Zhidx (WeChat public account: zhidxcom)
Author | ZeR0
Editor | Moying
Zhidx reported on November 5 that the second International Engineering Intelligence Conference of the 2026 World Laureates Forum was held recently. At the conference, Tao Mei, a foreign member of the Canadian Academy of Engineering and founder and CEO of Zhixiang Yuanyi (HiDream.ai), delivered a speech proposing the construction of agentic omnimodal world models (Agentic Omnimodal World Models), advancing toward next-generation world intelligence.
Data from AI model aggregation platform OpenRouter shows that global large-model usage is entering a phase of explosive growth. As of October this year, China’s weekly token-call volume reached 93 trillion, 4.16 times that of the United States, and has rankedfirst globally for 26 consecutive weeks.
Behind these figures is AI’s rapid evolution from a “tool” into a new type of social infrastructure, like water and electricity.
Meanwhile, model iteration continues to accelerate. In 2023, large models were essentially iterated on a 12-month cycle. As recursive self-improvement (RSI) has pushed AI onto a self-accelerating trajectory, model iteration has moved to a weekly cadence.
The more capable AI becomes, the more impossible its shortcomings are to ignore.
In the field of digital intelligence (Digital AI), from text interaction and code writing to video generation and even mathematical research, AI has reached and even surpassed the level of human experts. For example, in mathematical research, OpenAI recently made public 377 mathematical results at once, covering hundreds of long-standing unresolved open problems.
The development of physical intelligence (Physical AI), by contrast, has been relatively slow, constrained by weak physical understanding, a lack of interaction data, and inadequate closed-loop execution capabilities.
How to bridge the chasm separating Digital AI from Physical AI is precisely the defining question of the era that next-generation AI must answer.
World models are precisely the bridge across this chasm.
At the core of a world model is enabling AI to understand how the world operates and reason about events that may occur in the future. Through unified understanding and generation of information across all modalities, it models the state of the world, physical laws, and causal relationships, thereby enabling perception, inference, and prediction in the real world.
Different technological paths lead to world models, making this the next competitive frontier in AI development. The industry has now formed three evolutionary paths: large language models, multimodal models, and embodied large models. Tao Mei believes thatthe three technological paths are not isolated from one another and will eventually converge in the long run, moving toward agentic omnimodal world models (Agentic Omnimodal World Models).
An agentic omnimodal world model has three main components:
First, the LLM, which serves as the agent’s decision-making brain, responsible for understanding intent, generating reasoning, and devising task plans.
Second, the Omni-Model, or general-purpose world model. It is the world simulator: based on the LLM’s intent, it simulates the operation of the physical world, reasons through the consequences of actions, and predicts the next action.
Third, the Embodied Context, or embodied context. It interfaces with the real physical environment, executes embodied actions, and feeds real observation data back to calibrate the world model.
Through the synergy of the three, “One Token Space, One World” is realized—the world of thought and the physical world are unified within the same token space. This is precisely the key to the next generation of world intelligence.
Reflections on world models have been internalized into Zhidx’s product philosophy and technological DNA. Zhidx independently developed a native omni-modal (Unified Transformer) model architecture,enabling unified representations across all modalities, including text, images, video, space, pose, force and touch, sensors, and more.
Building on the UiT foundation, Zhidx has created a comprehensive world model matrix: starting with image models and extending to video models, the interactive model HiDream-O1-World, and the embodied model HiDream-O1-Embodied, enabling capabilities to evolve from static images and dynamic video to spatial interaction and embodied action.
Zhidx has established a “1+1+3” commercialization architecture: the foundation is a native omni-modal world model matrix; the middle layer is the enterprise-grade model platform HiHarness, which brings together Zhidx’s proprietary HiDream-O1 omni-modal model family and the world’s leading models; and the top layer consists of three agents: HiBurst, an AIGC-powered commercial marketing agent; “帧赞,” an audiovisual content creation agent; and vivago, a content creation agent.
At the same time, Zhidx is actively advancing the application of world models in frontier fields. In embodied intelligence, world models can synthesize high-precision, millimeter-level action videos at scale, overcoming the bottlenecks of high cost and limited scale in real-robot data and providing data for real-robot training. In the life sciences, Zhidx and its partners are using unified representations of multiscale biological images to understand and predict cellular evolution, providing an upfront reference for drug-response evaluation and screening out candidates more worthy of validation before experiments.
From generating images to building worlds, an agentic world model is not a linear upgrade in capabilities, but an entirely new direction in AI evolution. When the laws of physics become the underlying language of AI, the real world will truly open itself to it.
jQuery(document).ready(function($) { $.post('https://www.zhidx.com/wp-admin/admin-ajax.php', {action: 'wpt_view_count', id: '607782'}); });No text provided.
