Shulou(Shulou.com)11/24 Report--

The tuyere has come, "I believe that one day we can take advantage of the long wind to break thousands of miles of waves, hang up the cloud sails high and forge ahead in the sea."

The mobile phone industry in 2023 has come to a long-lost intersection.

According to The Information, OpenAI CEO Sam Altman is discussing with Jony Ive, a famous iPhone designer, the possibility of developing a new AI hardware device, further pushing the industry's imagination of the combination of AI models and mobile phones to the best part. MediaTek also announced recently that Tianji 9300 has landed the 7 billion-parameter AI great prediction model on the end side, and has successfully run the 13 billion AI big prediction model on the end side. Two new products released by Qualcomm this year, Snapdragon X Elite and Snapdragon 8 Gen 3 chips, correspond to mobile devices represented by PC and smartphones, respectively, announcing the local operation of 10 billion parameter models.

From the cloud to the terminal. The development of AI large model is surging, and the end-to-side landing large model has become inevitable. It is agreed in the industry that the big model "loaded into the mobile phone" is expected to become a new growth point for a significant upgrade of the user experience. Is it true that as long as the big model is on the mobile phone, you can catch the trend and become popular all the way?

Loading a large model into a mobile phone is only the first threshold, and the road to experience optimization on the end of the big model has just begun.

The big world, will it be the "big model" and "a thousand faces"?

On November 16, the 2023 OPPO developer Conference will be held at the Shanghai World Expo Center, when ColorOS 14 will be released. AndesGPT, which makes its debut on the same stage, is a large model for independent training of OPPO, which contains more than one billion to hundreds of billions of model specifications with different parameter sizes, which can flexibly support diversified application scenarios. As the core engine driving OPPO's AI strategy, AndesGPT is comprehensively empowering OPPO smart terminals, bringing users a new and unique intelligent experience.

The upcoming AndesGPT highlights three levels of labels: "dialogue-enhanced language model with hundreds of billions of parameters", "agents with knowledge, memory, authoring and tool capabilities" and "end-cloud collaborative full scene intelligent scheduling".

In fact, "end-to-end cloud collaboration" has become a common idea for mobile phone manufacturers to cross the "first threshold", and to some extent, it is a necessary element for the industry to promote large model experience. The cloud can solve the problems in scenarios that require big computing power and large parameter models, while the mobile phone is responsible for lightweight large model scenarios such as high data security requirements, high real-time requirements, weak network environment and so on. AndesGPT will also be designed for intelligent terminals to innovate technology and architecture, providing end-to-end cloud division, end-to-end cloud complementarity, end-to-end cloud collaboration and other work modes.

And "individuality" is put into the positioning level of AndesGPT. The reason why it emphasizes the "big model of personality", or because OPPO sees an inevitable evolution of the big model of mobile phones-to serve each individual consumer and achieve a truly "thousand people and faces" experience.

On November 7, AI's "Science and Technology Spring Festival Gala" and OpenAI developer Conference announced a series of GPT-4 Turbo upgrades. As Dan Shipper, a tech blogger attending the conference, wrote: "OpenAI makes it easy for anyone to build their own custom ChatGPT version with built-in private data-without programming. Anyone can set up a ChatGPT version with their own personality and the ability to access private knowledge. This is a major development."

What OPPO has to do, and what the industry's leading players are doing, is to be truly individualized at the level of the big model. And make targeted enhancements based on dialogue, the most commonly used form of communication.

The long way harvest comes from a direction that has been going forward for a long time.

As a generative model based on hybrid cloud architecture built by the main force of OPPO Andean intelligent cloud team, AndesGPT focuses on continuous learning, instruction fine-tuning, manual feedback reinforcement learning and knowledge enhancement in Chinese dialogue data.

Among the top ten basic competencies of SuperCLUE in September, OPPO Andean Model (AndesGPT) "knowledge and Encyclopedia" ability is second only to GPT4, ranking first in domestic big model, and "role-playing ability" in the top three.

OPPO once proposed that the big model should not be toB, but toC. Through the integration of "user profile, context awareness, personal data" and other information, to provide users with a "personalized customizable experience". Hope to make users feel that the large model is exclusive and specially tailored for him, in order to meet the needs and expectations of users to the maximum extent.

