Ming Min from the concave non-temple qubit | qbitai
Why is the trend of "large models entering mobile phones" so popular?
Vivo once again gave an answer to the outside world.
Vice President of Vivo, Vice President of OS Products, and President of Vivo AI Global Research Institute, said at the Meet Smart Future Conference
At present, the best scenario for large models to complete the closed-loop experience and business closed-loop is to land on the mobile phone and create an intelligent twin.
The implication is that not only smartphones need large models to do intelligent upgrades, but also large models need a huge mobile phone market to help large models land in applications.
It is not difficult to understand why, starting from the second half of this year, global mobile phone manufacturers have accelerated the pace of large models into mobile phones, quickly and firmly.
Taking vivo as an example, this year, vivo officially released the self-developed AI large model matrix Blue Heart large model, as well as the new mobile phone operating system OriginOS 4.
Among them, the blue heart large model contains three parameter orders of magnitude: billion, 10 billion, and 100 billion, with a total of 5 large models, of which 7 billion parameter versions are open source. At the same time, it also announced that the 13 billion parameter model will be run on the end side.
Behind these practical advances, how can manufacturers in the fierce competition understand the technology itself?What do you think about the trend development?
The latest sharing around vivo can be used as a reference in the industry.
A large model is like a dictionary that you can carry around.
Why can large models be so magical and bring great improvement to production efficiency?
It is believed that the core reason is that the large model abstracts the knowledge of human civilization for thousands of years at a high latitude, and compresses it into knowledge and information that everyone can obtain, and then applies this knowledge and information to solve problems.
This turns the large model into a "new dictionary" that everyone can use. It contains knowledge of human history, culture, civilization, etc., which can be carried with us, and it will answer professionally whatever we ask.
But it doesn't stop there.
It is believed that the big model should also have the same logical thinking, emotions and values as humans.
Take this year's hotly discussed large-scale model advanced application form agent as an example.
When it comes to the discussion of their capabilities, the "emotional intelligence" section is very high. For example, for the mobile phone intelligent body, it is believed that it should have the following capabilities:
Actively perceive the environment and behavior, actively understand user intentions, give judgments and decisions, actively call system capabilities, meet user needs, remember user Xi, provide more personalized services, have memory, bring warm companionship to users, corresponding to the actual scene, mobile phone intelligent body can fully simulate the user's control of the OS system and application under the full authorization of the user, "just like a virtual butler in the mobile phone, to help you complete various things." ”
For example, making travel strategies and booking more suitable air tickets and hotels;Two-way communication throughout the shopping process to assist users in selecting their favorite products.
On this basis, it is emphasized that mobile phone agents should have personality and memory, understand human joys, sorrows, and sorrows, and also have their own emotions.
It chats with us like a friend, giving us professional service and caring companionship.
In a word, the agent should be warm in anticipation, not cold.
And in the foreseeable future, agents will appear on various terminal carriers, with a variety of images, whether robots, pet dogs, or smart cars.
Through a series of sharing around, it can be clearly felt that mobile phone manufacturers pay more attention to the actual experience brought by large-scale model applications.
For example, comparing the large model to a "new dictionary" for each person is a concern for portabilityEmphasizing the agent's ability to understand intentions and emotionally accompaniment, it is a user-friendly focus.
In fact, this is the underlying logic followed by the development of human-computer interaction.
Throughout the history of human-computer interaction, from command line interface to touch interface, from multi-interaction to AR VR, it has always followed the direction of richer experience, more personalized, and more user-friendly.
The application of large models is the latest stage of the development of human-computer interaction.
And why can vivo and the surrounding areas grasp this underlying logic so accurately?
The answer can be found in the history of vivo in the surroundings.
Mr. Wei joined vivo in 2005 and has been engaged in smart phone software development, familiar with various Linux smart machine open source projects, and served as the general manager of vivo software research and development.
