Generational changes in technology often have profound effects on the business landscape. Right now, with the deepening of investment in large models from all walks of life in the industry, an industry consensus has gradually emerged: large models are changing the cloud ecology and will bring new singularities to the cloud computing industry.
In the past period of time, in the forecast report by the end of 2022, the market research organization IDC lowered the average growth rate of China's cloud computing industry in the next five years by 10%. This adjustment is affected by the slowdown in the overall growth rate of the Internet industry, and the government-enterprise integration business is also in the process of adjustment. These factors have collectively led to Internet cloud vendors generally lowering their business expectations.
In this context, the AI model is the technological singularity, and Alibaba Cloud is the main driver of this round of transformation. On April 11, at the Alibaba Cloud Summit, Zhang Yong, chairman and CEO of Alibaba Group, made his debut as the CEO of Alibaba Cloud Smart, and its large-scale model product "Tongyi Qianwen" was launched immediately.
At the event site, Zhang Yong said that all Alibaba products will be connected to the "Tongyi Qianwen" large model in the future. DingTalk and Tmall Genie have taken the lead in accessing the Tongyi Qianwen test, and will officially release new functions after evaluation and certification ; Alibaba Cloud Intelligence Group CTO Zhou Jingren said that Tongyi Qianwen will be open to enterprises based on Alibaba Cloud infrastructure, and enterprises can train their own industry and vertical models on Alibaba Cloud.
It can be inferred from the known information that Alibaba's strategic deployment of the "Tongyi Qianwen" project does not only focus on the field of GPT-like models, but focuses on broader and far-reaching application scenarios in the future, showing An ambitious long-term plan.
Tongyi Qianwen will be "packed" into all Ali products
At this summit, Zhang Yong solemnly announced that in the future, all products under Alibaba will be integrated with the "Tongyi Qianwen" model to achieve comprehensive innovation and upgrading.
"The emergence of AI large models is an epoch-making milestone. Human beings will enter a new era of intelligence. Just like the industrial revolution, large models will be widely used in all walks of life, bringing about a huge increase in productivity and profound changes. Our way of life." Zhang Yong said.
Taking DingTalk and Tmall Genie as examples, DingTalk has added nearly 10 new AI functions after accessing “Tongyi Qianwen”. Through settings similar to Microsoft’s Copilot (co-pilot), users can wake up AI to assist daily work.
Specifically, according to the demo demo previewed by DingTalk on the day, after accessing "Tongyi Qianwen":
In the group chat scene, when you are added to a new group chat, DingTalk can intelligently generate a chat summary based on the past conversation records in the group. This feature is designed to help users quickly grasp the topic and context of the discussion, eliminating the tedious manual review process and improving communication efficiency.
In the DingTalk document scene, "Tongyi Qianwen" shows a full range of office assistance capabilities, whether it is writing poetry, novels, writing emails, or formulating marketing planning plans, etc., it can be handy. In addition, in addition to generating various copywriting content, it also has the ability to intelligently generate creative images according to instructions, bringing more convenience and inspiration to work.
In the DingTalk meeting application, "Tongyi Qianwen" has the ability to generate meeting minutes in real time, and can automatically sort out meeting minutes and to-do items to improve meeting efficiency. In addition, DingTalk video conferencing also supports real-time subtitles, so that participants can not only see the content of the speech, but also clearly understand who is speaking. At the same time, new meeting members can quickly grasp the content of previous discussions through smart summaries. After the meeting is over, DingTalk can also generate key summaries and to-do items for participants to help them arrange follow-up work efficiently.
What is eye-catching is that DingTalk has demonstrated a scene of generating applets based on taking photos. Users only need to upload a functional design sketch to realize the rapid generation of low-code business applications, and create a practical application without writing code. Such innovation has further lowered the threshold of application development and brought more convenience to users.
In addition to DingTalk, the performance of "Tongyi Qianwen" on Tmall Genie is also good.
