WAIC2026基座大模型架构创新与生态合作论坛

文章语言：

简

繁

Minutes

原文

会议摘要

SenseTime emphasizes the reshaping of a domestically sourced supply chain and AI innovation, showcasing achievements such as optimized operator performance and enhanced data center energy efficiency, and has launched the “Galaxy Plan” to accelerate the commercialization of the domestic ecosystem. In terms of global collaboration, the discussion explored the potential for cooperation between China and countries such as Japan, South Korea, Singapore, and Malaysia, emphasizing complementary strengths and cross-cultural understanding. The urgency of AI governance and ethics has been highlighted, with calls for an open and inclusive global AI ecosystem. The presentation showcased how AI empowers individuals and enterprises, driving industry transformation—for example, the Cycle agent, a multimodal model, lowers the barrier to content creation. SenseTime continues to innovate its infrastructure, building a domestically produced inference cluster to accelerate the global expansion of AI. The forum focuses on the iterative advancements of large‑model technologies and ecosystem collaboration, showcasing the latest AI developments and future prospects, such as the U1 Pro model’s breakthrough in unified understanding and generation.

会议速览

2026 World Artificial Intelligence Conference: Jointly Exploring Large Model Innovation and Ecosystem Collaboration

At the 2026 World Artificial Intelligence Conference Forum on Foundation Model Architecture Innovation and Ecosystem Collaboration, participants engaged in discussions on the iterative advancement of large‑model technologies, their industrial deployment, and ecosystem synergy. They emphasized aligning with national strategic priorities, deepening independent innovation, coordinating internal and external collaboration, and jointly building a secure and thriving AI industry ecosystem to promote the widespread adoption and high‑quality development of artificial intelligence.

Shanghai is driving innovation and ecosystem development in foundational large models, jointly charting the blueprint for the future of AI.

At the Forum on Architectural Innovation and Ecosystem Development for Foundation Models, participants jointly discussed Shanghai’s latest advances in the field of artificial intelligence, including enhanced model‑supply capabilities, the accelerated strengthening of its intelligent infrastructure, and the rapid emergence of a thriving industrial ecosystem. Looking ahead, Shanghai will strengthen efforts to tackle key infrastructure challenges, accelerate the development of a robust intelligent foundation, and deepen its open ecosystem, with the aim of advancing the development of large‑scale foundation models in artificial intelligence and jointly forging a path of high‑quality growth.

Multimodality and Language Models: A Discussion on the Core Path of Future AI Development

The roundtable forum delved into the central role of multimodal processing in today’s state-of-the-art large models, as well as its and language models’ respective roles in the future development of AI. The experts in attendance believe that multimodality not only enhances models’ ability to understand complex scenarios but also serves as a crucial bridge between the digital world and physical space. Although language models play an important role in cognitive understanding, the integration of multiple modalities is regarded as a key direction for advancing AI technology, underscoring the need to align multimodal information with language and to build the infrastructure for developing sophisticated contextual understanding.

Exploring the Next Generation of Intelligence: The Convergence of Multimodal Models and Autonomous Learning

Four pathways for the development of intelligence were discussed, with emphasis placed on the importance of multimodal models and the pivotal role of autonomous learning in realizing the next generation of AI. It was pointed out that intelligent learning should not be confined to language or vision alone, but rather focus on comprehensive cross-modal learning capabilities.

Language and Vision Agents: From Static Learning to Dynamic Self-Evolution

The discussion focuses on the success of language models during the static learning phase, as well as the analogous challenges they face in dynamic scenarios alongside visual models. It emphasizes the importance of agent self-evolution in interaction, highlights the limitations of language and vision in dynamic learning, and looks ahead to the pivotal role of multimodal data in the future development of intelligent agents.

Exploring the Bottlenecks and Future Paths of AI: Innovation in Data, Architectures, and Learning Paradigms

The discussion identified three major bottlenecks in AI development: at the data level, more effective recording and annotation are needed; architecturally, native support for multimodality is required; and in terms of learning paradigms, the challenge of continuous evolution must be overcome. It identifies the challenges that AI must overcome in the future, including the social organizational issues of data collection, the exploration of multimodal architectures, and the development of methods for achieving true continuous evolution.

In-depth Exploration of Multimodal Agent Memory Mechanisms and Online Learning Challenges

The discussion focuses on the challenges faced by multimodal agents in long-context memory management and online learning, highlighting the current models’ shortcomings in handling dynamic information, compressing multimodal signals, and constructing robust, long-term memory mechanisms. It emphasizes the importance of online learning, exploratory learning, and efficient learning, particularly the potential of world models to address the challenges of interaction efficiency and sparse knowledge density. Meanwhile, we propose a potential solution based on on-policy learning and world-model construction, as well as the need to innovate existing memory mechanisms.

AI Architecture and Training Paradigm Innovation: From Data-Driven to Algorithm Optimization

The importance of data collection, training paradigms, and algorithmic architectures in the AI field was discussed, with emphasis on the scarcity of data in the physical world and the challenges of efficient data acquisition. At the same time, it highlights the driving role of continuous innovation in AI technologies, including advancements in training methods and algorithmic architectures, as well as the future trend of deep integration between software and hardware.

Five-Year Outlook on AI Development: Overestimations and Underestimations

The discussion focused on aspects of the AI field that may be overestimated and underestimated over the next five years, highlighting the potential overestimation of AI’s autonomous capabilities in enterprise‑level applications, as well as the underestimation of AI’s impact on transforming how society operates. It emphasizes future paradigms of AI design, such as next-generation AI designed by AI and the development of multimodal models, arguing that technological progress is achieved through gradual accumulation rather than overnight.

AI’s New Generation: Original Rap Showcases the Journey of Scientific Research and Witnesses the Times

Through an original rap, the piece showcases the tenacious spirit of AI researchers in the face of challenges and setbacks, as well as their relentless pursuit of scientific passion. Meanwhile, it reviewed how the SenseTime Scholarship has nurtured and supported AI talent over the past decade, highlighting the strength of the new generation of AI professionals and the progress of the times.

SenseTime Scholarship’s Decade-Long Impact: AI Rising Stars Together Shape the Blueprint for the Future

Since the establishment of the SenseTime Scholarship in 2017, over the past decade it has supported 280 outstanding undergraduate students, witnessing their diverse development from academic research to industrial applications; today, these awardees have become the backbone of China’s AI sector. The past decade has been not only a time for reflection but also a period of forward-looking vision. SenseTime continues to push the boundaries of AI with original technologies, inviting the next generation of young people to take up the mantle and jointly explore the future of AI integration across all industries.

The Evolution of Intelligent Technologies and the Exploration of Multimodal Applications: Bridging the Gap from Professional Use to Mass Adoption

The dialogue delves into the evolution of intelligent technologies from specialized domains to widespread public use, analyzing the shift from model‑as‑a‑service to ubiquitous applications, as well as the breakthroughs of multimodal technologies in content generation and system‑level delivery. Through case studies, the text demonstrates how technology progressively resolves the tension between specialization and universality, thereby realizing broader social value.

AI Empowers Super Individuals: Unlocking New Industry Applications and Personal Creativity

Through the deep application of AI technology, it is not only possible to scale professional expertise—such as codifying the experience of landscape review engineers—but also to spark interest-driven innovation, like story maps created by history enthusiasts. These cases demonstrate how AI expands the boundaries of capability, fosters the emergence of industry applications, and at the same time provides ordinary people with platforms for creation and expression, thereby enhancing individual value and ensuring the efficient utilization of social resources.

AI Boundary Expansion and Empowerment of Super Individuals: Innovative Practices from Visual Impairment Education to Content Creation

The conversation revolves around the expanding frontiers of AI technology and its profound impact across various sectors of society, with particular emphasis on AI’s applications in education for the visually impaired and on how intelligent agents can empower individuals and enterprises to democratize content creation. It emphasized the shift of AI from a tool to a capability, as well as the importance of collaborating with partners to unlock the potential of super individuals, showcasing the wide-ranging applications and future prospects of AI technology in fields such as education, enterprise services, and content creation.

AI empowers personal growth and life optimization: from health assistant to financial planning

It discussed how AI, through applications such as health, photography, and financial management assistants, helps users achieve their life goals and enhance their quality of life, emphasizing that AI not only replaces human labor but also expands the boundaries of capability and fosters personal growth.

Large-Scale Facility: Exploration of Indigenous AI Infrastructure Innovation and Scalable Applications

The speaker outlined the large-scale facility’s explorations and achievements in the field of AI infrastructure over the past year and more, including the launch of the first commercial PD‑separated inference service, heterogeneous decoupling optimization, and performance enhancements for domestically produced chips. It underscored the large-scale deployment of a domestically developed inference cluster, with an average daily invocation volume of 2.4 trillion and projected to reach 10 trillion by year-end. At the same time, it shows the localization support and application expansion under the scenes of AIGC and scientific discovery, and accelerates the maturity of the independent ecosystem.

AI infrastructure innovation and space computing power strategic cooperation

Strategies for enhancing the environmental sustainability and efficiency of AI infrastructure were discussed, including the application of the TPW metric and optimization of AI scheduling systems, as well as future technology trends such as photonic computing, quantum computing, and space-based computing power. Meanwhile, SenseTime and Guoxing Aerospace announced a strategic partnership in the field of space-based computing power, aiming to establish a constellation of one thousand satellites within five years, deliver ten thousand units of space-based computing power, and support the global expansion of Chinese AI applications.

Co-creating a domestic AI ecosystem: SenseTime’s Galaxy Program is launched to drive the comprehensive development of intelligent infrastructure.

Against the backdrop of advancing the development of a domestic AI infrastructure ecosystem, the Galaxy Initiative has been launched with the goal of building multiple domestically developed clusters within two years. By collaborating with partners across ten key technological domains, it aims to serve over 200 AI startups and accelerate the commercial deployment of the domestic ecosystem. At the same time, it emphasizes the application of AI infrastructure in the fields of physical intelligence and scientific discovery, and cooperates with scientific research institutions to promote the construction of AI-driven scientific discovery platform, and jointly build a sustainable domestic AI ecosystem.

New Advantages and Role Positioning of Chinese Innovation in the Context of Global AI Ecosystem Reconfiguration

Following the event’s launch ceremony, focusing on the advantages and role of Chinese-style innovation amid the global AI industry’s restructuring, a roundtable discussion explored how China can assume a new role in the global AI ecosystem. Several Asia-Pacific tech leaders participated in the dialogue, jointly envisioning China’s innovative positioning and contributions on the world stage.

How Chinese Innovation Is Reshaping Regional Cooperation: Opportunities and Challenges from an Expert Perspective

This discussion focused on the impact of China’s innovation on regional cooperation. Drawing on their respective fields, the participants highlighted how China’s rapid technological advancements, particularly in the field of artificial intelligence, are opening up new opportunities for global collaboration. The Malaysian representative emphasized that China’s open AI models have made regional collaboration possible, while the Korean representative pointed out that China’s strengths in implementation efficiency and market data are encouraging Korean companies to shift toward innovation. Singapore’s representative believes that China’s technology exports have become a global focal point, and that Chinese innovation is increasingly integrated into the international value chain, thereby advancing regional economic integration. The Japanese delegation shared case studies of China–Japan innovation cooperation, demonstrating the potential for two-way technology exchange. Overall, China’s innovation is gradually reshaping regional cooperation models and fostering closer economic and technological ties.

New Opportunities for China–South Korea Scientific and Technological Cooperation and Prospects for Global Innovation Trends

The conversation centered on the complementary strengths of China and South Korea in fields such as semiconductors, AI, and data centers, emphasizing the importance of deepening cooperation for global standard-setting and supply chain optimization. It proposes a “China Technology Procurement” strategy to promote the internationalization of Chinese technology companies, while helping them adapt to overseas markets and achieve a win-win outcome for both technology and market development. In the future, China’s role in global technological innovation will become even more prominent, and it is essential to strengthen its engagement with the international community to jointly advance scientific and technological development.

Chinese Innovative Enterprises Enter the Japanese Market: Opportunities to Address Aging Populations and Labor Shortages

The discussion explored the potential of Chinese innovative enterprises in the Japanese market, particularly in addressing Japan’s aging population and labor shortages. It examined how companies in sectors such as food, robotics, and unmanned retail can leverage technological innovation and collaboration to deliver solutions tailored to the Japanese market, including automation upgrades across manufacturing, warehousing, and the retail industry.

The Role and Challenges of East Asia in Global Technological Innovation: A Perspective on the Globalization of Chinese Technology Enterprises

The discussion focused on East Asia, especially the role and challenges of Chinese technology companies in promoting global technological innovation, especially in the field of artificial intelligence. It highlights the challenges posed by market heterogeneity that Chinese technology firms encounter in their global expansion, as well as East Asia’s advantages in technological innovation, such as its wealth of practical experience and meticulous attention to detail. At the same time, the importance of maintaining neutrality in the global technology competition and the need to build a more secure technology architecture to support long-term development are pointed out.

East Asian Science and Technology Cooperation and Innovation: Building Global Influence

The dialogue explored how East Asian countries can complement each other’s strengths in areas such as semiconductors, AI technologies, and supply chains, emphasizing that through cross-regional collaborative innovation, East Asia can exert a significant influence on global technological development. Referring to the phenomenon that many AI scientists are of East Asian origin, the discussion examines the impact of cultural factors on scientific and technological contributions, while also highlighting the importance of communication and open exchange, as well as how China’s innovation is shaping its neighbors and even the entire world.

AI Development and Governance: A Global Dialogue on Balancing Innovation and Risk

The positive role of AI technology in enhancing quality of life was discussed, while the importance of AI governance and risk management was also explored in depth. Participants emphasized the need to establish multilateral cooperation mechanisms and formulate global AI guidelines, as well as the critical role of preventive measures in mitigating potential risks. The dialogue points out that, in the face of AI’s rapid advancement, it is necessary to establish an open and collaborative governance framework to ensure that technology develops for the common good, while also remaining vigilant about the challenges posed by potential technological失控.

Chinese-style innovation and AI development: the 'bittersweet lessons' in technological evolution

The discussion focused on China’s innovation in the field of artificial intelligence and global cooperation, emphasizing the importance of openness and collaboration. By reviewing the history of technological development, particularly the “bitter lessons”—namely, the contrast between knowledge construction and large-scale computing—we reveal the intrinsic principles governing progress in the AI field. The speaker argued that although history has repeatedly demonstrated the importance of large-scale computation and learning, whether this principle will continue to hold in the future—particularly with respect to foundation models—remains an open question that calls for ongoing observation and exploration.

Large Models and Foundation Models: The Limitations of Knowledge Repositories and Their Future Vision

This paper examines the limitations of large models and foundation models in knowledge acquisition and creation, arguing that foundation models should serve as repositories of human‑shared knowledge rather than striving for comprehensive knowledge mastery. It emphasized that the complexity of the world exceeds the capabilities of any single model, advocated for building models that accurately reflect consensus knowledge to deliver services in a customized manner, and looked ahead to future directions in AI development, including the coexistence of human-like intelligence and specialized-function AI.

要点回答

Q：What is the theme of this forum?

A：The theme of this forum is architectural innovation and ecosystem collaboration for foundation large models, with a focus on the iterative advancement of large‑model technologies, their industrial deployment, and ecosystem synergy.

Q：What key points did Li Bin, member of the Party Leadership Group and Vice Chairman of the All-China Federation of Industry and Commerce, emphasize in his address?

A：In his address, Vice Chairman Li Bin emphasized three key points: first, we must grasp the direction of the times, keep the nation’s strategic priorities at heart, and seize development opportunities by aligning with national strategies; second, we should deepen independent innovation, strengthen our core competencies, and consolidate the foundations for growth through tackling tough challenges; third, we must coordinate internal and external collaboration, improve our governance system, and jointly build an open, inclusive, secure, and prosperous artificial intelligence industry ecosystem.

Q：What achievements has Shanghai made in the field of artificial intelligence, as mentioned by Pan Yan, Deputy Director of the Shanghai Municipal Commission of Economy and Informatization, in her address?

A：Deputy Director Pan Yan noted that Shanghai has achieved significant results in the field of artificial intelligence, including breakthroughs in foundational large‑model technologies, the development of a robust computing‑power infrastructure, and the building of a thriving ecosystem. At present, Shanghai has seen multiple large-scale models successfully registered and made open source; some of these models boast world‑leading comprehensive capabilities, while a multi‑tiered computing power supply framework and a robust industrial ecosystem have taken shape.

Q：In the development of large models, how will multimodality influence and drive the future of AI?

A：Multimodality plays a central role in the development of large models; it not only enhances their overall capabilities but also serves as a key enabler for building intelligent agents and advancing artificial intelligence technology. In the future, multimodality will be more deeply integrated into model design and applications, helping AI better understand and perceive the world.

Q：What are the differing views of the experts in attendance regarding the role of multimodal AI in large models?

A：At the conference, experts held two views on the role of multimodality in large models: one argues that language models are the core of future AI, with other modalities such as images and videos being processed through language models; the other maintains that multimodality plays a crucial endogenous role in the interaction between the digital world and physical space, serving not only as an extension of language models but also as an indispensable component for agents to understand the world.

Q：In human–machine interaction scenarios, why do multimodal models struggle to comprehend complex environments and contexts?

A：This is because current multimodal models lack the ability to understand complex real-world scenarios. Beyond textual context, they must also process multimodal information such as objects, behaviors, and actions in the real world, and establish the causal relationships and interaction logic among them—tasks that a single-language model is far from capable of accomplishing.

Q：Is the development of next‑generation intelligence related to the source signals that give rise to intelligence?

A：Indeed, the development of next‑generation intelligence is linked to multiple approaches, including large language models, simulations of open‑space robots, digital life, and brain–computer interfaces—each of which is tied to the fundamental signals that give rise to human intelligence.

Q：When considering the next generation of intelligence, should the focus be on language models or multimodal models? Why have large language models achieved success at this stage, while vision models have performed relatively more weakly?

A：The core challenge of next‑generation intelligence lies not in choosing between a language‑centric or a vision‑centric approach, but in the learning paradigm—namely, how to enable intelligent agents to achieve autonomous or acquired learning. Both language models and vision models have their limitations, and a multimodal approach is needed to drive future advancements. The primary reason is that language model training datasets are vast and highly dense in information, making them well-suited to this learning paradigm, which relies on imitation learning and rule-based reinforcement learning. However, visual models, due to signal sparsity and high dimensionality, perform poorly under the current learning paradigm.

Q：What challenges arise in the data layer, architectural design, and learning paradigms when building the next generation of intelligent systems?

A：At the data level, in addition to accumulating extensive linguistic resources, we must also address the challenges of organizing, annotating, and effectively leveraging multimodal data such as videos and images. Architecturally, we need to explore designs that natively support multimodal input and output, tackling difficulties in the encoding and representation of long texts and sparse information. In terms of learning paradigms, a more elegant solution is required to overcome the limitations of current language‑model‑centric architectures, enabling online and autonomous learning in dynamic, multimodal scenarios.

Q：How can we address the challenges of information acquisition and self-evolution faced by intelligent agents when interacting in real-world environments?

A：Intelligent agents need to learn efficiently from the rich feedback they receive from their environment and achieve self-evolution, but current models lack this capability. The key to solving this problem lies in improving the learning paradigm, enabling the model to autonomously learn and improve rapidly from sparse real-world environments.

Q：Regarding the continuous evolution of intelligent agents, what challenges do you believe need to be overcome? Can the context‑management mechanisms relied upon by current mainstream models support a long‑term, unbounded evolutionary process?

A：I believe that the path of self‑evolution through the continuous accumulation of skills, which is currently widely accepted in the engineering community, has its limitations. True, sustained evolution—such as the decades-long process of learning and growth seen in humans—is still in its infancy for machines. Today, machines can handle tasks that once took hours, but they still fall far short of human capabilities in terms of long-term growth and development. To achieve true, sustained evolution, we must address the challenges of autonomous and efficient learning, with particular emphasis on three key issues: online learning, curiosity-driven exploration, and modeling. Although existing models such as the transformer excel in short-term memory, they face challenges in modeling long-term memory, particularly in effectively capturing both transient and enduring memory representations. Although the parameter count of existing models is scalable, the current memory mechanisms may need to be reimagined to facilitate the transition from short-term to long-term memory, thereby aligning with the vision of continuous, unbounded evolution.

Q：What are the main limitations of existing methods for multimodal information processing?

A：A major challenge in multimodal information processing is the sheer volume of data and its difficulty to compress, especially when the information must be updated dynamically; in such cases, conventional text-compression methods yield limited results. The more fundamental issue is that relying solely on external memory cannot enable an agent to achieve genuine internal evolution, just as individuals with memory impairments cannot sustain learning and progress by depending exclusively on notes.

Q：How can we understand and address the core challenges of online learning?

A：The core challenges of online learning include modeling urbanized memory—addressing the effective representation of both short-term and long-term memory—and reforming the learning algorithms themselves. In the realm of algorithmic learning, it is necessary to explore efficient learning paradigms beyond the current approaches of imitation learning and reinforcement learning. For instance, leveraging on-policy methods—where the policy itself serves as a model—to distill and learn from self-generated data can significantly enhance learning efficiency.

Q：For the modeling of massive-scale memory, does the current model architecture require fundamental innovation?

A：Existing model architectures, particularly those based on transformers, are limited in their ability to model memory at different levels, such as short-term memory, working memory, episodic memory, and long-term memory. They typically manifest as a single loss list structure, making it difficult to effectively select and compress the memory to retain only the most relevant information. Therefore, it is indeed necessary to reform the architecture so that it can proactively select and compress information, thereby better emulating human memory mechanisms.

Q：In the field of AI, what misconceptions do people currently have about the algorithms themselves?

A：Currently, there is a significant misunderstanding about AI algorithms, especially regarding their use as universal tools. Many people have not fully grasped the principles underlying algorithms or the importance of data.

Q：What are the main challenges AI faces as it moves from the virtual world into the physical world?

A：When AI ventures into the physical world, the first challenge it faces is the data problem. Compared with the abundant data resources available on the internet, data in the physical world is exceedingly scarce. For example, training a model typically requires large amounts of data to achieve high accuracy, yet acquiring such high-quality real-world data poses a significant challenge.

Q：What advantages does China have in the field of AI technology?

A：China possesses a natural and significant advantage in AI technology, primarily because its diverse and highly demanding industrial, service, and home-appliance sectors provide abundant real-world application scenarios, generating vast amounts of data to support model training. Furthermore, Chinese researchers tend to share their technologies and algorithmic architectures when publishing papers, thereby fostering the collective advancement of the field.

Q：Is there a contradiction between technology sharing and the innovativeness of the technology itself?

A：Technology sharing does not imply that technology is unimportant. In fact, technological advances such as reinforcement learning, self-attention mechanisms, and low-level optimization techniques have all driven the industry’s progress by being shared with the community through open‑access papers and code. Even within the current foundation model landscape, innovation in algorithmic architectures remains relentless; for instance, Loop’s Transformer and Loop’s language model are among the latest breakthroughs.

Q：Regarding predictions about the future development of AI, what are some things that might be overestimated or underestimated?

A：Predicting the future development of AI is extremely challenging. What may be overestimated is the extent to which AI will replace humans and how well automation performs in solving real-world problems; meanwhile, what may be underestimated is the impact of AI on the way society operates—such as the modes of interaction among intelligent agents and the payment systems—which could have far-reaching consequences.

Q：What was the original intention behind establishing the SenseTime Scholarship? How have the recipients of the SenseTime Scholarship fared in their subsequent development?

A：The original intention behind establishing the SenseTime Scholarship is to provide support to undergraduates at their most energetic and inquisitive years, just as Professor Tang once led us through the fog of scientific research and opened the door to a new world of AI. We, too, hope to help nurture your dreams, witness you become leaders and pioneers in the field, and pass down this passion for AI from generation to generation. Over the past decade, the 280 winners of the SenseTime Scholarship have come to epitomize the nation’s top undergraduate students in the AI field. Starting out in computer vision and deep learning, they have since ventured into cutting-edge fields such as large models, multimodal agents, and embodied intelligence. More than 90 percent have chosen to pursue advanced studies, with some dedicating themselves to academia and others moving into industry.

Q：What significant impact has the SenseTime Scholarship had over the past decade?

A：The past decade has been a pivotal period in the growth of outstanding students. The SenseTime Scholarship has served as a guiding light, illuminating their research journey and helping them to continuously explore and thrive in the field of AI. At the same time, it also reflects Teacher Tang’s appreciation for talent in education and his long-term vision for the future development of AI.

Q：Where will AI head in the next decade?

A：Although no one can provide all the answers, it is certain that among those who tackle these questions will be future recipients of the SenseTime Scholarship. In the next decade, AI will be integrated into more fields and become a key driver of social development.

Q：Where does intelligence come from, and what is the driving force behind its self‑evolution?

A：Intelligence and self-evolution arise from multiple sources, including data, experience, and cross-validation. Currently, the transformations brought about by the industrial revolution indicate that services with high model‑usage volumes are increasingly evolving into widely recognized, general‑purpose applications. This shift is driven by the advancement of core technologies that effectively address real‑world problems and by innovations that genuinely expand the scope of applications to reach a broader user base.

Q：How does a unified generative model achieve system-level delivery?

A：Through continuous iteration and refinement of a unified generative model, we have transitioned from generating isolated content to delivering comprehensive, system‑level outputs. For example, the introduction of the U1 Pro mode has significantly boosted model throughput, enabling more refined, stylistically consistent generated content that can be tailored and edited for diverse application scenarios. This ultimately yields a deployable AI foundation capable of powering various domains and producing high‑quality, distinctly styled outputs.

Q：What is the main content of the image generated this time?

A：This chart illustrates the nine-year evolution of the World Artificial Intelligence Conference, from the integration of AI technology with various industries in 2018 to the governance framework and global governance progress proposed in 2024, as well as the latest achievement—the establishment of global AI governance cooperation in Shanghai.

Q：Can this delivery system be applied to content‑generation workflows?

A：The answer is yes: the system can decompose a screenplay into multimodal components—such as character designs and storyboard scripts—and generate corresponding visual assets based on this information. Creators can check the overall progress at any time and generate the next phase of video content.

