Wen/Chen Yongwei

The curtain of 2025 has come to an end, and undoubtedly, the AI industry has gone through a milestone in this year. From the innovation of technological paradigms, to the reconstruction of business logic, from the implementation of industrial applications, to the game of global rules, this year has seen both breakthroughs and many reflections.

Given the complex development of AI, we can only provide a brief review from ten aspects.

1? Multi mode fusion

In the past few years, AI models have made rapid progress in areas such as text and reasoning, but their multimodal capabilities have been relatively slow to develop, which greatly limits their ability to be fully utilized. For example, before version 4.0, although GPT was capable of writing poetry and programming, it could neither see nor draw. If users wanted it to analyze what an image said or generate an image according to requirements, it would be inadequate.

Although AI developers have been vigorously developing the multimodal capabilities of models since 2024, for a considerable period of time, these efforts have mainly focused on combining existing models - text as a system, image as a system, speech as another system, and then using engineering methods to assemble them together. This model can accomplish some multimodal tasks, but its capability limitations have always been evident due to coordination issues between various systems. By 2025, more and more developers will no longer be satisfied with this "assembly style" solution, but will start designing "native multimodal" models that can simultaneously process text, images, audio, video, and other information in the same system from the beginning of training.

In the process of designing native multimodal models, many people gradually realize that the real bottleneck of multimodal models is not whether they can see and hear images, but whether they can organize vision, language, time, and space into a unified representation that can be inferred and acted upon. The current large-scale models are highly mature in language, but still lack a systematic understanding of three-dimensional space, physical constraints, and causal relationships. Therefore, the breakthrough focus of the next generation of AI may not lie in a few more parameters, but in the ability to construct internal models of the real world that can "rehearse" the consequences of actions in the brain. Under this approach, the "world model" and "spatial intelligence" have been re emphasized and gradually become important theoretical basis for designing native multimodal models.

At the practical level, multimodal models have almost become the main battlefield for top AI companies. Enterprises are no longer satisfied with "being able to see pictures", but are pushing their ability to see accurately, comprehensively, and understand processes, and transforming visual understanding into executable actions: some strengthen the integration of visual and language expression, some extend their ability to video and long-term understanding, and some try to transform "understanding screens" into interface operations. The common trend is that models are no longer just answering questions, but increasingly intervening in real tasks themselves.

The flourishing development of multimodal models has provided strong support for the application and popularization of AI technology. McKinsey pointed out in its "Technology Trends Outlook 2025" that multimodal AI is becoming a key foundation for the new generation of AI systems and will be deeply integrated with agents, automation, and physical system control. Its impact will extend from the digital world to the real world. Gartner emphasizes that the impact of multimodal technology will go beyond the scope of AI and comprehensively reshape the current software ecosystem. According to its prediction, by 2030, 80% of enterprise software will incorporate multimodal AI capabilities into their products.

2? Embodied outburst

If multimodal fusion solves the problems of "how to see" and "how to understand" AI, then embodied AI needs to answer what AI does and how it does it in the real world. Although the development of embodied intelligence has been going on for decades, in the past, robots were mostly in the laboratory demonstration stage: completing cool actions, doing local automation in limited scenarios, and people only discussed what they could technically achieve. By 2025, the situation will begin to change - embodied robots will truly enter the market, and the industry narrative will shift from "can it be achieved" to "can it be scaled up, can it work stably, and can it enter positions".

The most direct signal comes from the change in production pace. In 2025, domestic companies such as Yushu and Ubiquitous, as well as foreign companies such as Boston Dynamics and Apptronik, will clearly announce that their products will enter the mass production and commercialization pilot stage, rather than just a single prototype or demonstration prototype. In major markets represented by North America and China, embodied robots have begun to plan their supply chain, manufacturing, and delivery according to the "hundred level" and "thousand level". IDC estimates that the global pilot applications in warehousing, manufacturing, inspection and other positions will increase several times by 2025 compared to 2024.

At the same time, there has been a significant decrease in costs. A few years ago, humanoid robots that could interact with humans were priced at tens of thousands or even millions of dollars, making them almost impossible to commercialize; According to data from the Bank of America Research Institute, the price of a typical humanoid robot has dropped to around 35000 yuan per unit, a decrease of at least 40% from 2023, and there is still room for further decline in the coming years. This cost reduction is rapidly lowering the threshold for use, making it easier for robots to enter factories and homes.