For example, the small cloth assistant on the OPPO phone is an important application of the "personality exclusive" ability of the big model. Under the blessing of AndesGPT, based on the memory characteristics precipitated in the communication with different users, Xiao Bu gives personalized recommendations and active suggestions of thousands of people, as well as content generation and creation. You don't have to describe your needs repeatedly every time, Brooks will be able to think of your every preference.

Each accumulation is also a progressive process of AI technology.

Andean smart cloud has long been an advantage for AndesGPT in many dimensions, such as data, manpower, algorithm, computing power, and so on.

Since 2020, the Andean intelligent cloud team has begun to explore and apply pre-trained language models, and the large-scale model OBERT,OBERT, which has developed 100 million, 300 million and 1 billion parameters, once ranked fifth in the Chinese language comprehension assessment benchmark CLUE1.1 and the first in the large-scale knowledge graph question and answer KgCLUE1.0 list.

In terms of data, OPPO has also completed the accumulation of large-scale and high-quality corpus data in the process of building a C-end product Xiaobu assistant with more than 100 million monthly active users; and relying on its services to 600 million users worldwide, it has completed the accumulation of relevant basic data such as end-cloud user portraits and context awareness.

Layout time, do not compete first only for reasons

In fact, OPPO has always maintained a "long-term doctrine" for the technical route of AI. Consumer electronics is an important field of carrying agents, and agents will also help consumer electronics to enter a new stage of development.

In the early stage of research and development, OPPO developers worked tirelessly and repeatedly on the bottom of the large model in order to have deeper and more detailed accumulation on the R & D path. "do not compete first" to make the technology more down-to-earth, including complex technical practices such as "pre-training, instruction fine-tuning, human feedback intensive learning, massive Chinese dialogue data, high-quality knowledge graph, high-quality QA question and answer data" and so on.

According to a technical article in July this year, the main solution of OPPO's pre-training model is knowledge enhancement. Knowledge enhancement can be understood as data enhancement. And the pre-training model lacks a priori knowledge, so integrating knowledge into the pre-training model is a very effective method. When doing model pre-training, OPPO also suggests using open domain knowledge graph based on open domain corpus to help the model improve its overall generalization ability.

In terms of data volume, OPPO Research Institute has built a high-quality knowledge graph of nearly 556 million, and from the "pre-training" and "knowledge integration" two modules to better integrate knowledge into the pre-training model.

Moreover, as a leading intelligent terminal manufacturer in the industry, relying on the experience of serving 600 million users around the world and years of related technology precipitation, OPPO has built a solid large model technology base.

Looking back at consumer electronics in the past 10 years, the expression of "wisdom" has also appeared, but the attribute of "wisdom" is more reflected in taking whatever you want and guessing what you want. On the other hand, the agent under the future big model is the independent perception, independent decision-making and active behavior under the framework of legal security. More anthropomorphic and humanized intelligence will show more diversified forms and functions to meet the needs of different fields and industries. Efficient, reliable, can complete more tasks in a shorter time, but also can better ensure the accuracy and stability of the task.

In the future, OPPO will combine large model capabilities with agents to lay out deeply into the underlying architecture of terminal products and system applications, bringing a new intelligent interaction experience. As the OPPO conversational intelligent assistant and super terminal entrance, Xiao Bu assistant is also considered to be the best experimental site and technical breakthrough for the application of Andean model.

"Wisdom" looks to the future, and the scenery should be long-term.

Looking forward to the future, the AI model still has a very broad space for development on the mobile end, and there will be more and more AI App and agent applications such as "Xin Xiaobu". Needless to say, "machine" vendors naturally have the advantage of getting through the underlying capabilities of software and hardware. OPPO, as a head brand, will also work with developers to continuously integrate large model technology with Agent into the actual experience of users, and share their own ideas on large models and technological ecology to the industry and users.

The coming ODC has aroused widespread expectation in the industry. How will OPPO create a "personalized large model service", and how will it position its own value and role in the technological ecology of the large model?

Perhaps OPPO's answers to these questions will bring new imagination and inspiration.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.