Q：How does the expansion of AI capabilities affect ordinary people and industry applications?

A：As the boundaries of AI capabilities expand, many tasks once reserved for professionals have become feasible, giving rise to super-individuals who leverage AI tools to scale their personal expertise and catalyze the emergence of industry‑level applications. For example, landscape review engineers leverage AI models to codify and extend their expertise, while the office assistant “raccoon work” helps professionals streamline their workflows more efficiently.

Q：How can AI empower the integration of interest-driven demand and applications?

A：An enthusiast with a passion for history has used AI technology to create an interactive spatiotemporal story map, presenting the lives of historical figures to the public in a more intuitive and personalized way—this is a new application scenario born from the effective integration of interest and need.

Q：What is the specific impact of AI on the field of education for visually impaired individuals?

A：A visually impaired individual has leveraged multimodal AI technology to develop an educational system tailored for the visually impaired, transforming instructional materials into formats that meet their needs and significantly enhancing the quality and efficiency of accessibility services.

Q：What is SenseTime’s track record in AI technology development, and what are its plans for the future?

A：Over the past decade, SenseTime has been committed to bringing AI into industries, addressing the issue of proprietary AI, and continuously expanding the frontiers of AI. Now, it is empowering super individuals and everyday users with industry‑scale AI capabilities, thereby sparking broader business transformations. The future goal is to empower everyone with industrial‑grade AI productivity, enabling ordinary people to use AI tools for high‑quality content creation.

Q：Why do we focus on super‑individuals and high‑quality AI products?

A：We focus on super individuals because the impact they generate far exceeds expectations, and they can replicate similar use cases to accelerate AI adoption. A truly effective AI product does not merely replace existing workflows to enhance performance; rather, it extends into broader domains, making its capabilities accessible to a wider audience, expanding the boundaries of what is possible, and unlocking new application scenarios and user needs.

Q：How can a well-designed AI product help individuals grow?

A：A great AI product not only fosters user dependency but also focuses on personal growth. By using AI tools, users may develop better habits, enhance their ability to acquire knowledge, and deepen their thinking. For example, with the help of an AI assistant, users can create videos, infographics, and more, thereby fostering personal growth.

Q：What is the role of artificial intelligence in predicting the future?

A：The real problem that artificial intelligence addresses is not predicting the future, but rather forecasting the current consensus of information. Although predicting future events is challenging, AI has demonstrated its application in forecasting the consensus on current information, such as predicting Cape Verde’s qualification for the round of 32.

Q：How does SenseTime integrate AI into everyday life and help people achieve their goals?

A：SenseTime is committed to developing the kapi family of AI assistants that can understand users’ lives, goals, and skills, and help them achieve their life objectives. Among them, Kaimo Health serves as a member, offering round‑the‑clock health support by monitoring users’ health metrics, providing long‑term companionship, and helping them achieve their wellness goals.

Q：What achievements has the Large-Scale Facilities Business Group attained in the field of artificial intelligence infrastructure?

A：As an explorer and pioneer of AI infrastructure, the Large-Scale Facilities Business Group has achieved significant results over the past year and more, such as being recognized as the nation’s first 5A‑level intelligent computing center, securing a leading position in both the AI cloud market and the AI solutions market, and contributing to the definition of future infrastructure. At the same time, in response to industry changes, the large-scale facility is also poised to build a next-generation infrastructure capable of supporting future transformations and innovations.

Q：On domestically produced chips, what specific improvements have you achieved through optimization and tuning?

A：Through approximately three months of optimization and performance tuning, we successfully increased the average effective throughput of domestically produced single-GPU systems by 100%. In an environment built on a domestically developed heterogeneous hybrid inference cluster, domestic GPUs outperform NVIDIA’s H‑series in specialized workloads such as PREFU, achieving more than double the performance of the latter; internal benchmarks further indicate a performance improvement of over 200%.

Q：What impact does this performance improvement have on the large-scale deployment of domestically developed inference clusters?

A：Thanks to performance enhancements, we have achieved the large-scale commercial deployment of a truly domestically developed inference cluster, and this deployment is commercially viable. Over the past six months, our large-scale model has been actively promoted, and we currently deliver an average of 2.4 trillion token‑level API calls per day, with projections to scale up to the 10‑trillion‑call range by year-end.

Q：What is SenseTime’s strategy for AI infrastructure?

A：In addition to providing token services, SenseTime is also continuously optimizing its software algorithms, supporting a variety of AI-driven innovation scenarios and algorithms, such as text-to-video, text-to-image, multimodal models, and AI for science. Meanwhile, operator optimization enables multi-card parallelism to boost performance, and automated migration tools help customers cost-effectively port their existing AI models to domestic chip architectures.

Q：What work has SenseTime undertaken at the infrastructure and data center level?

A：At the data center level, SenseTime focuses on power consumption and efficiency, introducing a new metric, TPW—token output per unit of electricity cost. This metric has now achieved a more than 50% improvement over the industry average. Through AI-driven prediction and scheduling, electricity prices and the actual effective electricity price were further reduced, POE utilization was improved, and overall system-side output increased by more than 80%.

Q：How does SenseTime view the future development trends of artificial intelligence infrastructure?

