Large models have been the biggest hot spot in the global technology circle for more than a year, and it is understandable that mobile phone manufacturers want to use the edge of large models to create a high-end image and achieve new growth.
However, everyone has noticed that the more "AI powerhouse", the more weight is taken to treat large models.
To put it simply, it is not to emphasize too much about how powerful the large model itself is, but to disassemble this new technology into its own existing AI system and melt it into user needs and experience changes.
For example, Apple, in AI chips, AI algorithms have a lot of achievements, but for large models, but calmly observed for a long time, Cook once said: the use of generative AI "is very important to be thoughtful and thoughtful", not long ago Apple open-sourced a ferret model, the time of the iPhone landing large model is not yet known.
Looking at domestic manufacturers, Huawei and Honor are the pioneers of mobile phone AI.
Huawei has long identified the line of defense for mobile AI, and launched the first processor equipped with NPU to support special AI tasks, even earlier than Apple. Honor's exploration of mobile AI can be traced back to 2016, when it launched its first AI mobile phone, the Magic Generation, equipped with the platform-level AI solution Magic Live intelligent system, which took four years to develop, and took the lead in introducing the concept of AI into mobile phones.
So, how do Huawei and Honor deal with the relationship between mobile phones and large models?
In mid-2023, the Huawei Mate60 series will be connected to the large model, and more emphasis will be placed on the deep integration of the large model capabilities with the Harmony OS mobile phone system, which will bring the capability upgrade to the new system.
Honor is also very cautious about large models, and it is the last of the leading domestic manufacturers to release an end-side large model. In this regard, Deng Bin, president of Honor R&D management, once mentioned in an exchange: Honor focuses on AI itself, and always takes understanding the user as the central nerve of the entire system.
It is not difficult to see that taking user experience as the primary goal and finding a people-oriented interaction solution has become the common choice of Apple, Huawei, Honor and other "device-side AI pioneers".
The big model is just a tool, and the talent is the goal. This reminds me of the martial arts**, those masters whose martial arts have reached the realm of transformation, do not stick to ** and moves, do not have to be sharp, and shout "Kanglong has regrets!" before making a move! "It's a spring breeze and rain.
On January 10, 2024, Honor released Magic OS MagicOS 80, the first to create a personalized operating system based on AI intent recognition. Melt the device-side large model in Honor's platform-level AI system, MagicOS 80 has become the "spring breeze and rain hand" of a pair of mobile phone models, turning the next page of human-computer interaction.
We're starting with MagicOS 8Speaking of 0 changes, let's talk about what the end-side AI means to China's technology, and why MagicOS, iOS, and Hongmeng may be "divided into three parts of the world".
Mobile phone model, why did it start?
In 2023, I have participated in several press conferences of large mobile phone models, and gradually discovered some "routines":
First of all, I would like to share some cutting-edge trends of large models and generative AI from a far-sighted perspective. Then it was announced that we also have a similar technology, creating xxgpt, parameter xx, which can write copywriting and draw pictures; Then posted a ranking of large model benchmarks, surpassing Microsoft, Google, Openai, and ...... on the "XX track".
This AI version of the "running point war", manufacturers are still happy to do it, and the audience is already excited.
There's a saying in tech circles that consumers don't know what technology can do, they don't know what they want, but they must know what they don't want. When mobile phone manufacturers are vigorously peddling large models, users' hearts are full of entanglements, which are manifested in a lot of "unwanted":
Don't want an extra burden. The general large model has a mobile app, which can be called and run in the cloud by direct installation, and putting the large model in the mobile phone will increase the demand for power consumption and computing power, and a more advanced processing platform is needed to support it, is it necessary?
Don't want homogeneous features. Large models and generative AI have been running wild for a year, text writing, using literate diagrams, and interacting with intelligent assistants in natural language, these use scenarios have been turned over and over by various manufacturers to educate the market, and they have long been familiar, are there any new ones?
Don't want complicated interactions. If you want to generate a large model, users have to learn to write prompts first, and the threshold for human-computer interaction has increased.
From traditional machine learning to deep neural networks to large models, mobile AI technology is always advancing, and it will always be a tool, not an end in itself.
Any technology must be adapted to the needs of users, truly innovate the experience, and let users sincerely sigh "want", so as to implement product value and market value.
Resolve all kinds of doubts in the hearts of users, and smooth out all kinds of entanglements that they don't want to use, dare not use, and can't afford to use. The large model of the mobile phone is waiting for a pair of "spring breeze and rain hands".
IUI is the shortest path for people and large models
How to establish a real intimate relationship between the model and the user? The new interaction scheme and operating system are the shortest path for people and large models.
You may have heard a saying: the simple is left to the user, and the complex is left to yourself. The value of the operating system lies in this, which plays a role in the mobile phone. Coordinate software and hardware resources downward, so that the capabilities of various components such as sensors, cameras, and processors can be used by users; Upward support for a wealth of application software, so that users can easily obtain services.
Human interaction, which is more in line with natural habits, has always been the evolutionary trend of operating systems.
From the command-line interface (CLI) represented by DOS, to the graphical interactive interface (GUI) with keyboard and mouse input represented by Windows, to the touch-centered natural user interface (NUI) represented by Android and iOS, every change in human-computer interaction, the OS operating system will shield a large number of cumbersome and complex underlying technologies, making the device more and more "fool-like", and ordinary people can use it directly.
Therefore, theoretically speaking, users are entangled with mobile phone large models and device-side AI, such as computing power, power consumption, applications, security ......It can be eliminated through the operating system, leaving only a minimalist and user-friendly interactive experience.
Through intent-based UI and personalized operating systems, the complex AI software and hardware technology on the device side and the operation threshold brought by various components are eliminated, and the complex is simplified, so that the large model can approach, understand, and serve people.
However, in practice, it is not easy to melt the large model into the device-side AI system and create an intent recognition human-computer interaction IUI.
Take Glory Magic OS as an example, the iteration time of its technical capabilities is as long as 10 years, during which the amount of funds invested by Honor reaches tens of billions. Spare no effort and work hard to achieve magicos 80's "spring breeze and rain" eliminates the barriers to the entry of large models, and truly opens the next page of human-computer interaction.
Specifically, by integrating the large model into the device-side platform-level AI and establishing the shortest path between the large model and the user, MagicOS has done three things:
The first and most important thing is to see the whole person.
The traditional OS kernel is responsible for managing and scheduling the hardware resources of the system, such as CPU, memory, hard disks, and IO devices.
In order to build a truly human-centric experience and create a human-computer interaction IUI based on intent recognition, the OS needs a "new core" composed of "human-related factors", which can be analyzed and calculated to better understand and serve people.
We know that human beings are not one-dimensional beings, but complex and multi-dimensional creatures, some characteristics can be labeled, and some hidden attributes are difficult to be perceived. To this end, Honor has accumulated factors related to people with years of practice, including:
1.Personal knowledge base. Individual and behavioral factors, such as age, gender, health status, exercise habits, lifestyle, work and leisure time, etc.
2.The system can perceive the position and state of the person (walking, running, riding, riding, etc.).
3.Human habits and portraits are learned.
Through this new core of "human factors", Magic OS can see more complete people, accurately identify user intentions, and efficiently schedule system services, making it possible to "guess what you think and understand what you need".
The second thing is to remove the end-side bottleneck of the large model.
The powerful comprehension, analysis and creation capabilities of the large model have injected technical momentum into the creation of intent recognition human-computer interaction IUI. But here comes the problem again, the mobile phone model, not just the big model. A series of factors such as the tuning of the underlying hardware platform, algorithm optimization, model compression, power consumption control, and model data security need to be considered to jointly determine the end-to-end landing effect of the large model, which directly affects the upgrade of the human-computer interaction experience.
The magic model debuted on the magic6, and Honor has overcome the balance between performance and power consumption of the device-side large model.
On the one hand, Honor's self-developed magic model is an end-side model, which is different from the cloud-side model, which uploads data to the cloud for identification and analysis, and the magic model can make personal data uninterrupted and not on the cloud. Combined with the architecture of Honor's two locks and one core, it ensures that the model runs and stores safely on the device side, and is completely isolated from the device and outside the device.
In addition, Honor also cooperated with Qualcomm to develop in-depth debugging, and the large model with 7 billion parameters developed by Honor was quantized and pruned, and then deployed on the Snapdragon 8 Gen3 to achieve smooth operation on the device side.
Reliable and fast response, eliminating the obstacles to the landing of large models on the device side, only by allowing users to truly deliver trust and have a deeper, more accurate, and more detailed understanding of user intentions.
The third thing is the precise docking of platform-level AI.
We all know that large models and generative AI can do many things, such as writing copywriting, drawing, creating, doing math problems, storytelling, ......And user needs are also diverse, life travel services, entertainment, finance, education, conferences, etc., mobile phones carry almost all the needs of everyone's daily life, some needs may not be realized by the users themselves, such as making a birthday for children, making n mobile phone wallpapers for pets.
With the rapid growth of device-side devices and application services, the "supply and demand docking" has become more and more cumbersome, and some users are even overwhelmed and begin to take the initiative to carry out "digital withdrawal" to reduce the amount of labor.
At this time, Honor's platform-level AI provides a very good foundation for IUI. The so-called platform-level AI is not a single isolated AI application, but disassembles the AI capabilities of the device into specific scenarios, bringing a closed loop of services based on intent recognition.
Users don't need to bother to passively find services, and the magic OS platform-level AI supported by large models can provide users with active services that are just right and accurately matched to the scene.
The above obstacles are removed, and the "spring breeze and rain" of human-computer interaction is identified with intent, and the magic OS becomes the shortest path between humans and large models.
Via Magic OS 80, embrace the gentleness of the AI on the end side
Along Magic OS 80 paved a large model road, users can finally dispel all kinds of concerns and entanglements, and embrace device-side AI.
The temperature of human-factor interaction can be strongly felt through many small moments, such as:
A moving moment that simplifies the complex.
If you want to make a wonderful moment of the year for your family and an anniversary for your loved one**, but you can't operate editing software, many users have to "forget it". Magic OS 80 uses human-computer interaction based on intent recognition to "simplify" the professional threshold, so that the hidden needs of these users who are suppressed because of trouble can be released and satisfied.
Through the magic end-side large model, Magic OS 80 understands the user's dialogue instructions, combined with the user's unique understanding, it can be automatically edited into a video from the ** in the personal gallery, selected wonderful moments. All it takes is a natural language instruction to make the otherwise complex professional operation extremely simple.
A magical moment like a fairytale.
Friends of all ages who have seen Doraemon have fantasized about having any door, and with just one step, they can wander in a magical and magnificent world. However, in reality, most of the time we can only wander in the boring and repetitive work day after day.
When you receive a job offer, you need to copy the event and add the schedule. Copy time, add alarms; Copy the address, navigate and search; Duplicate your flight and wait for your departure ......A seemingly simple task, but it contains a lot of labor. Just focusing on these cumbersome human-computer interactions is enough to make workers physically and mentally exhausted.
magicos 8.0 Based on the platform-level AI technology, the "king bomb" interactive function - any door, you can fold a large number of repetitive operations, automatically identify the user's intention according to the dragged content, dynamically match the smart recommendation service in the door, and reach it directly across applications in one step.
For example, paper lists, invitation letters, invitations and visits from work partners in WeChat, group message notifications from school teachers, etc., these high-frequency documents in life contain complex information such as text, images, and other complex information, which only need to be photographed by mobile phones and dragged to any door, and the end-side large model will carry out semantic analysis, cross-modal understanding, automatically extract and enter key information, and generate schedules in one step, no longer requiring people to bother to disassemble and process.
In the surprise of "one-step direct" again and again, the Arbitrary Gate and Personal Affairs Center are like magic in fairy tales, liberating people from tired life and having more leisure to feel beautiful.
A powerful moment empowered by AI.
In the past, smartphones were an extension of the human eye and limbs, allowing people to see farther and remotely control a large number of IoT devices. With AI, mobile phones have an additional "brain" with strong comprehension capabilities, which has become an extension of people's thoughts and thinking. The carrier of this "AI brain" is the intelligent assistant.
With the support of the device-side model, the capabilities of HONOR intelligent assistant YOYO have been greatly improved, which can achieve in-depth understanding of user intentions and active services for complex tasks, and become a professional and intimate personal assistant in the hands of every user.
Through the YOYO intelligent assistant, the user can have a tacit dialogue with the large model, without mechanical command instructions and repeated confirmation, and the device-side large model allows YOYO to understand the user's words (pronouns), understand the user's intention (latent intention), and then carry out automatic prompts, and parse the user's simple prompts into detailed prompts to actively provide services.
