Alibaba Cloud made a big move

Author | Liu Baodan

Editor | Zhou Zhiyu

The application of AI large models is on the eve of an outbreak. Faced with the gap between basic large models and industry applications, Alibaba Cloud has decided to make a strong push.

On May 9th, Alibaba Cloud officially released Turing Q&A 2.5. According to the evaluation results of the authoritative benchmark OpenCompass, Turing Q&A 2.5 scored on par with GPT-4 Turbo, breaking a new record for domestic large models.

In addition, Alibaba Cloud also announced the open sourcing of the 110 billion parameter model Qwen1.5-110B. It is understood that this model has surpassed Meta's Llama-3-70B model in benchmark evaluations such as MMLU, TheoremQA, and GPQA. On the open source large model leaderboard Open LLM Leaderboard launched by HuggingFace, Qwen1.5-110B has surged to the top.

Alibaba Cloud, showcasing its muscle in large model capabilities, hopes to further promote the landing of large models through this.

Zhou Jingren, Chief Technology Officer of Alibaba Cloud, stated, "The potential of the model has not been fully tapped. At this point in time, we hope to open source the most advanced technology to allow everyone to explore in parallel."

Currently, Alibaba Cloud has attracted many developers to join. As of now, the Magic Platform community has 4,500 high-quality models, including various open source models. The number of developers has exceeded 5 million, making Magic Platform the largest AI model community in China.

If open source is aimed at promoting the development of the domestic AI ecosystem, avoiding redundant construction, and allowing capable enterprises to directly apply it, then the upgrade of the Hundred Refinements Platform is to better integrate large models with enterprise scenarios.

At the conference, after 8 months, the Hundred Refinements Platform ushered in a comprehensive upgrade to version 2.0, covering model development, application development, and computing power infrastructure. It not only introduces more models but also is the first to be compatible with open source frameworks such as LlamaIndex, allowing enterprises to freely replace capability components to adapt to their own systems.

Zhou Jingren introduced that there are three paradigms for enterprise application of large models at present: one is to use large models out of the box, the second is to fine-tune and continuously train large models, and the third is to develop applications based on models, with the most typical demand being RAG, enhancing knowledge of large models with enterprise data. Hundred Refinements has created a Model Center and an Application Center to meet the above needs.

This is an important shift, indicating that Alibaba Cloud's current focus has shifted from model capability research and development to the application end. Zhou Jingren stated that this year, applications will gradually explode, with more and more successful cases, and he believes there will be further breakthroughs in this area.

AI technology is driving incremental demand for cloud computing, which is a deterministic trend of the AI era and the core essence of Alibaba Cloud's firm implementation of the "AI-driven" strategy. Alibaba Cloud believes that the vast majority of AI computing will mainly run on the cloud in the future.

Therefore, over two months ago, Alibaba Cloud carried out the largest price reduction in history, involving over 100 products, with an average price reduction of over 20%. Liu Weiguang, Senior Vice President of Alibaba Cloud, believes that cloud computing is a business model with network effects and economies of scale. Alibaba Cloud provides a reusable global cloud computing network and resource pool for millions of customers. The more customers use it, the lower the cost Today, Alibaba Cloud is accelerating the application and innovation of large models in the enterprise sector, which largely implies Alibaba Cloud's consideration for revenue growth.

On February 7th, Alibaba released its third-quarter financial report, with Alibaba Cloud achieving a revenue of 28.066 billion RMB, with a growth rate of 3%. While the performance growth is weak, Alibaba Cloud is facing the high training costs of large models. Despite the firm support from Alibaba Group, Alibaba Cloud's business model ultimately needs self-validation in the AI era.

The application and innovation of large models in China are just beginning. As the only cloud computing vendor globally that develops basic models and open sources them, Alibaba Cloud is undoubtedly a company at the forefront of AI. Whether Alibaba Cloud can transform AI opportunities into performance potential, there is still a long way to go.

The following is a transcript of the communication meeting (edited):

Question: What was the process like when Alibaba Cloud decided to open source? As a cloud vendor, what is the significance of self-developing a model or the goal Alibaba Cloud wants to achieve?

Zhou Jingren: The contribution of open source to the global technology field is undeniable. Alibaba Cloud is the only cloud computing vendor globally actively working on AI models and daring to open source them comprehensively. The ultimate goal is to empower enterprises and developers with AI technology and capabilities in all aspects.

Some of the potential brought by models has not been fully explored yet. Many developers or enterprises may see a revolutionary change in their actual business needs. At this point in time, we hope to have an open mindset to open source the most advanced technology, allowing everyone to explore in parallel.