The explosion of embodied intelligence in 2025 is the result of multiple factors working together. On the one hand, the development of native multimodal AI has filled the gap of robots' inability to understand the world, allowing them to not only run and jump, but also understand the environment and make decisions based on the situation, thus possessing true practical value; On the other hand, the continuous rise in labor costs is also expanding market demand. Manufacturing, logistics, inspection, service and other fields are already labor-intensive. With factors such as aging population, young people's unwillingness to enter high-intensity positions, and increased compliance and safety costs, enterprises have begun to accept robot substitution, which has driven rapid growth in demand.

3? Competition in computing power

If the core of computing power competition in the past few years was' who can grab more GPUs', then by 2025, this competition has evolved from a capital driven resource struggle to a long-term, three-dimensional, and even geographically significant comprehensive game.

Firstly, the competition in computing power has shifted from a "scale oriented" to an "efficiency oriented" approach. In the past, people generally believed that when a model encountered a bottleneck, it would continue to stack parameters and computing power; But as the scale expands, costs and complexity rapidly increase, while marginal benefits continue to decrease. More and more companies are realizing that the difference is no longer determined by the total amount of computing power, but by how to transform limited computing power into effective capabilities. The domestic model DeepSeek is a typical example - through strategies such as multidimensional parallelism and mixed precision, it achieves near performance at lower investment, and is seen as a symbol of "efficiency shift".

Secondly, some developers have started to directly intervene in the chip process. For a long time, GPUs have been the sole mainstay of large model training, and Nvidia's supply and ecosystem have had a decisive impact. To reduce dependence, some companies have turned to self-developed chips. Google's TPU is a representative, which has now replaced GPUs on a large scale in core model training, gradually building a relatively autonomous computing power system.

Thirdly, computing power facilities have begun to be fully 'infrastructure oriented'. In the past, enterprise expansion only required purchasing servers, renting computer rooms, and adding cloud services; In the era of large models, this "modular" approach is difficult to meet the demands of high-density parallelism and stability. The intelligent computing center of 2025 has been designed around AI load from the beginning: the network topology is centered around parallel training, and even site selection requires simultaneous evaluation of power, energy consumption, and latency. Some companies have also ventured upstream through investments and mergers and acquisitions, betting on potential directions including 'space computing power'.

4? Paradigm controversy

At a time when the industry was showing its tricks to win the competition for computing power, the theoretical community began to reflect on the theoretical basis of sustained computing power investment - the "rule of scale". In the past decade, this has almost been a repeatedly validated iron law: more data, larger models, and stronger computing power often lead to better results, thereby shaping research paradigms, capital logic, and industrial structure. But by 2025, there will be a systematic divergence of confidence around this path for the first time.

For example, Turing Award winner Yang Likun has repeatedly pointed out that simply expanding autoregressive models will not naturally lead to general intelligence. The current models that focus on language prediction are essentially passive systems, lacking causal understanding, physical knowledge, and long-term planning capabilities; Continuing to stack parameters and data will only lead to diminishing returns, but will also mask structural weaknesses. Similar views have also been acknowledged by other researchers. Andr ? Capas emphasized that future breakthroughs are more likely to come from training paradigms, data structures, and inference mechanisms, rather than parameter scales themselves; Former OpenAI chief scientist Ilya Sutzkwei also warned that the 'infinite heap size' may be approaching its stage limit.

However, the rule of scale still has many supporters. They believe that the path centered on large models is still advancing, and the boundaries of capabilities are still expanding, especially in multimodal and complex reasoning tasks, where scale remains an important prerequisite. DeepMind co-founder Demis Hasabi believes that the key to achieving higher levels of intelligence lies in the world model, planning, and reasoning structure - these are not negating scale, but rather direction correction based on scale. In this perspective, the rule of scale may not necessarily end, but may gain new life after the emergence of new paradigms.

It will take time to verify whether the scale rule continues to be effective. But when we look at both supporting and opposing views together, we find that both sides are actually dissatisfied with the existing paradigm. Perhaps it is this controversy that will drive the AI industry to re-examine existing ideas and seek a better development path.

5? The Rise of Agents

In March 2025, startup Monica released an AI application demonstration video called Manus. In the video, it can automatically call external tools according to user requirements to complete complex tasks such as resume screening, real estate research, stock analysis, etc., without the need for manual intervention throughout the process. Manus quickly became popular and was regarded as another "phenomenal" product after DeepSeek; Although it later sparked controversy due to "insufficient technological innovation" and "excessive hype", it is still seen as an important symbol of the rise of AI intelligent agents.More interestingly, Meta subsequently acquired Manus in the billions, indirectly confirming this trend.

AI agents are different from traditional large models in that they are not just "smarter conversational interfaces", but rather entities capable of understanding goals, breaking down tasks, calling tools, performing operations, and adjusting strategies based on feedback. It has brought about a change in the way people interact with machines: from "people searching for functions" to "task driven systems". In the past, users needed to understand the interface structure and switch back and forth between multiple pages; In the intelligent agent mode, as long as a goal is proposed, the system automatically plans the path and requests confirmation at key nodes. This transformation may seem subtle, but it significantly reduces the cost of learning and use.

It should be emphasized that the emergence of intelligent agents is not accidental, but the result of multiple mature technologies: on the one hand, the ability of large models to reason, multimodality, and long context is enhanced, enabling them to understand complex tasks; On the other hand, protocols such as MCP, ANP, A2A, etc. have gradually standardized tool calls and external system access. AI is no longer limited to "just speaking", but can operate code, documents, databases, and even business systems and physical devices. Therefore, for the first time, it has the feasibility of completing tasks across steps and systems.

It can be foreseen that intelligent agents will have a profound impact on the industrial ecosystem. It will not only automate certain positions, but also restructure the process itself - the organizational structure will gradually shift from being designed around "people" to being organized around "tasks". At the business level, value measurement will also shift from "providing capabilities" to "completing tasks": companies will no longer pay based on call volume, but are more likely to pay based on tasks and results, thereby driving changes in business models.

Of course, the rise of intelligent agents also comes with risks. Issues such as employment substitution, privacy, and data usage boundaries may all become obstacles in the process of popularization. The controversy surrounding the Doubao AI smartphone has already revealed the tip of the iceberg for us, and similar discussions will continue in the near future.

6? The Age of Open Source Prosperity

By 2025, open source models have gradually evolved from a peripheral force in the AI world to a global infrastructure for innovation. During this year, open-source models have fully approached closed source systems in terms of performance, ecology, and adoption rates, and even surpassed them in some scenarios. Multiple authoritative reports have shown that in the past two years, open source or "open weight" models have accounted for the majority of newly released large models, and among the actively called models, open source models also dominate, especially in scenarios such as private deployment, fine-tuning, and intelligent agents.

The rise of open source models is not only due to their "free" nature, but also because the logic of AI innovation is undergoing changes: the cost of computing power is increasing, application demands are highly differentiated, and closed models are difficult to cover all scenarios; The open-source model, relying on community collaboration and rapid customization, has shown significant advantages in engineering efficiency and adaptability, attracting more and more enterprises and developers to join.

In this context, the division of labor structure in AI innovation is undergoing restructuring - the underlying model is no longer the "end product", but more like the underlying platform of an operating system or database; True innovation is increasingly occurring in fine-tuning models, toolchains, intelligent agent architectures, and industry applications. Open source is no longer just an idealistic choice, but a practical mechanism for lowering barriers and accelerating diffusion.

In this wave, China's power is particularly prominent. A group of models represented by DeepSeek and Qwen have formed distinct advantages in engineering efficiency, inference cost, and deployability. According to relevant statistics, by 2025, nearly 30% of global open source models will be sourced from China, making it one of the most important suppliers in the global open source ecosystem.

In the longer term, the "open source era" of 2025 is not simply a resurgence of the path, but one of the signs that AI has entered a mature stage: when innovation no longer overly relies on closed systems and capital stacking, but more on open collaboration and engineering wisdom, whoever can provide a "usable, modifiable, and scalable" technological foundation will occupy a more advantageous position in the next stage.

7? Business Innovation

In the past two years, the "difficulty in making profits" has been troubling the AI industry. Many companies, even if they produce good products, still struggle to generate stable income and are ultimately forced to exit the market. By 2025, the industry will gradually explore new business paths - players at different levels will find monetization methods that match their own abilities, and AI will shift from a single technology competition to a more clearly defined industrial ecosystem.