In 2018, the vivo AI Global Research Institute was officially established, and he served as the first president. At the beginning of its establishment, the institute had 12 research directions, including language recognition, NLP, machine vision, etc.
At that time, the company said that the concept of the vivo AI Global Research Institute was to build a platform based on AI technology to serve consumers.
In 2020, the vivo system OriginOS was released, and he served as the person in charge.
OriginOS has been iterated to *** in the latest version, which has integrated large model capabilities and launched a personal assistant for Blue Heart V.
While devoting himself to the exploration of AI basic technology, he is leading the research and development of mobile phone systems, and he is also a veteran of the mobile phone industry for nearly 20 years.
Therefore, it is not surprising that the surrounding area can accurately understand the trend.
Combined with his personal resume, a layer of truth can also be found that vivo's layout in AI and operating systems has been parallel for many years.
Since its establishment, the vivo AI Global Research Institute has always maintained a team of AI experts with a scale of 1,000 people. In addition, an artificial intelligence atlas research institute has been established, which has accumulated 13,000 TB of data.
Both of them start with user experience and begin to explore and implement technologies. There are already many functions in OriginOS before, and it is the spark of the collision of the two.
Now, with the outbreak of the trend of large models, these two lines have officially converged, and the results that can be seen directly are that vivo released 5 large models in one go this year, and directly brought landing applications, and ordinary consumers have been able to actually experience it.
This series of actions also provides a valuable reference for how the mobile phone industry can land large models and how large models can enter more scenarios.
An industry reference route has been given.
According to the surrounding circles, vivo's large-scale model strategy can be summarized in 5 points:
Large and complete, strong algorithm, really secure, self-evolving, wide open source.
In terms of actual actions, a month ago, vivo officially announced a series of large-scale model actions in just 15 days.
The self-developed large-scale model matrix "Blue Heart Large Model" was released, the operating system OriginOS 4 supported by the large model was released, and the R&D of the generative AI chip Dimensity 9300 was launched, and the large-scale model mobile phone Vivo X100 was released.
Behind this set of intense actions, there are two cores:
The development of large models and the implementation of large models first look at the opening of large models, which is the construction of the underlying basic capabilities.
vivo's route is self-developed + open source.
The self-developed AI large model matrix "Blue Heart Large Model", a total of five models, taking into account the device cloud:
1b device-side large model, 7b device-cloud dual-purpose model, 70b cloud main model, 130b cloud large model, and 175b cloud large model.
Among them, 7 billion versions are open source, making vivo the first mobile phone manufacturer to open source large models. At the same time, the 13 billion version is run through on the device side, which is also the first in the industry.
In terms of specific capabilities, each of the five versions of the model has its own expertise.
The smallest model with 1 billion parameters runs completely on the device side, with a memory occupation of only 1GB and a word output speed of 64 words per second.
The 7 billion parameter version is dual-purpose, and the first word response only takes 1 second, and the Chinese context capability is leading in the world.
For example, the upper-side model supports local processing on MediaTek and Qualcomm flagship platforms.
Tens of billions of large models can provide richer capabilities. The 70 billion parameter version is the main model of the device cloud, which supports role-playing, knowledge question answering, natural dialogue and other capabilities, and ranks first in multiple evaluation lists (data as of mid-November).
The 100 billion level includes 130 billion and 175 billion versions, which can perform more complex logical reasoning and task orchestration.
Let's look at the application of large models.
vivo has taken the route of integrating software and hardware. In terms of hardware, we have in-depth cooperation with chip manufacturers to accelerate the adoption of large models on mobile phonesIn terms of software, a variety of application forms have been launched and deeply integrated with the underlying system to allow consumers to get started faster.
This year, vivo and MediaTek have worked closely together on large models.
The two officially announced the cooperation on the front foot to achieve a 7 billion parameter large model and a 1 billion parameter visual large model on the end side.