According to the official demonstration demo of Tmall Genie, after accessing "Tongyi Qianwen", the new generation of Tmall Genie will have more flexible dialogue capabilities. It can generate corresponding content in real time according to the actual needs of users and the scene they are in, bringing users a more intelligent and personalized interactive experience.
For example, during the running process, users can ask Tmall Genie to "synthesize a playlist suitable for running based on 180 steps", and at the same time have in-depth exchanges with Tmall Genie about culture, life and other fields. Thanks to the strong technical support of "Tongyi Qianwen", the user experience of Tmall Genie will be more intelligent, convenient and diversified, and bring higher quality services to users.
Taobao Tmall's e-commerce business also has a large number of scenarios suitable for large-scale models. For example, search pictures by text, intelligent customer service, copywriting generation for selling points, etc. This will greatly change the shopping experience, and users can directly find suitable product recommendations through smarter interactions when their needs are vague.
It can be imagined that in the future, after all Alibaba products are successfully connected to the "Tongyi Qianwen" big language model and fully upgraded, various life scenarios may face a complete change and innovation in the application methods brought about by AI. This will bring unprecedented convenience and intelligent experience to daily life.
Large-scale model ecological "living water"
As mentioned above, Alibaba's "Tongyi Qianwen" is mainly combined with the group's internal products to improve product performance and user experience. However, in addition to the collaborative development of internal products, Alibaba's big language model also has an external ecological route.
At this stage, established Internet giants and artificial intelligence rookies are betting on large-scale model technology, hoping to obtain greater commercial value and technological breakthroughs. This prosperity has made large-scale model technology a new trend in the field of science and technology.
At present, large model technology has achieved impressive results, and large pre-trained models have shown amazing capabilities in natural language processing , text generation, and knowledge question answering. These large models have begun to be commercialized in many fields, such as intelligent customer service, content recommendation, intelligent writing, etc.
However, with the prevalence of large-scale model technology, some concerns have also been raised. For example, large models require a large amount of computing resources and data, which makes the research and development threshold increasingly higher, which may result in only a few technology giants being able to participate in the competition. To some extent, this has formed an important bottleneck in the development of domestic artificial intelligence.
In this regard, at the cloud summit, Zhang Yong said that facing the AI era, all products are worth redoing with large models. For this reason, Alibaba Cloud also hopes to help more companies use large models, so that every enterprise can "Tongyi Qianwen" has an exclusive large model with its own industry capabilities.
In other words, with the support of Alibaba's powerful computing infrastructure, advanced machine learning platform, and large-scale model services, companies no longer need to waste time and energy on "reinventing the wheel" or fall into the general Unnecessary competition for capabilities.
Instead, they can focus more on developing models with original and exclusive features to drive continuous innovation in the industry.
Zhou Jingren further explained that on the future Alibaba Cloud platform, enterprises can not only make full use of the comprehensive capabilities of Tongyi Qianwen, but also train and customize their own enterprise-level large models according to the characteristics of their industries and specific application scenarios. For example, each company may have its own models for intelligent customer service, intelligent shopping guide, voice assistant, copywriting assistant, AI designer, and autonomous driving.
As proof, Alibaba Cloud announced at the press conference that it will join hands with OPPO Andes Smart Cloud to build OPPO's large-scale model infrastructure, and use Tongyi Qianwen to realize continuous learning, fine-tuning, and front-end prompting projects for large models. In addition, many well-known companies such as ZTE, Geely Automobile, Zhiji Automobile, Chery New Energy, Momo Zhixing, Swire Coca-Cola, Bosideng, and Handy Technology also indicated that they will discuss technical cooperation with Alibaba Cloud in the field of large-scale models and co-innovate.
In order to promote the new development method of model-based AI development, in addition to opening large models, Alibaba Cloud is also actively building an open source AI community ecology.
In November 2022, Alibaba Cloud took the lead in proposing the innovative concept of "Model as a Service", and then launched the first domestic AI model community "Magic Build". On this platform, developers can easily download various open source AI models, and at the same time directly use Alibaba Cloud's powerful computing capabilities and one-stop AI large model training and inference services.
In just a few months, more than one million users have downloaded 16 million models of various types. At the same time, the number of models available on the platform has rapidly expanded from the initial three hundred to more than eight hundred. Among these models, there are more than 30 large models with more than one billion parameters, and 10 large models with more than ten billion parameters.
From this point of view, the chapter on the ecology of large models has just begun.
The confidence of Ali's large model: the internal strength of AI infrastructure
From a long-term perspective, to realize the wide application of "Tongyi Qianwen" in all walks of life, it is inseparable from the support of intelligent technology, as well as technical optimization in terms of computing power, algorithms and big data. These technological advantages are the result of Alibaba's unremitting efforts and technological accumulation over the years.
Take large model technology as an example. According to public information, Alibaba Cloud has been researching multimodal large models and language large models at the same time since 2019. In 2021, Alibaba DAMO Academy successively launched the language large-scale model Plug and the multi-modal large-scale model M6. Then in September 2022, the Ali large-scale model was merged into the "Ali Tongyi large-scale model series".
Not only that, but the training and operation of large models also requires a strong computing power foundation. These computing power requirements include the following aspects:
High-performance computing hardware, large models require a large number of computing resources, such as high-performance graphics processing units (GPUs), dedicated AI chips; distributed computing, due to the high number of parameters and computational complexity of large models, a single computing device is very It is difficult to meet its needs; efficient algorithms and optimization techniques. In order to reduce the training time and computing costs of large models, researchers need to develop efficient algorithms and optimization techniques.
In short, a large model requires a strong computing foundation, including high-performance computing hardware, distributed computing frameworks, efficient algorithms and optimization techniques, and a large amount of training data and storage resources. These factors together support the successful operation of the large model.
Alibaba Cloud has made sufficient preparations for this. At the summit, Zhang Yong said that over the past ten years, Alibaba Cloud has accumulated the full-stack technical strength of "AI + cloud computing" from the Feitian cloud operating system, chips to intelligent computing platforms, and has built a solid foundation in IaaS (Infrastructure as a Service) On the basis of PaaS (Platform as a Service) and PaaS (Platform as a Service), a third-tier MaaS (Model as a Service) system has been established.
Specifically, the "Feitian Intelligent Computing Infrastructure" intelligent computing IaaS service covers a series of excellent features, including a single cluster scale that can support up to 100,000 GPUs, and a powerful intelligent computing system that allows multiple trillion-parameter large models to be trained online at the same time. Computing clusters, and the self-developed RDMA network architecture that provides congestion-free communication for Wanka-scale AI clusters, etc.;
The intelligent calculation PaaS service is centered on the machine learning platform PAI, which can support Wanka's single-task distributed training scale. The efficiency of AI training can be increased by 10 times, and the efficiency of inference can be increased by 6 times. It covers the full link of AI development tools and Big data service deeply supports the research and development of generalized large models;
Smart Computing MaaS services include general-purpose large-scale model series and enterprise-customized large-scale models based on it.
The above shows that Alibaba Cloud is not only committed to building a solid technical foundation for its own large-scale model "Tongyi Qianwen", but also will share rich computing resources and large-scale model capabilities with more enterprises and all sectors of society to jointly promote AI. technology development.
It can be seen that, relying on its long-term accumulated technical strength and R&D investment, Alibaba has performed well in the competition of large language model-related technologies, and has demonstrated excellent comprehensive strength and competitive advantages in China and the world.
It is also on the basis of this technology, in addition to the big model as an appetizer, whether it is Alibaba’s comprehensive implementation of the big model to transform all its products’ intelligent process, or the large model route for the external ecology, these two sets "Mengcai" together constitute Alibaba's ingenious strategic layout.