The path towards super artificial intelligence
The Industrial Revolution integrated steam engines into production lines, amplifying human physical strength. The Information Revolution brought computers into factories and offices, enhancing human information - processing capabilities. Wu Yongming, the CEO of Alibaba Group, stated at the 2025 Yunqi Conference that AI will be the starting point of a new leap. It will amplify human intelligence and ultimately lead to Super Artificial Intelligence (ASI) capable of self - evolution.
"Artificial General Intelligence (AGI) is not the end of AI development but a brand - new starting point. AI will not stop at AGI; it will progress towards Super Artificial Intelligence (ASI), which surpasses humans and can self - iterate and evolve," said Wu Yongming in his keynote speech.
This prediction is not far off. In the daily operations of some enterprises, more and more Agents are replacing manual comparison and processing, automating repetitive tasks and making complex processes faster and more accurate. Agents are one of the ways to make large models truly applicable in industries and release their practical value.
In May this year, PwC surveyed 300 executives globally. The results showed that 79% of the surveyed companies have applied AI Agents in some of their businesses. Among them, 66% reported increased productivity, 57% saw cost reduction, 55% felt faster decision - making efficiency, and 54% improved customer experience.
Therefore, Agents have become a frequently mentioned term in almost all technology giants' press conferences this year. From OpenAI's Agent Mode that can directly execute tasks within ChatGPT, to Microsoft's use of Copilot to create the "Agentic Web", then to Google's newly released Jules Agent, as well as ByteDance's Coze and Baidu's full - platform intelligent Agents, hardly any major Internet company has been absent.
The positioning of Alibaba Cloud's Bailian Agent Development Platform is also related to Alibaba's prediction of ASI. It is not an experience - based application targeting consumers to follow the trend or hype concepts. Instead, it is a platform that enables enterprises to conveniently and effectively build enterprise AI Agents. It systematically integrates all the necessary steps for enterprises to develop intelligent agents, from model invocation, framework construction, to resource scheduling and compliance operation and maintenance.
Behind it is a complete industrial chain: from Tongyi series of basic models, to the Bailian enterprise - level development platform, and then to underlying infrastructure such as computing power, storage, and databases. It not only enables enterprises' Agents to operate effectively today but also lays the foundation for future development towards Super Artificial Intelligence.
The grand future goal and current specific applications are intertwined, which is also the consistent style of Alibaba Cloud - it can solve current pain points and bet on the next round of intelligent leap.
01
Only with a good foundation can Agents truly thrive
At the just - concluded Yunqi Conference, Wu Yongming announced that Alibaba Cloud has been officially upgraded to a "full - stack artificial intelligence service provider" and launched an all - out attack on the future computing paradigm. He proposed that large models are the new operating systems, and super AI clouds are the new computers. In the future world, a few "super AI platforms" will jointly support it.
Based on this prediction, Alibaba Cloud's strategy has remained unchanged: it has spared no effort in building infrastructure and refining base models in the past, present, and future, and will continue to increase investment in these two areas in the future. This is not only the confidence for Alibaba Cloud to enter the next stage but also the fundamental guarantee for enterprise customers to obtain long - term value.
Stable infrastructure and powerful base models can enable Agents to be truly applied in business scenarios.
Many people tend to focus on the front - end interaction experience when talking about Agents. However, enterprises' real evaluation criteria are often more practical: Is the platform stable enough? Can it support large - scale invocations? Can it be smoothly integrated with existing systems? Only when the underlying infrastructure and basic models are strong enough can an Agent platform be truly useful.
The strength of the basic model almost determines its position in the competition. Alibaba's self - developed Tongyi Qianwen series, with its capabilities in Chinese understanding, tool invocation, and multi - modal tasks, has become one of the most competitive open - source systems in China and even globally. In several domestic and international authoritative evaluations, the Qwen model has achieved top - tier global results in dimensions such as comprehensive Chinese ability, mathematical and reasoning tasks, and code generation. For example, in benchmark tests such as MMLU, CMMLU, and AGIEval, the latest version of Qwen has repeatedly outperformed international strong competitors such as GPT - 4 - turbo and Claude 3 Sonnet.
At this Yunqi Conference, Alibaba Cloud updated six models at once and launched a new brand: Qwen3 - MAX, with trillions of parameters and top - ranked international performance in coding and tool invocation capabilities, benchmarks international first - tier standards in reasoning, code, and vision. At the same time, derivative versions such as Qwen - Plus and Qwen - Flash cover the full range of scenarios from lightweight reasoning to heavy - duty generation; the new - generation native full - modal large model Qwen3 - Omni; Qwen3 - VL, which can understand and respond to the world; the image model Qwen - Image, which can "change words without distorting the face and change clothes without losing shape"; Qwen3 - Coder, with a significantly improved TerminalBench score; Wan2.5 - Preview, which can generate synchronized audio - video; and Tongyi Bailing, an enterprise - level voice base large model.
