Technology cloud report original.
No one can predict the future, but we can follow common sense to capture the rhythm of technological innovation and evolution.
The hottest trend in 2023 is the large model.
At the end of 2022, ChatGPT, a chat application developed by the American start-up company OpenAI, will detonate the market, and generative AI will become a hot spot in the technology market. Behind ChatGPT is a large deep learning model, and its ability to understand and generate text exceeds that of previous AI products.
The world's major cloud computing companies have joined the competition to compete for computing power, develop and sell large-scale models, and the cloud computing market has ushered in a new round of competition, covering all levels of computing power, algorithms, and data required for AI computing.
Today, the fire of large models has been burned from natural language to thousands of industries. Compared with spending effort to make large models deal with various tricky problems more smoothly, MaaS is becoming the focus of AI large model competition.
There are many challenges in implementing large-scale enterprise models
At present, the combination of large industry models and industries is still in a relatively early stage.
On the one hand, the development of large model technology is changing with each passing day. Whether it is the level of computing power, model miniaturization, data labeling, data training or Kaiyuan model level, they are all in the process of rapid change; The understanding of the method is still in its infancy.
Generally speaking, if an enterprise wants to build a large-scale industry model, several steps are essential: select a model suitable for its own usage scenarios, use professional tools of cloud vendors to build integrated model services, sort out test cases, and establish a service quality evaluation process. Ensure the security, compliance and controllability of data and model applications.
However, if enterprises want to deploy and use large-scale model services in actual business scenarios, they face many implementation difficulties such as cost, data, and security.
First, there are fewer computing resources. Large models require relatively high computing and storage resources. Compared with general servers, the stability of GPU servers is relatively low, and the number of GPUs required for large-scale model training often reaches thousands of cards.
At the same time, in the training cluster, the network speed required to connect hundreds of GPU servers is extremely high. Once the network is congested, it will affect the training efficiency. Therefore, many algorithm teams tend to choose professional cloud service vendors for cloud service operation and maintenance support.
Second, the data quality is relatively poor.
Building a large model itself is a very costly systematic project that requires a large amount of high-quality data for training and optimization, as well as cleaning and preprocessing to eliminate noise, fill in missing values, and ensure data quality. If the quality of the imported data is low, the trained model will also have problems.
Third, the input cost is high, which is also one of the main difficulties faced by large-scale models. Large models require massive investments of data, computing resources, expertise, and time to train, debug, optimize, and deploy.
Fourth, less professional experience. The development and implementation of large models requires more technical and human resources. The deployment of large models requires consideration of computing resources, network bandwidth, security and other issues. Many companies lack relevant technologies and talents, resulting in the failure of large models to be successfully implemented.
MaaS reshape cloud computing service paradigm
MaaS, or "Model as a Service", integrates the functions of data processing and machine learning models into existing businesses through cloud services, and provides enterprises with intelligent and automated solutions. The large model of MaaS can learn from massive and multi-type scenario data, summarizing and learning general characteristics and rules under different scenarios, and becoming a model base with generalization ability.
The data set warehouse, model warehouse, and computing power platform are open to everyone by providing zero-threshold model experience, fast model use, complete link model customization, and cloud model deployment.
In terms of form, MaaS is a typical integration of cloud and intelligence. The so-called "integration of cloud and intelligence" is a concept put forward by Baidu Smart Cloud in its strategy release three years ago. Through the integration and innovation of cloud computing and artificial intelligence, computing power, frameworks, models, and scene applications are built into standardized products, thereby reducing the cost of enterprise acquisition. And the threshold of using artificial intelligence.
Baidu, Ali and even more cloud computing companies are now focusing on "integration of cloud and intelligence", which confirms that AI application capabilities are the core capabilities of cloud computing industry infrastructure after maturity. This capability lies in the level of "smartness", the general-purpose AI product capability on top of intelligent infrastructure.
From the perspective of the cloud computing service paradigm, the iterative upgrade of the large model has also reshaped the traditional cloud computing service. In the past, cloud computing focused more on computing capabilities, and the service model was concentrated on the three layers of IaaS, PaaS, and SaaS; today, cloud computing has stronger integration capabilities driven by large models, embedding large-scale computing power, algorithms, and application layers. Model, and then strengthen the "intelligence" of cloud computing, integrate applications with intelligent bases and unify external output, and realize the liberation of productivity at the scene end.
Although the current scenario of MaaS is still in the field based on natural language processing (NLP), with the iteration of technology and the extension of its core idea, MaaS in a broad sense may play an important role in addressing the above-mentioned difficulties.
In terms of data cleaning and integration, MaaS can help enterprises clean, integrate and transform data from different systems and departments to form a complete and reliable data set.
By consolidating data from different departments into one data warehouse, MaaS can ensure data integrity and consistency. In addition, MaaS can also automate the processing and analysis of massive data, and quickly discover business abnormalities and trend changes.
