The large model of Xiaomi handing in the paper, the new Xiaoai classmate is here to test

Source: Qubit

Author: Meng Chen Mingmin

In Lei Jun's annual speech sharing, he talked about his study experience at Wuhan University, shared his experience of being a good student in 4 years of university and 2 years of completing courses; he talked about being ignited by "Silicon Valley Fire", diligently practicing to write the best code, and starting his first business. In the end, he joined Jinshan, founded Xiaomi, insisted on his dream, realized his dream, and continued to grow in the unity of knowledge and action...

As a result, Lei Jun's super-burning speech swiped the screen, and the record set by his own large model and the heated discussions caused by it were all overshadowed.

Not only that, a series of the latest developments in Xiaomi technology at the press conference were also overshadowed - but looking back, they are actually very important and worthy of attention, such as AI large model loaded into mobile phones, small Ai classmate's large model upgrade, Bionic quadruped robot new landing...

And Lei Jun also personally announced Xiaomi's technology strategy and AI formula**, and AI is about to launch an all-round refresh of Xiaomi.

This is why Xiaomi's new AI development is worth discussing in detail.

Let's start with the large model of Xiaomi that is four or two thousand pounds.

AI is fully empowered, and the large model is stuffed into the mobile phone

Xiaomi’s self-developed large-scale model MiLM appeared on GitHub the day before Lei Jun’s speech, and it didn’t roll.

Regardless of the size, the main versions are 6B and 1.3B (6.4 billion parameters, 1.3 billion parameters), which are proper lightweight players.

Focusing on performance, the smaller 1.3B parameter version is stuffed into the mobile phone, so that it can match the effect of large cloud models in many scenes.

The Xiaomi large-scale model team is led by Luan Jian and reports to Dr. Wang Bin**, Director of the AI Laboratory of the Xiaomi Group Technical Committee. The research and development started in April this year, and the papers have been handed in in just 4 months.

How about the results, first the evaluation results:

6B parameter version, ranked first** in the same parameter level on the C-authoritative list, and CMMLU Chinese to the large model won ranked first.

Specifically, the C-dataset is a comprehensive Chinese basic model evaluation dataset, covering 52 subjects and four levels of difficulty.

The CMMLU dataset is a comprehensive Chinese evaluation benchmark specifically designed to evaluate the knowledge and reasoning capabilities of language models in Chinese contexts. CMMLU covers 67 topics ranging from basic subjects to advanced professional level.

Since ChatGPT hit in November last year, cloud computing companies such as Microsoft, Google, and Amazon were the first to act.

However, as the technology matures further, large end-to-end models become the industry's next focus. There are already rumors about Apple, and Xiaomi is the first to enter the first echelon.

**What are the benefits of putting a large model into a mobile phone? **

First of all, local operation does not need to worry about poor signal, no network on the plane, etc., and the response speed is faster.

At the same time, user privacy is more guaranteed, and it is also convenient to provide more personalized services.

To sum it up: enjoy the advanced productivity brought by AI anytime, anywhere.

At the same time, the ** qubit has also obtained the qualification for the new version of Xiao Ai to invite students to test**.

It's still the familiar voice, but with the blessing of the large model, Xiao Ai can really be said to have evolved into a smart assistant.

It is not a problem to recommend according to the ability first, and explain the superconducting concept that has become popular recently.

Do you still remember the mentally handicapped question "Why didn't my parents invite me to get married" that stumped many domestic and foreign models?

Xiao Ai said that she should not be fooled by this at all.

If you don't know what else to do with the little love model, you can also come to the exploration zone to find inspiration.

It can also be seen from here that the Xiaoai model has also been integrated with the system to a certain extent.

For example, in Xiaomi’s gold medal customer service mode, the flagship folding mobile phone that has just been released for a few days is also well-introduced, and can jump to the mall with one click.

According to Xiaomi, Xiaomi's large-scale model is aimed at system-level capabilities. Maybe the entire mobile phone operating system will evolve into a "smart operating system" like this in the future.