The same goes for AI products. We will showcase technological advancements in an open-source manner to everyone, not only providing strong capabilities but also offering multiple choices. We hope to explore with enterprises through an open-source system, not only enhancing model capabilities but more importantly, deeply exploring the prospects and potential of models.

Question: By 2024, many native applications developed directly based on MaaS will be booming. What is the relationship between SaaS software and these AI native applications in the future?

Zhou Jingren: Today, every industry will redefine itself and develop itself again due to AI upgrades. In this process, new applications may emerge. Technological developments will give rise to new business scenarios. Scenarios that were not fully covered or supported in the past may become possible during the technological iteration process. This is an inevitable trend in the development of human technology.

Question: In the past, SaaS grew on PaaS platforms. What will be the relationship between MaaS and PaaS in the future, and will the importance of PaaS become increasingly thin?

Zhou Jingren: No, MaaS itself is built on top of PaaS. Today, the training, development, and inference of models all rely on Alibaba Cloud's PaaS layer. We believe that in the AI era, from IaaS to PaaS to MaaS, the so-called bottom, middle, and top layers define a new architecture of cloud computing, and these three layers mutually depend on and promote each other.

Question: Since the release of Tongyi Qianwen, what is the trend of industrial landing?

Zhou Jingren: One year after the model was released, several fields are becoming clearer. Firstly, the capabilities of the model are rapidly iterating, from single-language models to multimodal ones, the overall development of the technology system is on a fast-rising channel Secondly, the development of technology needs a good ecosystem. Only with the development of the ecosystem can we receive more feedback overall today, have better business innovation, and stimulate the application of more models.

Thirdly, the trend of implementation is also very important. Behind the BaiLian platform are various models of support. We hope that with such a platform, we can truly establish a good link between the capabilities of the models and the needs of the business scenarios. Only in this way can we help large models land in various industries.

Question: What are the differentiation and characteristics of Alibaba Cloud? It is very popular now in terms of implementation. What is your focus on work regarding implementation?

Zhou Jingren: The ultimate competition for large models is about capabilities, and differentiation lies in capabilities. There are so many companies today working on large models, telling the same story. In the end, it depends on the technical capabilities and product capabilities, and the ability to implement.

The second question, I think at least in the past year, there may have been many expectations in the early stages, but the implementation process was not smooth. It involves the capabilities of the basic models themselves, as well as various standards, data, etc. This inevitably requires a platform, which is why so much has been said about BaiLian.

Question: What is the progress of using large models to redo all of Alibaba's products internally, and what are the scenarios and extent of internal use of BaiLian?

Zhou Jingren: It should be said that the Alibaba system is relatively large, and AI capabilities should be rapidly demonstrated in various scenarios, such as DingTalk, where all AI capabilities behind DingTalk are provided by TongYi.

Question: Is this a logic of transforming one product after another, or first building a basic product platform and then reflecting it at the product level?

Zhou Jingren: Similar to the products provided externally, Alibaba Cloud provides basic model capabilities. Taobao, DingTalk, Quark Search, etc., all have their own business scenarios, which will combine the scenarios with the model capabilities for business innovation and development. This not only involves AI capabilities but also various aspects of business innovation and product innovation.

Question: How does Alibaba Cloud continue to maintain its leading position in the AI era, and what is its core competitiveness?

Zhou Jingren: In the AI era, Alibaba Cloud's technical system is at the forefront. We need to combine advanced capabilities with industrial applications to truly empower everyone and apply them to specific scenarios. Only in this way, in the AI era, Alibaba Cloud will still be No.1.

Question: In January, Alibaba Cloud released a series of AI products in Singapore. Will today's product upgrades involve iterative upgrades related to overseas business? What is Alibaba Cloud's layout and planning for overseas markets?

Zhou Jingren: Related model capabilities, including BaiLian, will gradually serve overseas markets because Alibaba Cloud itself is a cloud computing vendor oriented towards the international market.

Question: Alibaba Cloud has brought positive significance to customers and developers. Has this brought some positive impact to Alibaba Cloud itself, such as an increase in demand for cloud services?

Zhou Jingren: Our overall approach is not entirely utilitarian, otherwise, your starting point is not very pure. We believe that the development of the entire ecosystem is very important today. This is why we not only talk about TongYi but also about MoDa Community, to allow more people to participate. Only by making this cake bigger can we bring about the development of the entire industry Q: How can Tongyi help small and medium-sized enterprises and growth-oriented enterprises with their business?

Zhou Jingren: BaiLian is the most important first step we have taken in this direction. It is not easy for small and medium-sized enterprises to use models in various scenarios, direct calls, fine-tuning and secondary optimization of models, deep instruction PROM optimization, stable support systems, etc. These are all capabilities that we aim to provide to help small and medium-sized enterprises