In terms of data analysis and mining, MaaS can analyze and mine business data, discover relationships and rules between businesses, provide enterprises with more comprehensive and accurate financial data, and support better decision-making and management.
For example, when an enterprise needs to analyze sales orders and inventory, MaaS needs to input sales order and inventory data, which can come from different departments of the enterprise, such as sales department, inventory management department, etc. After that, MaaS will use NLP technology to automatically process and analyze these data, for example, to identify and classify the status of the order, calculate the inventory turnover rate, etc.
In the process of data processing, MaaS will quickly discover business abnormalities and trend changes, such as slow sales of orders and backlog of inventory.
In terms of intelligent decision-making support, MaaS can provide enterprises with intelligent decision-making support, predict future business trends and financial conditions through machine learning models, and help enterprises make more informed decisions.
MaaS will transform the analysis results into reports and charts that are easy to understand and operate, and provide customized business solutions. Enterprises can choose different machine learning models and data processing algorithms according to their own business needs and data characteristics to achieve more intelligent and personalized business processes.
Competition in the MaaS market is fierce
In March of this year, Baidu took the lead in launching the Wenxin Yiyan model. Baidu CTO Wang Haifeng said during the Zhongguancun Forum that MaaS will become the mainstream business model of cloud computing in the future, and various applications will be developed based on large models, and each industry needs to create its own large model. The large model will be deeply integrated with the real economy, empower thousands of industries, accelerate industrial transformation and upgrading, and promote high-quality economic development.
In April, Ali also launched its own Tongyi Thousand Questions model. Zhang Yong, then chairman and CEO of Alibaba Group and CEO of Alibaba Cloud Intelligence Group, announced at that time that all Alibaba products will be connected to the Tongyi Qianwen model in the future for a comprehensive transformation.
Alibaba Cloud CTO Zhou Jingren also proposed at the Zhongguancun Conference in May this year that the concept of MaaS is being widely accepted, and models will be used as an important production element in the development of business and development systems.
Behind this, AI has always been burdened with the problem of commercialization. In recent years, the B-end market has increasingly become an incremental market for Internet giants.
In May, Tencent released its financial report for the first quarter of this year, showing that its current revenue was 150 billion yuan, an increase of 11% year-on-year; its net profit was 25.84 billion yuan, a year-on-year increase of 10%.
Net profit (Non-IFRS) was 32.538 billion yuan, a year-on-year increase of 27%. Among them, the financial technology and enterprise service sector in the B-end market saw a 14% year-on-year revenue increase to 48.7 billion yuan in the first quarter of this year. According to the reporter's understanding, this sector has accounted for more than 30% of Tencent's total revenue for eight consecutive quarters.
Previously, various cloud businesses have lowered prices one after another, which not only reflects the fierce competition in the B-side business, but also reflects the impact of the development of large models on the B-side business cost to a certain extent.
Zhang Yong once said that in the future, he hopes that the cost of training a model on Alibaba Cloud can be reduced to one tenth, or even one percent, of the current one. Even small and medium-sized enterprises can obtain the capabilities and services of AI large models through the cloud platform.
At present, the prospect of large-scale industry models depends on the technical maturity of large-scale industry models, and our competition is mainly concentrated on industry data sources. Different industries have different corpus. The dominant industry in which the manufacturer is located can form the corpus required for training AI. The more complete the corpus, the more advantageous the AI product.
At the same time, on the track of large-scale models, it has never been "the latecomer prevails". Only a richer supply can bring more customers; more customers can help improve and iterate in data feedback, thus producing a "flywheel effect".
For example, Hugging Face, a community that has accumulated in machine learning models for a long time, launched HuggingGPT based on the existing open source models in the community. It uses a large model to call multiple AI models, and quickly transforms the long-term accumulated model ecology into a larger industry. Influence.
From this stage, the role of ecology will appear. And this is why building an ecology determines the height of MaaS. But no matter which stage it is at, the core is that the large model is still a costly new thing.
No matter in the research and development, iteration or use stages, large models are a "luxury" that consumes a lot of resources and is not cheap to use.
Therefore, only by building an ecology can we truly reduce costs through economies of scale, help iterative improvements, and finally realize the real commercial sustainability of large models and MaaS, which requires more ecology. Therefore, players who really value large models and MaaS will definitely spare no effort to build an ecosystem.
In the game rules of the new MaaS paradigm, the large model determines how fast it will go at the beginning, and the ecology determines how far it will go in the end.
[About Science and Technology Cloud Report]
Focus on original enterprise-level content experts - technology cloud reports. Founded in 2015, it is the top 10 media in the cutting-edge enterprise IT field. Recognized by the Ministry of Industry and Information Technology, Trusted Cloud, one of the official media designated by the Global Cloud Computing Conference. In-depth original reports on cloud computing, big data, artificial intelligence, blockchain and other fields.