However, the large model of the mobile phone is not the only direction of Xiaomi AI's efforts. The Robot Dog II Iron Egg that was unveiled together this time hides a lot of AI secrets.

Why is the second-generation iron egg smaller and more athletic than the first generation?

The CyberGear force-controlled micro-motor self-developed by Xiaomi behind it is simulated by AI algorithm, which makes the motor 20% lighter while optimizing its performance.

In order to be more bionic in terms of behavior, Xiaomi also simulated 30,000 robot dogs for parallel training on the Xiaomi Robot Reinforcement Learning Platform based on the movements of real dogs.

AI technology leaves the pure virtual environment like this, and comes to mobile phones, robot dogs, and more smart home devices in the future. It is not just a kind of "empowerment", but also discovers new directions for innovation and optimization.

We have seen in the past that Xiao Ai's voice interaction technology is used on robot dogs, which is quite different from the environment of mobile phones and speakers.

Robot dogs can move by themselves, and often appear in outdoor environments with high noise, which puts forward higher requirements for speech recognition and noise reduction algorithms, which in turn promotes the evolution of algorithms.

Large language models and multimodal large models will also encounter similar situations in the future, and eventually the features required by new application scenarios will be internalized into the model itself.

This is the embodied intelligence that the academic circle is also exploring, and it is one of the most important new directions after the big model.

We can see that in addition to the new models in the Xiaomi conference this year, the proportion of AI is significantly higher.

For the deeper thinking behind it, Lei Jun started with a formula.

**Xiaomi's technology strategy, why is there an AI power? **

The (software × hardware) ᴬᴵ formula given by Xiaomi implies a layer of meaning:

AI is the most critical variable, it can bring about exponential change.

This further reflects Xiaomi's emphasis on AI.

The reason behind it is not difficult to understand. Combined with the development of the AI trend itself and the trend of the terminal market where Xiaomi is located, the answer can be found.

First look at the field of AI.

The reason why the large-scale model trend is vigorous is that it has a profound impact on the underlying technology and industry applications.

On the one hand, the birth of ChatGPT proved that when the scale of the model breaks through a certain level, there will be a phenomenon of intelligence emerging. This feature enables the large model to have the ability to "infer other cases from one instance", and further makes it possible for AI applications to be implemented on a large scale.

At the same time, ChatGPT and Stable Diffusion also promote the development of AI in a generative direction. When AI has richer creativity, the way of human-computer interaction will also change. The most intuitive manifestation is that when the smart assistant is connected to the large model, it can promote the development of human-computer interaction to a more natural and convenient language communication method.

On the other hand, ChatGPT has sparked heated debates in the whole society around the world, making the concept of AI quickly break through the barriers and enter the cognition of thousands of industries and the general public, which accelerates the pace of AI application and commercialization.

It is generally believed in the industry that industry large models built on top of basic large models will become the main components of the future large model market and one of the main forces driving practical applications.

On the large-scale model side, manufacturers have been actively promoting cooperation with the industry in the past six months; in the vertical field, there are also players who use large-scale models as the underlying technology to show their prominence, such as InflectionAI, a chat robot company that focuses on emotional companionship. up to $4 billion.

Through these latest trend changes, it is not difficult to feel a fact——

As an underlying, basic and core technology, AI is rapidly changing the current era.

In the previous technology cycle, the representative technologies that can assume such a role are steam technology, electric drive, and computer.

In other words, today's AI, like them, can become future productivity and bring exponential growth.

Then, how to make good use of this future productivity has become the most concerned thing for various industries.

Taking the terminal market where Xiaomi is located as an example, the trend of large models is evolving here.

Now, Mobile phone access to large models has become a "big move" for many manufacturers.

Not only to combine smart assistants with large cloud models, but also to vigorously develop end-to-side models from their own scenarios.

For example, PaLM 2, which is regarded as a benchmark against GPT-4 by Google, launched the smallest "Gecko" version when it was released, which can run quickly on mobile devices, including offline, and can process 20 tokens per second.

The lower chip level is also preparing for the deployment of large models on the end side.