The update of models shows the "upper limit" of intelligence. However, to make these capabilities truly practical, it is inseparable from the support of computing power, storage, and scheduling systems.
Large - scale elasticity is the primary requirement for AI platforms in the industry. In real - world business, the traffic fluctuations of Agents are often unpredictable. There may be only a few hundred invocations today, but it could expand to hundreds of thousands tomorrow. Without elasticity, enterprises may either come to a halt due to insufficient computing power or be forced to hoard resources for a long time, incurring huge costs. Alibaba Cloud's solution is the serverless elastic architecture of ACS GPU. Enterprises can launch instances of models with tens of billions of parameters in seconds, and achieve cross - cluster and cross - regional scheduling in combination with the container service ACK. Coupled with inference engines such as vLLM and KServe, it significantly reduces the cold - start delay, making computing power as readily available as water and electricity.
High availability and stability are the prerequisites for enterprises to integrate Agents into core processes. A risk - control error or a system outage may cause huge losses. Relying on the Object Storage Service (OSS), Alibaba Cloud maintains low - cost and high - concurrency storage and retrieval capabilities even with a data volume in the hundreds of billions. At the computing level, it ensures the continuous and stable operation of computing tasks through automatic isolation of faulty nodes and minute - level self - healing capabilities. At the database level, products such as the cloud - native database PolarDB provide real - time and reliable support for financial - grade scenarios. These underlying guarantees enable MoE models with trillions of parameters to run stably in the Alibaba Cloud environment for a long time.
Long - and short - term memory and retrieval are the keys to whether an intelligent agent can evolve. The industry generally recognizes that if an Agent can only provide one - to - one responses without remembering past conversations and documents, it cannot form real productivity. Alibaba Cloud has incorporated Tablestore and Lindorm into its system. The former stores short - term conversations and long - term knowledge bases, while the latter supports unified management and retrieval of multi - modal data. It helps enterprises accumulate SOPs, contracts, and business knowledge, enabling Agents to draw on past experience in continuous tasks and gradually form real "organizational memory".
Inference optimization and acceleration are the core factors determining whether an Agent can be cost - effective. Enterprises generally care about two questions: How long does the inference take? Can the cost be borne? Alibaba Cloud's Artificial Intelligence Platform PAI has launched specialized engines for architectures such as MoE, DiT, and reinforcement learning. PaiMoE triples the training acceleration ratio of Qwen3; paiFuser shortens the sample processing time by nearly 30% in video generation tasks; PAI - RL achieves multi - round reinforcement learning optimization; and the inference service PAI - EAS significantly reduces the cold - start and expansion delays, with a throughput increase of more than 70%. This means that Agents can be launched faster and support large - scale user interactions at a lower cost.
Overall, the four characteristics of Alibaba Cloud's infrastructure are large - scale elasticity, high availability and stability, long - and short - term memory and retrieval, and inference optimization and acceleration. These keywords have been repeatedly mentioned in Alibaba Cloud's Apsara Release Moment column in the past year. They address the issues that enterprises care most about, such as "can it be expanded", "can it be relied on", "will it forget", and "can the cost be calculated clearly".
02
Let AI Agents grow from the enterprise business soil
At this Yunqi Conference, Wu Yongming explained the three - stage evolution route towards ASI: In the first stage, "intelligent emergence", AI acquires generalized intelligence by learning a vast amount of human knowledge; in the second stage, "autonomous action", AI masters tool - using and programming abilities to "assist humans", which is the current stage of the industry; in the third stage, "self - iteration", AI connects to the physical world and achieves self - learning, ultimately "surpassing humans".
Wu Yongming also mentioned that we are currently in the second stage. "Current Agents are still in their early stages, mainly solving standardized and short - cycle tasks... In the future, there may be more Agents and robots than the global population working with humans, having a huge impact on the real world. In this process, AI can connect most scenarios and data in the real world, creating conditions for future evolution."
From "intelligent emergence" to "autonomous action", and then to future "self - iteration", the task boundaries of large models are constantly expanding. The upgrade of Alibaba Cloud's infrastructure is to provide enterprises with a stable, controllable, and cost - effective operating environment when facing increasingly complex reasoning and planning tasks.
Many domestic manufacturers have already started their own attempts. However, compared with "attached" extensions, when Agents need to be integrated into complex enterprise processes, they often encounter problems such as high thresholds, difficult integration, and uncontrollable costs. Alibaba Cloud's Bailian has chosen another path: building a complete system from the bottom up, integrating models, development frameworks, and enterprise operation and maintenance requirements.
To solve these problems, Alibaba Cloud has proposed a "1 + 2+7" enterprise - level Agent system: one set of model services, two development modes, and seven key capabilities.
