Introduction: The Shifting Landscape of Artificial Intelligence

The artificial intelligence landscape is undergoing a profound transformation. Recent breakthroughs, such as Google's release of Genie 3, a world model capable of real-time interactive 3D environment generation, and OpenAI's launch of GPT-5, have once again ignited industry discussions. These advancements underscore a critical shift: the future of AI is increasingly intertwined with the Internet of Things (IoT).

For years, the notion that 70% of the industrial value of "AI+" would ultimately reside within the IoT was considered bold, even radical. However, as the industrialization of AI accelerates, this prediction is proving to be increasingly accurate. Far from being marginalized, IoT is emerging as a core driver for the practical implementation of AI, empowering diverse industries. By 2025, global IoT terminal connections are projected to exceed 27 billion. Crucially, these vast numbers of IoT devices, deployed across manufacturing, transportation, healthcare, and urban environments, provide a staggering 67%-72% of the raw data essential for AI applications. The IoT has become the most robust and expansive data foundation for AI's evolution and application.

This trend is further validated by the latest advancements in foundational AI models. New-generation AI systems, exemplified by GPT-5 and Genie 3, are moving beyond a sole reliance on virtual data like internet text and images. They are progressively shifting towards actively perceiving, understanding, and even manipulating the physical world. Behind these technological updates, the value of IoT is becoming increasingly prominent. It serves not only as a data collector but also as an indispensable bridge for AI to interact with the real world, receive feedback, and engage in continuous learning.

Whether it's more powerful world models or autonomous intelligent agents, they all depend on the massive amounts of real-time, multimodal, and embodied data generated by IoT terminals. This data is not only vast in quantity but also rich in physical attributes, scene characteristics, and behavioral semantics, becoming key to helping AI models overcome hallucinations and move towards true intelligence.

Virtual Intelligence Limits vs. The Dawn of Physical Intelligence

In recent years, the concept of "Scaling Law" has been a guiding principle for the rapid advancement of artificial intelligence. Since GPT-3, the development of large language models has largely followed a "brute force aesthetic" logic: the larger the parameters, the more data, and the stronger the computing power, the closer intelligence gets to being general-purpose. GPT-4, GPT-4o, and the recently released GPT-5 have each pushed the boundaries of scale and capability. From text generation to multimodal understanding, these models have indeed brought about astonishing leaps in ability. However, behind these larger and more powerful models, their limits and bottlenecks are also becoming increasingly apparent.

As the dividends from data diminish and computing costs grow exponentially, the improvement in model accuracy and generalization capabilities has slowed, even showing diminishing marginal returns. OpenAI's highly anticipated new-generation model, GPT-5, faced an unexpected initial reaction upon its release, with some early users complaining that its performance was "clumsy" and even inferior to previous versions. OpenAI CEO Sam Altman quickly responded, stating that Plus users would be allowed to continue using the previous version of GPT-4o.

More alarmingly, the phenomenon of hallucinations in large models within the virtual world remains difficult to curb. Many facts indicate that AI still "talks the talk but doesn't walk the walk." They excel at filling in blanks or imitating within existing data distributions but struggle to break out of the virtual world's sandbox to truly understand and respond to complex and ever-changing real-world scenarios. It has been proven that by simply stacking data and computing power, AI finds it difficult to overcome the ceiling of virtual intelligence. This makes the so-called "AI + IoT" no longer a mere embellishment but a cornerstone of the intelligent agent era. AIoT not only connects everything but also imbues everything with intelligence, becoming the essential path for AI to break through its boundaries.

It is against this backdrop that data from the physical world is becoming the new goldmine for AI evolution. As the value of text and image data approaches its limit, real-world data collected by IoT terminals is becoming the "fountain of life" that drives AI capabilities forward. As the video above demonstrates, the introduction of Genie 3 has allowed world models to achieve real-time interaction in 3D physical environments for the first time. The research and implementation of embodied intelligent agents also emphasize AI's ability to actively perceive, operate, and receive feedback from the physical world. The essence of these latest cases is a paradigm shift in AI capabilities from the virtual to the physical.

Only perception, interaction, and feedback data from the physical world can provide AI with true generalization and causal reasoning capabilities. This type of data is not only large in quantity and high in quality but also contains rich scene diversity and dynamic changes, which are crucial for intelligent agents to adapt to complex environments. Although collecting, labeling, and generalizing physical world data face significant technical and cost challenges, the "scene generalization" value it brings far exceeds the data accumulation in the virtual world. The path of AI evolution can no longer avoid a deep embrace of the physical world.

World Models × AIoT: The Rise of New Species of Intelligent Agents

In the progression of AI development, "big data" was once considered the panacea for intelligent evolution. Countless models, built upon massive accumulations of text, images, and audio data, achieved unprecedented expressive and understanding capabilities. However, as AI capabilities approach the limits of the virtual world, this "quantity over quality" paradigm is gradually becoming ineffective. In its place, there is an extreme desire and competition for "good data." In the future, what will truly drive AI implementation and evolution will no longer be the absolute scale of data, but rather the quality and structure of "good data."