Qualcomm is currently able to apply Stable Diffusion at a speed of 15 steps and 20 steps to generate a 512×512-pixel image on a mobile phone equipped with the second-generation Snapdragon 8 mobile platform through quantization, compilation, and hardware acceleration optimization.

Behind this, it is reflected that the large-scale model and the mobile phone terminal are rushing in both directions.

For large models, the mobile phone scene has built the most abundant and extensive mobile applications. It is the screen closest to the general public and the key area for the accelerated landing of large models. Therefore, in order to adapt to the mobile phone scene, a model that is more lightweight and can be deployed on the device must appear.

For mobile phones, large models are accelerating the reconstruction of mobile applications and changing human-computer interaction, and the mobile phone industry will definitely be affected. Under the drastic changes, grasping the trend can grasp the new growth momentum. Therefore, for mobile phone and even terminal manufacturers, it is inevitable to embrace large models.

In summary, it is not difficult to understand why the AI content is so high in Xiaomi's latest trends.

Whether it is launching a large-scale model version of Xiao Ai, or releasing a self-developed end-to-side model, it is the result of starting from Xiaomi itself and embracing the trend.

In his personal speech this year, Lei Jun proposed the formula (software×hardware) ᴬᴵ, which further revealed Xiaomi's current understanding of the latest AI trends.

This year, Lei Jun selected the theme of "growth" as the theme of his annual speech. In the three-hour speech and annual new product release, he shared several key growths and insights he has experienced in the past 30 years.

Correspondingly, in this year that will definitely be marked in the history of AI development, Xiaomi is also standing at a new key node——

Xiaomi Group announced the upgrading of its technology strategy, announced its technology concepts, and clarified the principles of technology exploration.

In these new changes, the AI content is very high.

AI becomes the key variable of the new millet

In Xiaomi's latest upgraded technology strategy, four principles have been established:

• Deep cultivation of underlying technology
• Long-term continuous investment
• Deep integration of software and hardware
• Fully empowered by AI

A layer of information is conveyed inside and out: AI will become a key variable for the new Xiaomi.

In fact, inside and outside the trend, Xiaomi was born directly in the wind this time.

First of all, the changes triggered by the latest wave of AI are "naturally compatible" with Xiaomi's technical layout and longboard.

In 2016, when AlphaGo was born to push the concept of AI to the public, Xiaomi established its first AI team.

In the trend of AI at that time, the fields of deep learning and computer vision developed rapidly, which was not only reflected in the basic algorithm, but also in the implementation of applications. For example, the domestic AI Four Tigers all started with computer vision.

The first AI team established by Xiaomi is also anchored in the direction of computer vision.

Then in 2017, "Attention is all you need" came.

The emergence of the Transformer architecture not only directly promotes the development of the NLP field, but also affects computer vision, speech and other fields. It is also the core underlying technology of today's AIGC trend. You must know that most of the large models currently on the market are based on the Transformer architecture. .

This year, Xiaomi made bigger moves in AI. Officially established an AI laboratory, with a multi-directional layout, covering fields such as acoustics, voice, NLP, knowledge graphs, and even large models.

In the next few years, the trend of AI is in full swing. With a relatively solid underlying algorithm as the foundation, industry applications will be implemented faster, especially in the field of segmentation.

From the mobile phone scene, two typical examples are smart assistants and AI images.

They all need to be deeply integrated with the product in order to complete further exploration. In this regard, Xiaomi AI has a layout, and it can be said that it has a long board.

Xiao Ai, who was fully upgraded this time, was born 6 years ago, applied Xiaomi NLP technology, and accumulated massive scene training data for today's large-scale model access applications. In terms of AI imaging algorithms, Xiaomi has also set up a dedicated team. For example, the "Night Owl Algorithm" has been applied to mature products.

** In April this year, the Xiaomi large-scale model team was quickly established, completing the latest expansion under the trend. Up to now, the Xiaomi AI team has exceeded 3,000 people. **

Therefore, Xiaomi was born in the wave of AI large-scale models, and the accumulation and power accumulation have already been completed. All that is needed is to use good steel on the "blade" and gather resources to achieve greater breakthroughs.

