How to combine large models with smart terminals? Alibaba Cloud has a pattern

In 1961, Professor John McCarthy, the "Father of Artificial Intelligence", proposed an epoch-making idea – Utility Computing (Public Computing Service).

In his view, computing services may become as popular as telephone services and become a new and important industrial foundation in society.

Later, McCarthy's prediction seemed to come true, and cloud computing began to gradually become a reality.

The overlapping identities of the father of artificial intelligence and the prophet of cloud computing seem to indicate the fate between AI and cloud computing.

Today, Ai Faner learned at the Alibaba Cloud Shenzhen AI Summit that mainstream mobile phone, PC, and automobile manufacturers have in-depth cooperation with Alibaba Cloud in the field of large models to enhance the intelligent product experience.

When AI is everywhere, cloud computing is also everywhere.

Infrastructure for the AI ​​Explosion

Han Hongyuan, vice president of Alibaba Cloud Intelligence Group and chief solution architect of public cloud, gave an opening speech on cloud computing accelerating the AI ​​explosion at the AI ​​Smart Leaders Summit held today.

Han Hongyuan said that most of today's AI-related work is actually carried on the cloud. Most organizations use these cloud capabilities and make them work more effectively by using them on the cloud.

In terms of computing, data, development, and deployment, today’s generative AI poses many new challenges to all technical capabilities. These challenges include the need to upgrade computing capabilities to the EFLOP (sound) level we see today. , including continuously running a training task, which may take weeks or months to get effective results.

As AI applications continue to deepen, cloud computing is also constantly innovating itself.

As a cloud computing company, Alibaba Cloud continues to pursue continuous improvement in the direction of ultimate performance and elasticity to effectively support the improvement needs in all aspects from storage computing network to software capabilities.

For example, the database responds to the business pressure during the annual Double Eleven and all promotional activities, including the processing of streams in big data hundreds of millions of times per second and the storage volume of hundreds of millions of terabytes per second.

The development of cloud computing is evolving in a direction that does not require users to directly manage servers.

Simply put, when users use these cloud service capabilities, they do not need to be aware of the existence of these physical servers and can use this service more effectively, thus greatly simplifying the difficulty of enterprises using IT computing power and making it easier for everyone to develop new application.

Han Hongyuan believes that since the birth of cloud computing in 2006, cloud computing has experienced an iterative process of rapid development. The cloud and the applications carried on the cloud have promoted each other, forming today's booming trend of cloud computing.

When looking at cloud computing from an application perspective, what we access are AI model services, not simple original inference services.

Behind the inference service, there are complex technical chains and processing processes hidden. Every API call needs to go through layers of stringing links to finally realize the inference and output of the model.

For users, they are more concerned about how to establish a good interface with the model and how to effectively utilize the capabilities of the model, rather than going deep into the interior of the model to explore how it is trained.

Using a base model doesn't mean you need to train it from scratch.

From IaaS and PaaS to IaaS and PaaS+MaaS, it is like the landlord not only provides a place to live and facilities, but also provides additional services such as cleaning and security, making your stay more worry-free.

In this AI era full of possibilities, there are too many ready-made services for us to choose and use, and Alibaba Cloud is one of the most cost-effective manufacturers.

How to combine large models with smart terminals

The combination of software and hardware is the development direction of large models.

Especially as the capabilities of multi-modal large models continue to increase, smart terminals such as mobile phones, personal computers, headsets, cars, and robots are expected to usher in a new explosion.

Lifestyle blogger "Brother Bao and his guide dog" who lost his sight seven years ago used a video to record the entire process of traveling by high-speed rail using "vivo see".

Through this app, he "saw" the scenery outside the high-speed train window, the water glass on the table, and "differentiated" the toiletries in the hotel.

"vivo sees" describes the rose flowers on the roadside for him, which evokes familiar childhood memories for him. Behind these "warm" scenes is the support of vivo's self-developed blue heart large model.

Since last year, vivo has stepped up the research and development of large models. The pre-training performance of the kilocalorie large model based on Alibaba Cloud PAI machine learning is close to the LLaMA level.

At present, the vivo blue heart model has three parameter levels: billion, tens of billions, and hundreds of billions, and five different sizes.

With the support of large models, "vivo Seeing" can not only automatically broadcast screen content and text information when the camera is pointed at the surrounding environment and objects, but can also switch to multiple recognition modes such as text, cards, and barcodes, and conduct multiple rounds Dialog to get more screen details and support common item searches.

Hao Xiong also said that in the future, vivo will continue to cooperate with Alibaba Cloud in terms of computing power, large models and ecological applications to further enhance the intelligent experience.

In addition to mobile phones, cars are also becoming important smart terminals.

Xpeng Motors also announced at the meeting that it has added access to Alibaba Cloud Tongyi Qianwen in the smart cockpit scenario. The upgraded car assistant "Xiao P" based on Xpeng's self-developed large models XGPT and Tongyi Qianwen can accurately understand the user's intention and adjust the temperature in the car when the user says "a little cold."

After releasing an end-to-end large-scale model for mass production, Xpeng became the first car company to put large-scale models into the cockpit and smart driving at the same time.

Apple, Google, and Tesla have all announced their official entry into the game. In China, Lenovo AI PC has been integrated with large models such as Tongyi Qianwen and was officially launched on May 10.

Since its release in April 2023, the Tongyi Qianwen series has been continuously iteratively updated. By version 2.0, its capabilities have been comparable to the industry's leading level. Version 2.5 has greatly improved its understanding, logical reasoning and other abilities.

Moreover, Alibaba Cloud's open source contributions have injected strong impetus into the development of the entire AI community.

From 7B to 14B, it has been iterated to 72B, and then to the Tongyi Qianwen model with hundreds of billions of parameters opened in April this year. In Han Hongyuan’s words, Tongyi may be the largest, most complete and most systematic parameter open on the market today. model series.

In addition, Alibaba Cloud has also built a large-scale open platform – Moda Community.

This is an AI model open source community jointly launched by Alibaba DAMO Academy and the China Computer Federation Open Source Development Committee. It aims to lower the threshold of AI application and promote the popularization and innovation of AI technology.

Developers can upload, store, and manage their machine learning models in the community, with version control and annotations to better track and manage model evolution.

As an open source platform, Moda Community supports community members to participate in the development, improvement and sharing of models, and jointly promotes the openness of the model ecology, allowing the entire AI environment to develop more vigorously.

As Han Hongyuan said at the summit:

Today, I will combine my generative AI model service capabilities on the cloud to build a new generation of applications, or transform existing applications to add more intelligent capabilities. This may be what we most want to discuss and continue to develop with you. direction, and hope that we can have more opportunities to cooperate with you on this matter in the future.

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo