Alibaba Cloud CTO Zhou Jingren: Build the most open cloud in the AI ​​era

On October 31, at the 2023 Yunqi Conference, Alibaba Cloud CTO Zhou Jingren said that facing the intelligent era, Alibaba Cloud will upgrade its cloud computing system through full-stack technological innovation from underlying computing power to AI platform to model services. , to create the most open cloud in the AI ​​era. At the scene, Zhou Jingren announced the latest progress in the basic capabilities of cloud computing, upgraded the artificial intelligence platform, and released the large-scale model Tongyi Qianwen 2.0 with a scale of hundreds of billions of parameters, as well as the one-stop model application development platform Alibaba Cloud Bailian, Alibaba Cloud has initially built a full-stack cloud computing system in the AI ​​era.

"Currently, half of China's large model companies are running on Alibaba Cloud, and 2.8 million AI developers are active in the Alibaba Cloud Magic Community. In the future, Alibaba Cloud will join hands with thousands of industries to promote AI innovation and share technology dividends." Zhou Jingren said.

(Picture: Cloud computing technology system in the intelligent era)

Comprehensive upgrade of AI infrastructure

Large models are the core technology of this wave of AI. The quality of the basic model largely determines the prospects of AI industrialization. Training large models is a systematic project that includes complex technologies such as computing base, network, storage, big data, AI framework, and AI models. Only a powerful cloud computing system can train high-quality large models.

Zhou Jingren said that in 2009, Alibaba Cloud proposed the concept of "the data center is a computer". Today, the AI ​​era requires such a technical system even more. As a supercomputer, cloud computing can efficiently connect heterogeneous computing resources, break through the bottleneck of a single performance chip, and collaboratively complete large-scale intelligent computing tasks.

In order to ensure stable interconnection and efficient parallel computing for large model training, Alibaba Cloud has newly upgraded the artificial intelligence platform PAI. The bottom layer of PAI adopts HPN 7.0 new generation AI cluster network architecture, which supports cluster scalability up to 100,000 cards. The acceleration ratio of ultra-large-scale distributed training is as high as 96%, far exceeding the industry level; in large model training tasks, it can save More than 50% of computing power resources, leading the world in performance.

Alibaba Cloud Tongyi large model series is trained based on the artificial intelligence platform PAI. In addition to Tongyi Large Model, half of China's major model companies run on Alibaba Cloud. A large number of leading companies and institutions such as Baichuan Intelligence, Zhipu AI, 010000, Kunlun Wanwei, vivo, and Fudan University are all trained on Alibaba Cloud. Large model.

Wang Xiaochuan, founder and CEO of Baichuan Intelligence, said, "Baichuan has released 7 large models in just half a year since its establishment. The rapid iteration is inseparable from the support of cloud computing." Baichuan Intelligence and Alibaba Cloud have conducted in-depth cooperation. With the joint efforts of the team, Baichuan successfully completed the training task of the large kilocalorie model, effectively reduced the cost of model inference and improved the efficiency of model deployment.

Alibaba Cloud has become the public AI computing power base for China's large models. Up to now, many mainstream large head models in China have provided API services to the outside world through Alibaba Cloud, including Tongyi series, Baichuan series, Zhipu AI ChatGLM series, Jiang Ziya general large models, etc.

As the industrialization of AI gradually deepens, there will inevitably be an explosion in demand for large-scale intelligent computing power. Alibaba Cloud has built 89 cloud computing data centers in 30 regions around the world, providing more than 3,000 edge computing nodes. The advantages of low latency and high elasticity of cloud computing will be fully utilized. This year, Alibaba Cloud successfully supported the short-term and high-intensity traffic burst of Miaoya Camera, which is popular across the entire network.

Towards the cloud of autonomous driving

Zhou Jingren said: "With the integration of large model technology and cloud computing itself, we hope that the future cloud can drive autonomously like a car, greatly improving the developer's experience of using the cloud."

It is reported that more than 30 cloud products on Alibaba Cloud have access to large model capabilities. For example, Alibaba Cloud's big data management platform DataWorks has added a new interactive form - Copilot, which allows users to generate generated data simply by inputting in natural language. SQL, and automatically execute corresponding data ETL operations, the overall development and analysis efficiency can be improved by more than 30%, comparable to "autonomous driving".

Alibaba Cloud containers, databases and other products also provide a similar development experience, which can realize functions such as NL2SQL and SQL annotation generation/error correction/optimization. In the future, these capabilities will be integrated into other Alibaba Cloud products.

Better large models, accelerating application innovation

At the 2023 Yunqi Conference, Zhou Jingren announced the latest progress of Alibaba’s self-developed large model and released Tongyi Qianwen 2.0 with a scale of hundreds of billions of parameters. Tongyi Qianwen 2.0 comprehensively surpassed GPT-3.5 and Llama2 in 10 authoritative evaluations, accelerating to catch up with GPT-4.

Thousands of industries want to use large models to achieve changes in production and service methods, but the high threshold for using large models has blocked most people from the technological wave. Whether it is customizing an exclusive large model or building innovative applications based on large models, there are high requirements for talent, technology, and funding.

At the Yunqi Conference, Zhou Jingren released a one-stop large model application development platform - Alibaba Cloud Bailian. The platform integrates domestic and foreign mainstream high-quality large models and provides model selection, fine-tuning training, security suites, and model deployment. Services and full-link application development tools simplify complex tasks such as underlying computing power deployment, model pre-training, and tool development for users. Developers can develop a large model application in 5 minutes, and "refine" an enterprise-specific model in a few hours. Developers can focus more on application innovation.

In order to make it easier for large models to be integrated and implemented in thousands of industries, Alibaba Cloud "prototyped" 8 industry large models based on Tongyi, and announced the progress on site: personalized character creation platform Tongyi Stardust, intelligent investment research assistant Tongyi Dianjin, AI reading assistant Tongyi Zhiwen, etc. made their debut; intelligent coding assistant Tongyi Lingma has been widely adopted within Alibaba Cloud and has been widely praised; work and study AI assistant Tongyi Tingwu processes more than 50,000 sounds every day Video, with more than 1 million cumulative users.

Large models are triggering a new round of innovation in thousands of industries. Currently, companies such as CCTV.com, Langxin Technology, and AsiaInfo Technology have taken the lead in developing exclusive models and applications on Alibaba Cloud Bailian. Langxin Technology trains on the cloud We have developed a large model dedicated to electricity and developed an "intelligent assistant for power bill interpretation" and "power industry policy analysis/data analysis assistant" to improve customer reception efficiency by 50% and reduce complaints by 70%.

Use firm openness to jointly promote ecological prosperity

"Promoting the prosperity of China's AI ecosystem is Alibaba Cloud's primary goal. Alibaba Cloud will firmly build the most open large model platform in the AI ​​era. We welcome all large models to connect to Alibaba Cloud Bailian and jointly provide AI services to developers ." Zhou Jingren said.

Alibaba Cloud is the first technology company in China to open source self-developed large models, setting off a wave of open source large models in China. At present, Alibaba Cloud has open sourced the 7B and 14B versions of Tongyi Qianwen, with downloads exceeding one million. At the scene, Zhou Jingren announced that the Tongyi Qianwen 72B model will be open source and will become the open source model with the largest parameters in China.

In addition to sharing self-developed new technologies with developers, Alibaba Cloud also strongly supports the development of large-scale models from three parties. In the Alibaba Cloud Magic Community, top players in the industry such as Baichuan Intelligence, Zhipu AI, Shanghai Artificial Intelligence Laboratory, and IDEA Research Institute have all open sourced their core large models, and Alibaba Cloud provides developers with "early adopters" of large models. Provides free GPU computing power, which has exceeded 30 million hours so far.

Zhou Jingren revealed that the Moda community has now gathered more than 2,300 AI models, attracting 2.8 million AI developers, and the number of AI model downloads has exceeded 100 million, making it the largest AI community in China with the most active developers.

At the 2023 Yunqi Conference, Alibaba Cloud announced a blockbuster plan: the "Yun Gong Kaiwu Plan", which will provide a cloud server to all college students in China. In addition, Alibaba Cloud will also provide larger-scale computing resource support to contracted universities to help China's young scholars and students reach the peak of scientific research. At present, Tsinghua University, Peking University, Zhejiang University, Shanghai Jiao Tong University, University of Science and Technology of China, South China University of Technology and other universities have reached the first batch of cooperation.

The above is the detailed content of Alibaba Cloud CTO Zhou Jingren: Build the most open cloud in the AI ​​era. For more information, please follow other related articles on the PHP Chinese website!