Secondly, technology is powerful because of the scene. Behind the natural advantages of Xiaomi's technology, the core is the scene advantages of Xiaomi's business and products, and Xiaomi has been keenly aware of this trend.

On the one hand, only when technology goes deep into the scene can it exert greater value.

Leaving aside, let’s just say that today’s hottest concept “big model” started as early as 2017, but the AI research field has been focusing on the academic index evaluation list for a long time, and it has not caused too many waves.

It was not until ChatGPT came out with the attitude of facing ordinary consumers that the application scenarios of chatbots and intelligent assistants were opened, and the subsequent Cambrian explosion of AI species almost transformed every industry.

Xiaomi has been a practitioner of this concept for many years. This is why in the newly proposed formula, (software × hardware) is used as the base number.

Xiaomi's scene is born from the integration of hardware and software, such as the smart assistant Xiao Ai, who has 110 million monthly active users and supports more than 5,900 smart devices.

In terms of multiple terminals, "the mobile phone is the basic disk, and the IoT is the imagination space" is how many people describe Xiaomi.

On the mobile phone side of the main position, Xiaomi has established its position through the integration of software and hardware. In terms of AIoT, Lei Jun saw the trend of smart hardware and the Internet of Things in 2013, and took the lead in becoming the definer of AIoT smart life.

In 2021, Lei Jun personally took charge of Xiaomi's car building, completing the last link of Xiaomi's "people's car home", thus completing the construction of Xiaomi's full-scenario intelligent ecology.

Even in terms of bionic robots, Xiaomi has now launched two generations of CyberDog and the humanoid robot CyberOne, laying out the future home scene in advance.

On the other hand, there is room for integration and evolution only when the latest technology goes deep into the scene to develop.

For example, the well-received "Focus Tracking" function on Xiaomi mobile phones was originally derived from the "owner tracking" technology on the CyberDog robot dog; Xiaomi's autonomous driving is progressing very fast, thanks to the Xiaomi Artificial Intelligence Laboratory founded in 2017; Similarly, many technologies in the autonomous driving sector are also applied to CyberOne, which constitutes the brain of the bionic robot.

**Under the latest trend of large models, such as mobile phones, bionic robots, etc., are widely discussed landing scenarios. **

It is precisely Xiaomi's rich scenarios that make the technical application and iteration of AI more exciting. After all, AI goes back to its origin, whether it is deep learning algorithms or today's large-scale model paradigm, the core competitiveness behind it is still based on data iterations. The scenes, terminals, and data volume owned by Xiaomi are all excellent guarantees for the implementation and experience of AI technology.

And don't forget, Xiaomi also has the most reliable guarantee for AI iteration - Mi Fan Users.

After the blockbuster of ChatGPT, the reinforcement learning mechanism based on human feedback has been repeatedly mentioned, and this key mechanism and method element, no player has more methodology and voice than Xiaomi.

Finally, the most important reason why AI will refresh Xiaomi is its determination.

This kind of determination is not only the historical experience of Xiaomi's technical background and R & D investment, but also Lei Jun, who was born as a programmer, has always been firm in technology and united with knowledge and action.

For a long time, Xiaomi's product experience is too strong and the technology is forced to "eclipse", so many people ignore Xiaomi's strong technology and comprehensive layout.

Therefore, "New Xiaomi" does not mean that Xiaomi has just begun to change, but that it is a better display of the continuous efforts in technological innovation that have been made on the table.

At the scene of Lei Jun's speech, he specifically emphasized that Xiaomi's technology research and development** has deployed 12 technical fields**, including 5G mobile communication technology, big data, cloud computing and artificial intelligence. Smart electric vehicles, etc., involve 99 subdivision tracks.

Lei Jun disclosed that in terms of technical patents that most reflect strength, as of March 31, 2023, the number of Xiaomi's global authorized patents exceeds 32,000. Among them, Xiaomi's global patent family accounted for 4.1% of the 5G standard-essential patent statement effective global patent family companies released by the China Academy of Information and Communications Technology, and entered the top ten for the first time -- the youngest player.