For example, if you say "yoyo, go to the old place to eat", it will be analyzed on the end side to understand that the "old place" is **, with this "plug-in brain", there is no need for users to think hard, and human-computer interaction becomes more worry-free.
It should be noted that device-side AI is not equal to mobile phone AI, and if the large model can only be applied on mobile phones, but cannot be effective on other devices, it is equivalent to "lame". End-to-end AI must bring moments of "changing the environment at will".
With the support of large models, data and services should follow people's actions and scene changes, and flow safely and freely between mobile phones, cars, smart screens, laptops, tablets and other devices to avoid fragmentation or uneven experience.
Crossing device barriers, there is a need for extremely high technological innovation (such as HarmonyOS or complete ecology (iOS) behind it, and Honor has also brought its own solutions. Magic OS 8The trust ring technology of 0 is further upgraded, and through full-stack reconstruction, platform-level AI is introduced into the interconnection technology stack, and the industry's first intelligent decision-making layer brings the experience of service intelligence to the first time, which can perceive the equipment currently used by the user and automatically complete the service flow.
*Loading. When users go out, services such as audio content, navigation, calls, and online car-hailing orders can be seamlessly transferred to the car with one click, and they can enjoy the large screen and professional audio on the car; When working efficiently, PCs and tablets can directly call the mobile phone camera to shoot** and insert documents; When you go out for a run, you don't want to carry your mobile phone and are worried about missing important messages, the mobile phone information can be automatically synchronized with the watch remotely without delaying the ......
Dive into the night with the wind, moisturize things silently, and the platform-level AI capabilities blessed by the large model melt into the IUI operating system, magic OS 8The interactive innovation of 0 has made unprecedented changes in digital life. This double-sided AI's "spring breeze and rain hand" allows users to indulge in the gentleness of technology.
The "spring breeze and rain hand" of end-side AI: holding the future of China's science and technology
Masters who have attained the realm of kung fu often have the world in mind, rather than their own selfish interests. Honor's thinking about large models is not limited to IUI, nor is it limited to its own "one acre and three points of land".
At this press conference, another major event related to the large model is the launch of the "100 Model Ecological Plan" by Honor.
If Magic OS is to dismantle and integrate Honor's own magic end-side model into the device-side, then Honor hopes to condense the AI power of the whole of China and build an unprecedented smart service ecology in the 100-model ecological plan.
This is the first large model that incorporates the cloud-side general model and various industry vertical domain models into the device-side operating system. The open architecture of the central control provides an entrance to the end side for large-scale model manufacturers and enterprises after the "100-model war".
For example, when users travel in the future, the Yoyo assistant can understand the location information in the conversation, automatically pull up the AutoNavi map, and plan and generate itinerary navigation in one sentence; If you want to fly, Magic OS 80 will desensitize the local end-side data and access it to the Qianwan model of the air travel vertical and horizontal to obtain more specialized vertical services.
From this point of view, the Honor ecosystem is also an excellent soil for large model manufacturers to build differentiated experiences and escape from homogeneous competition.
For users, Magic OS also provides the first, and currently the only, "shortest path" to access the large model ecology, without additional installation or switching back and forth between different platforms, so that rich, professional, and high-quality various large model applications can be used for their own use.
Taking Wenxin, the first echelon of China's general model, as an example, users of Magic OS will not only be able to use the Wenxin model through the YOYO assistant in the future, but also the intelligent cloud Qianfan model platform will also become a strategic partner of the Honor model ecology, and developers can enrich the MagicOS mobile application ecology through more AI native applications, provide users with an unprecedented intelligent experience, and seize the next bonus period after the mobile Internet.
With the help of the magic OS, many Chinese AI participants, such as large model manufacturers and upstream and downstream enterprises of smart terminals, have also grasped the future. China's AI forces are gathering at an accelerated pace, writing a new chapter in the "meta-innovation" of China's science and technology.
Together with HarmonyOS, MagicOS has contributed to China's OS, becoming the third pole in addition to HarmonyOS and iOS, making China's AI the glory of the world and making this world full of intelligence.
When the large model melts into the device-side AI, you and I will melt into the gentleness of technology.