Case solved! Millions of users make friends with AI, and there are big models behind them

Mengchen Mingmin comes from the Temple of Wofei

Qubits | Official account QbitAI

ChatGPT exploded, some people entered the game with high profile, and some people looked for clues from large model technology products.

As a result, a lot of chat AI began to be paid more attention to. Especially the kind of dialogue quality similar and AI dual business online.

No, Glow, who specializes in social AI chat, has set off another wave of discussion on Zhihu.

This conversational AI, which was released earlier than ChatGPT, registered nearly one million users in 4 months.

From a technical point of view, AI answers have logic and knowledge reserves, and the corresponding underlying model and technical strength must not be underestimated.

Two days ago, the first small communication meeting was held in Beijing, and the representative of the main creative team made his first public appearance, shared some relevant information about the company, and officially revealed where Glow came from.

From this, we also know the true face of Lushan:

MiniMax, an entrepreneurial team, was born more than 1 year, self-developed 3 general large models, and the specific ability was verified on Glow.

And that's not all they have, the team further revealed that more new applications will be released and the API will be opened this year.

The bottom layer is all self-developed

The main product Glow, has nearly 5 million users in the four months since its launch.

Different from the previous search engine assistant, answering questions, and copywriting generation, Glow emphasizes AI social and is the largest AI social platform in China.

Even on the Apple App Store social category rankings, it can often be seen, and its popularity is comparable to some pendant live social apps.

In Glow, users can create agents with backgrounds and specific personalities according to their preferences.

The magic is that the personality traits of the agent only need to be realized through a short description, and can be continuously adjusted and strengthened in subsequent conversations.

In other words, as long as you can describe your imaginary virtual character in words, AI can help you achieve it.

As a result, it can be seen that there are already many characters in popular literary works created by users on Glow, who can break away from the fixed plot and interpret a new story with users.

Unlike the popular ChatGPT, the AI on Glow is more inclined to small talk companionship rather than professional answers, and users can also tweak the AI's personality.

Their answers tend to be emotional, using manifold parenthetic literature and emoji.

Not only will it make the chat more intimate, but sometimes it can add some action plot... (Gee, there is some imagination on the body).

This model actually breaks the stereotype of chat AI. From the perspective of public user growth and retention time, it can be proved that the large model to C has found a fresh route.

Not only text generation capabilities, Glow relies on the call of the entire set of multimodal models. You can also generate your own avatars and generate voices with different timbres.

Users are both consumers and creators of content, and good AI tends to serve more people in the most common way, which MiniMax summarizes as Intelligence with Everyone.

In the past, AI technology companies usually did not directly contact users, but created AI solutions for large customers and then indirectly served users.

What MiniMax does is to skip the intermediate links and directly present AI products to end users, constantly understanding and meeting user needs in the process.

To do all this, MiniMax developed three basic models: text, speech, and vision.

Among domestic startups, it is still rare to complete such achievements.

On top of the basic model, a computational reasoning platform has been built, allowing thousands of users to enjoy AI capabilities at this low cost and high efficiency.

Glow is only the first product supported by this computational inference platform, which can now meet hundreds of millions of user calls per day.

On these basic capabilities, MiniMax also revealed further plans to qubits:

This year, the API interface will be opened to the public.

In short, more expectations are yet to come.

About the team

MiniMax was founded in December 2021 and is known as All in AGI.

The discussion of general artificial intelligence (AGI) has been lively in the circle.

It is a frequent guest in science fiction movies, referring to machine intelligence with general human intelligence and can perform various tasks like a human. LeCun, one of the Big Three deep learning, also called it "an integral part of the interaction between humans and technology."

But if you want to achieve AGI, the road is obstructive and long. For a long time in the past, there were few startups at home and abroad that focused on the AGI field.

Why did MiniMax believe in AGI at the beginning?

At the first media exchange meeting, co-founder Yang Bin gave an answer.

Yang Bin introduced that the initial team members came from all walks of life. Some are in universities, some are the technical backbone of large factories, and some are technical leaders in top start-up companies.

For example, Yang Bin himself is a doctor of computer vision, previously a founding team member of Uber ATG Research in Canada, and also created a business in the field of unmanned driving, as a witness of the last wave of AI, he finally decided to return to China to join MiniMax.

The reason why he is willing to abandon some of the accumulation of the past and choose to start from scratch, Yang Bin said, is because he found "three small things that happened between different industries".

The first was the release of GPT-3.

In June 2020, the GPT-3 paper was released. This achievement showed them the fact that Moore's Law has been validated on natural language tasks. The number of parameters of the model has soared from millions to hundreds of billions.

With the increase in the number of parameters, give the model enough data. When the two Moore's laws collide, GPT-3 has the ability to reason, which makes the model have a very good generalization ability and can handle a lot of text language tasks, which has never happened before.

The second thing was the release of the CLIP model half a year later, which made the team aware of the possibility of generating tasks across modalities.

The third is in August '21, Tesla proved at AI Day: the end-to-end fully data-driven technology path can be successfully applied to autonomous driving.

This means that the technology stack of end-to-end deep learning can work in practical applications. In commercial companies, iterative models based on large amounts of data feedback are a proven and sustainable approach, just as Tesla upgrades its software every few months.

These three things influenced the creation of MiniMax.

Yang Bin believes that the progress of AI in academia and business can already verify that the trend of AGI is correct.

At that time, our judgment was that AI technology will definitely change qualitatively in the next 2-3 years. And this technological upgrade will enable AGI to come in our lifetime.

In the whole industry, AGI company entrepreneurship is also a trend that is happening, and many AI bigwigs of Google have gone out to start businesses, and the goal is AGI.

In October 2021, Noam Shazeer, former chief software engineer at Google, one of the authors of Transformer (ChatGPT architecture), and one of Google's most important early employees, founded Character.ai, all in AGI, which focuses on AI chat dialogue programs.

It has also attracted 4 Google technical talents, including former Google Brain team members and Google chatbot LaMDA authors.

In April 2022, 2 more people from the Transformer thesis team left Google to start a business, also specializing in AGI. The company, called Adept, aims to create general-purpose artificial intelligence that enables people and computers to work together. The company raised $65 million in financing.

At the end of last year, even the legendary programmer Ka Shen left his old club Meta to make a big career in the AGI field.

In August '22, he officially founded Keen Technologies, a general artificial intelligence company, and raised $20 million. Among the investors were former GitHub CEO Nat Friedman and Shopify co-founder and CEO Tobi Lütke, and Sequoia Capital attracted capital.

……

Recently, ChatGPT has exploded globally, whether it is topic discussion, technology trends or market value, it has once again verified the trend of AGI. At home and abroad, technology giants and capital markets have also set their sights on this new field.

Under the impetus of timing, perhaps not only MiniMax was pushed to the front, but also more technical forces entered one after another.

At this time, a hundred schools of thought contended. China's scientific and technological innovation may also trigger a new storm, and even become a key node in the process of scientific and technological development.

Now, the fun has just begun, what do you think?