At the technical level, capabilities are beginning to be commodified. Computing power, training, and reasoning are standardized as measurable and priced "production factors", and cloud vendors, chip companies, and basic model providers form a relatively stable revenue structure through computing power leasing, APIs, and reasoning services. With the transformation of computing infrastructure and the improvement of model efficiency, this layer is gradually shifting from an "arms race" to an "operational competition", with unit costs decreasing and unit call values increasing.

At the platform service layer, "Outcome-as-a-Service" (OaaS) has begun to rise. In the past, AI products were mostly focused on "selling functions", while the popularity of intelligent agents has made "completing tasks" the core value unit, and pricing has gradually shifted towards charging based on tasks, processes, or results. Although it is not a long time, this model is becoming the zone with the greatest profit potential and the most fierce competition.

At the application layer, scene differentiation is clearly mature. In 2025, an important change is the convergence of the imagination space for general applications, while the value of vertical industries is gradually released. Whether in software development, enterprise operations, financial analysis, or content and customer service, AI is deeply embedded in business processes, coupled with data and rules, gradually becoming a long-term system investment for enterprises, rather than a one-time tool procurement.

These three layers, when stacked together, mark a turning point: the business logic of AI is moving from "capability demonstration" to "efficiency realization". In the past, larger models were easier to attract attention and financing; In 2025, what will truly determine success or failure is who can steadily transform capabilities into measurable customer value, which will also prompt more and more companies to focus on engineering efficiency, deployment costs, and user retention.

8? Rule-based game

If the main contradiction in the development of AI before was concentrated on the technical level, then by 2025, another equally important front - AI governance - has been fully launched. It can be understood from two dimensions: the horizontal tension between innovation and rules, and the vertical game between different institutional systems.

From a horizontal perspective, the contradiction does not lie in whether to regulate or not, but in how to avoid rules prematurely locking in the yet to be finalized technological path. The uncertainty of AI poses a challenge to the traditional approach of "setting rules before admission": if formulated too early, it may solidify existing forms; If left too loose, it may accumulate systemic risks. A significant change in 2025 is that governance is gradually shifting from "static compliance" to "dynamic calibration", keeping pace with technological evolution through a layered, phased, and adjustable approach.

This shift stems from a renewed understanding of the essence of AI: AI is not a single product, but a continuously evolving system of capabilities. Therefore, the governance object extends from a single model to a complete chain of data, computing power, model training, deployment, and usage scenarios. Consensus is forming - the goal of governance is not to slow down innovation speed, but to avoid risk amplification at irreversible nodes, and mechanisms such as "sandbox", "hierarchical management", and "post correction" are gradually replacing the "one size fits all" approach.

From a vertical perspective, governance is evolving into a competition between the state and institutions. Rules are no longer just internal order tools, but have spillover effects: whoever's rules are more easily adopted gains greater institutional influence.

In this dimension, the United States tends to view governance as a "guardrail", prioritizing national security, critical infrastructure, and extreme risks, and minimizing pre restrictions on research and development; The logic is to exchange technological leadership for governance flexibility.

The EU emphasizes clarifying institutional boundaries before proliferation, shaping development direction through systematic rules, and participating in global competition with rules and standards.

China emphasizes the adaptation of development order and scenarios, synchronous adjustment of rules and industry promotion, and does not freeze technology routes in advance or overly rely on post correction, but puts forward higher requirements for governance capabilities.

When these different paths meet at the global level, governance becomes a game of rules: companies choose to land between different systems, technology evolves and differentiates in different rules, and standards are tested and diffused in competition. The global governance pattern is therefore more likely to present a state of multiple modes coexisting and influencing each other.

From this perspective, the governance game of 2025 is not a "decelerator" for the development of AI, but one of its signs of maturity - when technology is systematically governed, it means that it is both important enough and potentially risky. How to form a dynamic balance between innovation openness and rule constraints will determine whether AI can be embedded in social structures in the long run and stably.

