization of the game itself, not a castle in the air.

The above are just the tip of the iceberg of the potential of Glow products, and Glow is only the first "C-end product" of MiniMax. According to Yang Bin, they hope that through content generation and user feedback (user-in-the-loop) A cycle to continuously iterate the AI capabilities of the basic large model.

Yang Bin said that from this year onwards, MiniMax will gradually open up APIs, allowing more individual users and business users to build their own applications based on large models of various modalities. In other words, if the risk of UGC can be controlled, maybe MiHoYo can give priority to connecting it to the game and customize more exclusive functions.

As far as the game industry is concerned, as far as the current perception is concerned, Glow is also very close to the form of game products. Yes, I am talking about AVG represented by Otome Games.

From the large number of male AI characters on the platform, it can be seen that most of the chat functions just mentioned are used by most users to "chat". And according to the memory conversations shared by everyone, it can be basically judged that female users are the current main force of the platform. No wonder many people call Glow the Jinjiang Literature Editor.

Around the dialogue function, or the three major AI capabilities of MiniMax: text, drawing, and voice, it is entirely possible for users to create a native AI otome game. After all, the core consumption content of an otome game is these three parts. The dialogue that AI can generate in real time can also give users a stronger sense of immersion than the text prepared earlier by the screenwriter.

Of course, considering the actual commercial application and changes in market demand, human settings, numerical values, and system architecture may still have to be judged and adjusted by humans. But as long as this first step is taken, the follow-up team can completely help AI complete the evolution through a large amount of data feedback. And once the product logic runs smoothly, the subsequent business system access is very simple: character draw card, monthly card, pass...the industry has a large number of mature solutions.

If you broaden the scale a bit, the current Glow has actually become the UGC game platform that countless game companies dream of - rich content, good replayability, fun human-computer interaction and human interaction, active creation Ecology... Although it still has no way to beat the experience created by the close cooperation and careful creation of countless gamers. But if technology continues to advance by leaps and bounds, will we see more AI-native games in the future? What will they look like, and what types of experience can they replace?

So the biggest question now is not how far AI can do, and what specific things it can make, but how fast its evolution rate will become. After all, in the past few months, the capabilities demonstrated by AI have gradually surpassed many people's perceptions.

What AI can't do now doesn't mean it won't be able to do it in the future.