Immediately afterwards, the Dimensity 9300 with an all-large core architecture was released, and this chip was jointly defined, jointly developed, and jointly tuned by vivo and MediaTek.
The innovation of the architecture makes it able to complete tasks quickly, sleep quickly, and greatly reduce power consumption, making it inherently more suitable for generative AI scenarios.
This year, the vivo X100 series won the first launch of the Dimensity 9300 and became a veritable "large model mobile phone".
In terms of software, vivo released the system-level AI assistant Blue Heart Small V and the AI application Blue Heart Qianxun.
The Blue Heart V is deeply integrated with the system and embedded in OriginOS 4 to achieve global intelligent assistance, with natural dialogue, intent understanding, intelligent search, intelligent image processing, and ** generation capabilities.
It can already be experienced on vivo X100,Blue heart small V can not only have a natural conversation with users、Figure out complex brain teasers,It can also eliminate passers-by in ** with one click,AIThe magic is stacked。
Based on the ability of the device-side large model, it can also be used without networking (** model) to summarize documents and diagrams offline.
Lanxin Qianxun is launched as a separate app, which means that even non-vivo phones can be used through the app store**.
It adds the ability to write quantization frameworks, python language interfaces, and responds in 30 seconds.
The above is a series of layouts of vivo in terms of large models.
Among them, keywords such as device-side, developer-friendly, and software-hardware combination have also been verified in recent trends.
For example, Google, which has just released a super-powerful model Gemini, has just been released.
According to the news, a smaller version of Gemini, Gemini Nano, will be available on the Pixel 8 Pro.
It will be the first to bring two localization functions, intelligent summary recording and input method intelligent reply. These all work offline, so the speed and native feel should be good.
At the same time, Google has also launched the AI Core system service, which allows developers to add Gemini driver functions to their developed applications.
It is understood that Google is also planning to incorporate Gemini Nano into the entire Android system, and Qualcomm, Samsung, and MediaTek's chips will be compatible.
The large model leverages the mobile phone industry.
In the first year of the outbreak of large models, if there is any industry that has changed because of this, mobile terminals must be one of them.
Behind this is the two-way rush of the large model industry and the mobile phone industry.
On the one hand, mobile phones are one of the first landing scenarios that many major technology companies think of.
Technology giants such as Google and Microsoft have taken the lead in laying out small models on the end of the device. And the reason why the mobile phone scene is attracting attention is also very clear:
In terms of the market, in the era of a smart phone per person, the successful use of large models in the smartphone industry means that the market of 100 billion US dollars has been leveraged.
In terms of user acceptance, the first-generation intelligent voice assistant represented by Siri has completed user education in advance, and the large model can be directly upgraded and innovated on this basis without looking for new application forms, improving the user experience and implementing it efficiently.
On the other hand, the mobile phone industry has also acted quickly in the past six months, actively embracing large models, and even driving a new trend of terminals on large models.
The most obvious representative is the domestic manufacturer.
Since the second half of the year, almost every month, domestic mobile phone manufacturers have officially announced the progress related to large models. This competitive pursuit accelerates industry innovation, and within a few months, the scale of large models that can run through the device has soared from billions to tens of billions of parameters.
Not only mobile phone manufacturers, but also Qualcomm and MediaTek have successively launched flagship chips for the generative AI era, and the performance and efficacy have been greatly improved, creating more space for the development and innovation of large-scale upper-layer applications.
This kind of upstream and downstream cooperation has also made the implementation of large models in mobile terminal applications extremely fast.
Recently, some industry views have suggested that the true value of a large model depends on the value of the industry it leverages.
Combined with the current situation, the mobile phone industry is further highlighting the value of large models and accelerating the evolution of large models.
As it was said around:
In the future, we hope to further reconstruct systems by leveraging the power of AI, and work together to move towards the era of intelligent twins through the spread of smartphones.
AI has entered thousands of households, and reshaping mobile phones is only the first step
What do you think?