A：SenseTime believes that in the future, AI infrastructure should not only prioritize efficiency but also environmental sustainability. Meanwhile, as domestically developed computing‑power systems achieve large‑scale commercial deployment, the entire supply chain will undergo a reshaping, giving rise to new technological innovations and applications. For example, emerging fields such as optical computing, quantum computing, and space-based computing power will provide stronger support for AI algorithm iteration, and we plan to launch a forward-looking initiative for space-based computing power. The first batch will consist of four satellites tasked with technology validation and iterative development, with the goal of establishing a constellation of one thousand satellites within the next five years, delivering ten thousand units of space-based computing power to serve China’s downstream AI application ecosystem.

Q：How is SenseTime advancing ecosystem development, and what are its future plans?

A：SenseTime has launched the Galaxy Initiative, a national industrial innovation ecosystem program, which aims to build five domestically produced clusters, each with over 10,000 AI chips, within approximately two years. The initiative will foster deep co-creation across ten technological domains, support more than 200 AI startups, and accelerate the large-scale, commercially sustainable deployment of the domestic AI ecosystem. In addition, SenseTime is deploying AI infrastructure in various locations both domestically and internationally, such as Shanghai Lingang, Hong Kong Island, and the Middle East, and is committed to leveraging these deployments as core support for the real-world application of intelligent scenarios, thereby driving the intelligent transformation of various industries.

Q：In the action plan released by the State Council last year, why was the integration of AI with scientific research designated as the top priority for development? How has SenseTime collaborated in building an AI‑driven platform for scientific discovery?

A：Because the country places great hopes on leveraging AI to achieve deep integration with traditional scientific disciplines and drive transformative innovation in fundamental capabilities, it has prioritized AI‑augmented research as the top priority among all future breakthrough areas. As an infrastructure‑as‑a‑service provider, SenseTime collaborates with national-level research institutions to deliver computational power and underlying platform capabilities, jointly advancing research in emerging fields, cultivating talent, and facilitating industrial application, thereby establishing a fully integrated closed loop.

Q：How was the strategic brand partnership for scientific discovery announced?

A：Through the organization of an on-site signing and launch ceremony, industry experts including Mr. Shi Boming, Assistant Director of the Shanghai Artificial Intelligence Laboratory, joined SenseTime in jointly signing a cooperation agreement.

Q：What was SenseTime’s original intention in establishing the Large-Scale Scientific Facility and Computing Power Alliance?

A：SenseTime is committed to bringing together domestic software and hardware vendors to achieve large-scale commercialization in the development of foundational AI infrastructure.

Q：What are your views on the future prospects of artificial intelligence infrastructure?

A：Over the next few years, China’s domestic infrastructure hardware ecosystem is expected to experience robust new growth—driven not only by technological competition but also by the successful deployment of sustainable, profitable, and scalable business models. This requires a collective effort across the entire industry to build a more robust and resilient AI infrastructure.

Q：How should we view the development and cooperation of artificial intelligence technologies in the context of globalization? What are the main topics of discussion at the roundtable?

A：Globalization is an inevitable trend in the development of artificial intelligence; only through openness and symbiosis can we ignite innovation and illuminate a broader world. By launching overseas innovation partnerships, the aim is to help Chinese-style innovation secure a new position within the global AI ecosystem and, together with other countries and regions, build more robust cooperation models. The main topic of discussion is how China’s innovation is reshaping regional cooperation models, particularly in AI and other high-tech fields, and how to draw on China’s successful innovation experiences to promote development and transformation in this region.

Q：What are the key characteristics of China's innovation that differ from Japan?

A：The key difference between China's and Japan's innovation lies in speed and scale. China's innovation is characterized by rapid speed and fast pace, which is critical for creating new value, while Japan prioritizes quality and reliability, resulting in a slower pace.

Q：What does the term 'China vision' signify and how has its definition evolved over time?

A：'China vision' is a term used to describe the concept of Chinese-style innovation. It is an old concept that emerged 15 years ago, which is considerably older than the concept of Samsung. The original definition aimed to understand emerging Chinese players in the international market. Today, the definition is changing and becoming a journey to discover a new world, reflecting China's transformation from being known as a copycat nation to being recognized for innovation.

Q：How has China's perception changed from being a 'copycat nation' to being innovative?

A：China's perception has significantly changed from being known as a 'copycat nation' to being globally recognized as very innovative. This transformation has occurred over the last ten to fifteen years as China has developed its own unique brands and technologies, leading to discussions around its innovative capabilities.

Q：What potential collaboration opportunities exist between Malaysia and China in the field of innovation?

A：Potential collaboration opportunities between Malaysia and China in innovation include working in hyper-scale data centers, where Malaysia is expected to be one of the top five players by 2023. Practical cooperation would involve leveraging local knowledge, understanding regulatory requirements and culture, and utilizing local talent to address real problems with innovative solutions.

Q：How should innovation in China be approached to solve real problems?

A：Innovation in China should be approached by understanding the real problems and utilizing vertical AI and local knowledge to provide solutions. It involves a deep understanding of local regulations, culture, and leveraging local talent, as well as adopting a cooperative mindset rather than just exporting technology and innovation.

Q：What is the significance of the Korea-China cooperation in technology sectors such as semiconductors and AI?

A：The significance of Korea-China cooperation in technology sectors lies in the complementary nature of their industries and the strategic timing presented by investment projects like the 'korean three mega project.' This cooperation can lead to significant synergies in global standard setting, supply chain integration, and the creation of a robust ecosystem encompassing manufacturing, data centers, and application capabilities.

Q：What role does NCS play in connecting Chinese innovation with global markets?