In Alibaba Cloud's Bailian "1 + 2+7" system, "2" represents two development modes. The low - code ADP allows enterprises to quickly verify prototypes like assembling Lego bricks and implement ideas in specific scenarios. The high - code ADK, on the other hand, opens up underlying interfaces, supporting in - depth customization and large - scale deployment of complex businesses.
However, such a dual - track design is rare in the industry. Many manufacturers either only emphasize "low - code", focusing on ease of use, but enterprises will soon reach the ceiling of scalability; or they only provide "high - code", with too high a threshold, requiring enterprises to invest a large amount of development resources, making the cycle and cost unacceptable.
Bailian allows enterprises to "get on board in stages". In the early stage, they can use ADP to quickly test and find effective application scenarios. Once the effectiveness is verified, they can seamlessly switch to ADK to support in - depth customization and large - scale operation.
This not only reduces the early exploration cost but also avoids repeated investment in "reconstruction" or "secondary development", ensuring business continuity. For traditional enterprises lacking a strong R & D team, this dual - track architecture is particularly friendly - they can conduct quick experiments like Internet companies and enjoy the stability and scalability of a mature platform in the large - scale stage.
To make these Agents truly "operate", Bailian further supplements seven key capabilities, including memory management, tool connection, security sandbox, log tracking and evaluation, dynamic inference, payment and transaction entry, and file and data management. These seemingly simple technical points actually address the "stumbling blocks" when enterprises implement Agents, from testing to operation and maintenance, from integration to closure. Enterprises no longer need to piece things together.
"Alibaba Cloud provides the one - stop model service platform Bailian, which supports model customization and rapid development of Agents. At the same time, it provides an operating environment for Agents like AgentBay and a series of developer kits such as Lingma/Qoder, enabling developers to conveniently use model capabilities and create and use Agents," introduced Wu Yongming at the Yunqi Conference.
The value of Alibaba Cloud's Bailian has also been demonstrated in some enterprises.
In the financial field, MYbank uses Alibaba Cloud's Bailian to take over the risk - control process. 26 types of vouchers and more than 400 types of fine - grained objects can be automatically recognized, with an accuracy rate of over 95%. The task processing time has been reduced from 3 hours to 5 minutes, and the transfer efficiency has increased by 50% - 300%. The recruitment platform Yupao.com uses more than 10 data - processing Agents to automatically clean up millions of job and resume information, increasing the job - candidate matching efficiency by 80%. The intelligent learning machine "Tinglixiong" has derived more than 50 interactive skills on Bailian, covering millions of teenage users, with the daily interaction volume ranking first among similar products.
As of now, more than 200,000 developers have built more than 800,000 Agents on Alibaba Cloud's Bailian platform, and the model invocation volume has increased by more than 15 times year - on - year. The value of Agents has been continuously verified in specific businesses.
Behind these achievements is Alibaba Cloud's choice - to build a solid foundation for enterprises to develop AI Agents, enabling different enterprises, developers, and scenarios to generate visible value.
Today's implementation results need to be understood in the context of long - term accumulation. According to a report by research institution Omdia, Alibaba Cloud has ranked first in the domestic cloud computing + AI market share for many consecutive years. In the first half of 2025, its business volume exceeded the sum of the second to fourth - ranked companies. This means that it not only has a leading edge in technology but also has established industry thresholds in terms of scale and customer trust. Hundreds of thousands of enterprise customers operate their businesses on Alibaba Cloud, making it the most proven AI + cloud platform in China.
The capabilities demonstrated by Alibaba Cloud's Bailian today are actually the result of Alibaba Cloud's long - term layout in the combination of AI and cloud computing in the past few years, which is now beginning to take shape.
03
A globally leading full - stack artificial intelligence service provider
If large models solve the problem of "where intelligence comes from", then Agents represent the next stage of AI truly entering the business world. It no longer stays at the level of generating text, pictures, and videos but has the ability to perceive, understand, execute, and provide feedback, capable of undertaking tasks and creating value in real scenarios. For enterprises, Agents have become productivity tools that can be directly integrated into processes and drive growth.
As Wu Yongming said, "Everything is just beginning. AI will reconstruct the entire infrastructure, software, and application systems, becoming the core driving force in the real world and triggering a new round of intelligent revolution."
On this path, Alibaba Cloud's layout is clear and complete: models provide intelligence, Bailian lowers the development threshold, and infrastructure ensures large - scale implementation. The Tongyi Qianwen series of models have ranked among the global first - tier in multiple authoritative evaluations, providing the underlying reasoning and decision - making capabilities for Agents. The Bailian platform breaks down the complex development process into reusable modules. Enterprises can either quickly build applications through low - code or conduct in - depth customization through high - code. The infrastructure solves the problems of large - scale operation, covering the full - chain requirements of Agents from training to deployment, including computing power, storage, and inference acceleration.
The combination of "model - platform - infrastructure" forms a complete closed