In the physical world, "good data" has become the core bottleneck for AI perception, understanding, and decision-making. What constitutes "good data"? Firstly, it must possess physical authenticity, meaning the data originates from real environments, real operations, and real feedback, accurately reflecting the laws and dynamics of the physical world. Secondly, it must have semantic interpretability; it's not just low-level sensor signals, but data with clear labels, structures, and semantic information that facilitates higher-level cognitive processing by models. More importantly, it needs scene generalizability, meaning the data can cover diverse scenarios, complex environmental changes, and edge cases, ensuring the model possesses transfer and generalization capabilities.

In the era of intelligent agents, "good data" is the true fuel for AI evolution and the foundation for all technological breakthroughs. This is because the awakening of intelligent agents requires embodied intelligence and world models as leverage, relying on the AIoT intelligent agent network to achieve collaborative evolution. Many mistakenly believe that embodied intelligence is synonymous with humanoid robots. In reality, the essence of embodied intelligence is to empower AI with the ability to actively perceive, physically interact, and self-learn. AIoT intelligent agents are the best carriers of this capability. Whether in factory automation, smart cities, unmanned delivery, or smart homes, AIoT intelligent agents are quietly permeating every corner of the physical world in a distributed and networked form.

The evolution of world models is enabling AI to transition from "knowing how to say" to "knowing how to do," evolving from pixel/text processing capabilities to physical causality and abstract reasoning. Taking the new generation of world models advocated by computer scientist Yann LeCun as an example, AI no longer passively reconstructs data but actively predicts environmental evolution, deduces the consequences of its own actions, and achieves counterfactual reasoning and zero-shot planning. The essence of this capability is a deep understanding and generalized application of the laws of the physical world. All of this is made possible by the active perception, distributed decision-making, and real-time feedback supported by the AIoT intelligent agent network.

Every embodied intelligent agent acts as an "eye" and "hand" in the physical world, forming a collaborative, shared, and evolving super-intelligent agent ecosystem through the IoT network. Ultimately, the generalization and adaptability of intelligent agents must rely on the closed loop of the physical world provided by AIoT. World models are the foundation of cognition, and AIoT is the sinews and bones of action; their synergy will lead to the awakening of intelligent agents in the physical world.

From the Hundred Model War to the Intelligent Agent Economy

As AI technology rapidly evolves, the industrial landscape is reaching an unprecedented turning point. Over the past two years, AI has rapidly expanded amidst the "Hundred Model War," with countless large models, applications, and platforms vying for leadership in algorithms and scale. However, the windows of technological and traffic dividends are closing. The true competitive focus is shifting from a battle of model capabilities to the control of platformization, hardware-software integration, and data closed-loops. Large models are now infrastructure; only those who can achieve "intelligent agents as ecosystems" in broader industrial scenarios are likely to dominate the new round of intelligent revolution.

This shift in AI focus signifies a deep evolution of AI business models from "model-as-a-service" to "intelligent agent as an ecosystem." In complex physical world scenarios such as factories, logistics, cities, and healthcare, a single AI model API can no longer meet the end-to-end demands from perception and decision-making to execution. Enterprises and urban clients are increasingly eager for integrated hardware-software platforms that enable end-to-end data closed-loops and continuous evolution. For example, in automated factories, only by connecting the entire chain of equipment, sensing, AI decision-making, and robot execution can a self-learning, self-optimizing, and self-managing intelligent production system be formed. The logistics industry's demand for active collaboration and dynamic scheduling of intelligent agents also determines the irreplaceable nature of platform-level AI capabilities.

In this process, it is worth noting that the mission of AIoT is being redefined. It is no longer merely a networking tool or a data collection hub, but rather a means to evolve every physical device into an active intelligent agent capable of perceiving, deciding, and acting, continuously generating high-value data. The value of AIoT is elevating from being the foundation of digital transformation to becoming the new infrastructure of the intelligent agent era. In cutting-edge fields such as smart factories, smart cities, and digital healthcare, AIoT has become a super connector for the deep integration of AI and the real economy. The future physical intelligent economy will essentially be a global collaboration, data-driven, and intelligent emergence driven by AIoT.

This trend also drives changes in the industrial ecosystem. AIoT platforms, embodied intelligent models, and Agent ecosystems are forming a three-in-one resonant development. AIoT platforms provide a unified foundation for perception, communication, and execution; embodied models empower each intelligent agent with autonomous learning and reasoning capabilities; and various intelligent Agents continuously evolve and collaborate in specific scenarios, forming a self-organizing, self-adaptive intelligent agent network.

Conclusion: Embracing the Physical World for the Next Decade of AI

Looking back at the evolution of the AI industry, we stand at an unprecedented historical turning point. The craze for large models will eventually return to rationality, and the true value of AI is accelerating its migration to the physical world. The prediction that 70% of the value of "AI+" will come from the Internet of Things is not only being validated by an increasing number of real-world cases but is also becoming a strategic consensus to be firmly believed in for the next decade. As the AIoT infrastructure awakens and matures, the future of intelligent agents is being defined and dominated by the IoT.

For all industry decision-makers, developers, and academic researchers, now is the best time to embrace the integration of "AI + Physical World." Whether it is to promote the intelligent upgrade of the real economy or to build new infrastructure for the future, AIoT has become an indispensable key cornerstone. Looking ahead, only by deeply embracing the physical world can intelligent agents truly awaken. When AI is no longer confined to virtual space but deeply integrates with ubiquitous perception, ubiquitous connectivity, and ubiquitous intelligence, the entire society and industry may usher in the next golden decade led by intelligent agents. The next industrial miracle will be ignited by the sparks of AIoT.