A：NCS acts as a super connector, introducing local good startups and technical partners from China to broader communities. NCS is shaping a strategy called 'IT China Technology Sourcing' to explore collaboration opportunities with different Chinese technology companies and help them introduce their innovations into global markets while also assisting with localization, understanding regulatory constraints, and client requirements.

Q：What are the potential opportunities for Chinese innovation in the Japanese market?

A：Chinese innovation holds potential opportunities in the Japanese market by addressing issues like labor shortages and addressing societal challenges. Examples include the introduction of robots that solve labor shortages in food and dairy sectors, the deployment of cloud peak unmanned retail stores to tackle retail labor issues, and the use of human or robotics in Japanese warehouses and manufacturing factories.

Q：What is the significance of global expansion for Chinese technology companies?

A：Global expansion is significant for Chinese technology companies as it represents the starting point of their globalization journey. It presents opportunities to address new markets' unique challenges and requirements. It involves appreciating the benefits of diversity, understanding the distinct values and norms of different cultures, and adapting products and services accordingly.

Q：What are the potential advantages of a stronger influence from the east in the context of the valuation?

A：A stronger influence from the east could potentially provide a deeper knowledge base in producing IT that is more detailed than being just conceptual. The hands-on work and attention to details acquired from the east could be advantageous moving forward.

Q：What are the examples of Chinese innovation in robotics and AI presented?

A：Examples of Chinese innovation in robotics and AI include a robot company in Hangzhou that has successfully entered the Korean market and is using the robot for empower inspection and emergency response scenarios. Another example is a partnership between a South Korean company, Echo, and LG Electronics, which is an AI investment, showcasing global cooperation in the AI space.

Q：How is Singapore positioned in the context of AI strategy?

A：Singapore is seen as a demo space for AI in Southeast Asia and has announced the Singapore AI Strategy in 2020. The focus is shifting from implementing AI infrastructure to using AI in real business scenarios and supporting AI adoption across major industry sectors.

Q：What are the risks associated with technology competition?

A：One important risk is maintaining neutrality in technology competition, which is a topic that affects governments, enterprises, and consumers. When defining an AI strategy, enterprises need to consider this and adopt a more conservative architectural design to support long-term growth.

Q：How is IT AI fusion assisting clients with AI implementation?

A：IT AI fusion is a solution that focuses on helping clients implement or consume AI in a more efficient way without worrying about the specifics of the underlying language model. This allows for a more streamlined approach to AI integration.

Q：What is the significance of collaboration between east Asian countries in AI?

A：Collaboration between east Asian countries in AI is significant because it combines strengths such as robust semiconductor industries in Korea, advanced materials for semiconductors in Japan, scalable supply chains and AI technology in China, and a wealth of emerging robots and talented people across the region. This collaboration can create a significant global impact.

Q：What potential impacts could the establishment of an AI ecosystem in east Asia have?

A：The establishment of an AI ecosystem in east Asia could create great value and foster greater communication and collaboration among countries in the region. This could help mitigate risks and prevent nations from isolating themselves in the technology field.

Q：What is the risk management approach for AI?

A：Risk management for AI involves acknowledging the existence of risks, identifying them, and determining their impact on objectives. A mitigation plan must be devised, open-mindedness is essential for finding resolutions, communication is key, and global guidelines and collaboration are necessary to address the broad impact of AI.

Q：Why is global collaboration important in AI?

A：Global collaboration is crucial in AI due to the cross-border nature of the industry, the need for consistent guidelines and standards, and to prevent any single entity from having too much control. Open communication, working with multiple service providers, and engaging in government-to-government collaborations are part of a broader risk management strategy.

Q：What did the founder of the company establish under his leadership?

A：Under his leadership, the founder established the first AI ethics committee within the company, which was one of the first of its kind.

Q：What risks does the increasing intelligence of AI bring?

A：As AI becomes more intelligent and prevalent in our lives, it also brings an accumulation of risks that require management.

Q：How is AI reshaping our lives and perspectives on technology?

A：AI is becoming more open source, more accessible, and available to both consumers and enterprises, affecting various aspects of our lives and redefining how we view technology innovation.

Q：Why are Asian values and culture important in the context of AI and technology innovation?

A：Asian, particularly East Asian, values and culture are important because they are amenable and consistent with the current AI innovation, focusing on the human side of technology which includes values and the understanding of the importance of humanity.

Q：What is the 'bitter lesson' in AI research and how does it relate to large language models and foundation models?

A：The 'bitter lesson' is a concept that highlights the historical pattern in AI research where reliance on human knowledge initially helps but eventually plateaus and is superseded by methods based on scaling, computation, search, and learning. In the context of large language models and foundation models, there is a discussion about whether this pattern will apply, given that these models are based on human knowledge yet aim to scale and create knowledge through search and learning.

Q：What is the fundamental question regarding foundation models and their ability to encompass all knowledge?

A：The fundamental question regarding foundation models is whether they are meant to be containers of all agreed human knowledge, which is a complex notion, or all knowledge, which is unattainable. The idea is to acknowledge the limits of foundation models and set a realistic goal for them to be accurate stores of current agreed-upon knowledge.

有用

Feedback

免责声明：本内容基于第三方人工智能AI算法生成，内容仅供参考。本平台不对其真实性、合法性、准确性负责，不构成任何投资建议，对以上内容不做任何担保与承诺。若有异议请与本平台反馈。