9? Competition among major powers

By 2025, international competition for AI has risen from the enterprise level to the national level: who will define the technological path for the next generation of AI? Who can control the chip and computing power supply chain? Who has the ability to translate technological choices into globally accepted standards? Around these issues, China, the United States, and Europe have gradually formed a competitive landscape that is misaligned with each other but highly entangled.

At the core technology level, the United States still holds the strongest voice. This advantage is not only reflected in model performance, but also in the "problem definition power" - from large models and multimodality to world models and intelligent agent architectures, many key directions are often first proposed by the United States, and the accompanying evaluation methods and technical narratives naturally become the industry's default reference.

China's path is clearly different. We did not bet all our chips on "redefining paradigms", but emphasized more on transforming technology into replicable capabilities through engineering optimization, system integration, and real-world feedback within existing technological frameworks, gradually forming advantages in training efficiency, computing power scheduling, embodied intelligence, and industrial level applications.

The EU's presence in core technology originality is relatively limited, but it is not absent. It maintains its influence on key concepts and methodologies through basic research networks, cross-border scientific research projects, and evaluation systems, and reserves a position for subsequent standard development.

If the core technology determines the long-term limit, then the chip and computing power supply chain determine who can sustain the technology. The dominance of the United States in high-end chip design, advanced processes, and software ecology has gradually endowed computing power with strategic attributes; Under pressure, China is accelerating the construction of a multi-path computing power system. On the one hand, it is filling in local chip and manufacturing capabilities, and on the other hand, it is enhancing resilience through intelligent computing centers and algorithm optimization; Although Europe does not have an advantage in scale, it is still irreplaceable in equipment, materials, and some key process nodes.

As AI moves from software tools to infrastructure, the importance of standard setting power is rapidly increasing and has become the most covert and persistent battlefield of competition. The United States relies more on technological leadership to form "factual standards", China forms "standards used" in large-scale deployments, and Europe attempts to continue to speak out in the international standard system through institutionalized rules. Three parallel paths make the standard itself a part of the competition.

It is worth noting that this competition is not simply about confrontation. In reality, China, the United States, and Europe are still highly interdependent in terms of technology, supply chain, and market: the cutting-edge technology of the United States cannot be separated from the global manufacturing system, China's industrial capabilities are deeply embedded in the international network, and Europe's regulatory influence also requires the cooperation of technological ecology. As a result, the international landscape in 2025 is closer to "limited cooperation in high-intensity competition". The real contest is no longer about the victory or defeat of a single model or a generation of technology, but about who can build and operate a whole set of technology and industrial systems for a long time.

10? Young Marshal in charge of troops

Another important trend in the AI industry in 2025 is that a group of young scientists in their early thirties or even twenties are being given command by large companies, directly affecting the direction of engineering architecture, data strategy, and next-generation AI capabilities.

When Tencent adjusted its AI organizational structure, it appointed 27 year old Yao Shunyu as the Chief AI Scientist and was responsible for the infrastructure and big model teams; Earlier, Xiaomi hired "post-95" scientist Luo Fuli to be responsible for the core research and development of the MiMo series. In Silicon Valley, this trend is even more pronounced. Meta has introduced Alexander Wang, the founder of Scale AI, as its Chief AI Officer, and remains steadfast in his support despite conflicting opinions with senior members within the company, demonstrating a high level of trust in young technology leaders.

Behind this "youthfulness+empowerment" is a change in the logic of AI development itself: AI has entered the "second half". The technological boundary is no longer just about making models bigger and longer, but about redefining problems, reshaping evaluation methods, and determining future paths. More and more companies are realizing that the core of AI has shifted from "solving problems" to "asking what questions to ask and how to measure progress," and this ability often appears in young tech enthusiasts who have been immersed in frontline research for a long time.

Therefore, 'Young Marshal in charge of troops' is not a temporary personnel slogan, but a structural adjustment after the industry enters the deep water zone: when the technological paradigm becomes more uncertain and exploratory, organizations need more people who are willing to quickly try and make mistakes and dare to judge under incomplete information, rather than relying solely on experienced engineering managers. As a result, young scientists such as Luo Fuli, Yao Shunyu, and Alexander Wang not only stand at the forefront of research and development, but also begin to directly influence the company's strategy and technological path. It can be expected that for a considerable period of time in the future, the key decisions in the direction of AI are likely to be in the hands of this generation of young technology leaders.