World models, the next frontier in artificial intelligence, are attracting billions in funding from tech giants and startups worldwide, but the industry lacks a shared definition of what they actually are. On April 16, Chinese tech giants Alibaba and Tencent released their own world models, Happy Oyster and HY-World 2.0, while earlier in the year, World Labs and AMI Labs each secured billion-dollar funding rounds . Yet despite the hype and capital flowing in, asking different players what a world model actually does yields conflicting answers: some describe it as an interactive 3D world, others as a causal model understanding physical laws, and still others as advanced video generation .

Why Did AI Companies Suddenly Bet Everything on World Models?

The answer lies in a fundamental weakness of large language models (LLMs), the AI systems behind ChatGPT and similar tools. Over the past two years, LLMs have demonstrated remarkable language abilities, but they do not truly understand the physical world. Ask an LLM what happens when you push a cup off a table, and it can recite the correct answer from memory, but it doesn't genuinely grasp gravity, acceleration, or collision dynamics .

This limitation becomes catastrophic when AI moves beyond text. Autonomous vehicles cannot afford to "approximately correctly" identify obstacles, and industrial robots cannot "roughly" predict part trajectories. A study in early 2026 found that hallucinations in LLMs are not a data or training issue, but an inherent flaw in their architecture . The industry needs AI that understands causality in the physical world, not just language patterns.

The commercial stakes are enormous. LLMs have dominated AI commercialization for the past two years, but their applications remain confined to information processing: writing copy, translating text, and generating code. The next growth engine lies in the physical world, including embodied intelligence, autonomous driving, and intelligent manufacturing . Whoever enables AI to truly understand and predict physical dynamics will dominate the next industrial cycle.

What Are the Three Competing Approaches to Building World Models?

The world model race has split into three fundamentally different camps, each with its own philosophy and trade-offs. Understanding these approaches reveals why the industry is so fragmented and why comparing them directly is nearly impossible.

The Abstract Reasoning Approach: Yann LeCun's AMI Labs has taken the most counterintuitive path, deliberately rejecting pixel-level realism. LeCun's JEPA architecture makes predictions only in abstract latent space, discarding visual details that humans find intuitive. The newly released LeWorldModel contains just 15 million parameters and can be trained in only a few hours on a single GPU, yet achieves planning speeds 48 times faster than traditional methods . The trade-off is stark: the output is incomprehensible to humans. You cannot "see" the future it predicts; you can only trust its calculations. This is a purely academic approach betting that true intelligence doesn't require simulating every leaf falling in the wind, only understanding the causality that "wind blows leaves down" .

The 3D Reconstruction Approach: Li Fei-Fei's World Labs believes intelligence must be built on explicit understanding of three-dimensional space. Her Marble model can generate an editable, navigable 3D world from a single photograph or text description, allowing users to freely move their viewpoint within it . World Labs has also open-sourced the rendering engine Spark 2.0, enabling standard web browsers to smoothly load hundreds of millions of 3D points . However, Marble excels at reconstructing what space looks like while showing relatively weak understanding of what happens within that space. You can walk into a generated room, but you cannot push the chairs or knock over cups on the table. It is a replicator of static worlds, not a simulator of dynamic physics .

The Generative Video Approach: The most commercially active camp includes Google's Genie 3, Alibaba's Happy Oyster, and Tencent's HY-World 2.0. Their logic is straightforward: as long as generated images look convincing, the model has learned something meaningful about how the world works . This approach prioritizes immediate commercial viability over theoretical purity.

How to Understand the Differences Between World Model Approaches

• Technical Philosophy: LeCun's approach abandons visual realism entirely, Li Fei-Fei's focuses on spatial accuracy without dynamic understanding, and the generative camp prioritizes visual plausibility as proof of physical understanding.
• Computational Efficiency: AMI Labs' LeWorldModel trains in hours on a single GPU, while 3D reconstruction and video generation typically require far more computational resources and data.
• Commercial Timeline: Chinese companies like Alibaba and Tencent are integrating models into paying products immediately, while U.S. research labs like AMI acknowledge that commercial products may not appear for several years.
• Output Usability: Abstract models produce incomprehensible predictions, 3D models generate navigable spaces, and generative models produce video that humans can evaluate visually.

The divergence between U.S. and Chinese approaches reflects different risk tolerances. U.S. teams like DeepMind, World Labs, and AMI Labs focus on fundamental science, betting on breakthroughs a decade away. China presents a different picture: Alibaba and Tencent immediately integrated their models into commercial applications. Happy Oyster targets paying users in film production and game development, while HY-World 2.0 directly outputs 3D assets compatible with industry-standard software like Unity and Unreal Engine . Sand.ai's VidMuse, focused on generating videos from music, achieved annual revenue in the millions of dollars within just a few months of launch . The Chinese team's logic is pragmatic: a world model must first be a profitable product .

What Critical Challenges Remain Unsolved?

Despite the hype and funding, the industry faces severe unresolved problems. The absence of technical standards has led to misaligned evaluations across different approaches. Core technical challenges include data scarcity, long-term temporal inconsistency, and insufficient physical accuracy . A model might generate convincing video for five seconds but fail catastrophically at ten-second predictions. Another might understand gravity but struggle with friction or collision dynamics.

More critically, accountability frameworks, ethical guidelines, and legal regulations are severely lagging behind the technology. Real-world risks are increasingly plausible, including autonomous driving misjudgments, industrial operational errors, and the spread of synthetic content . If a world model trained on biased data teaches an autonomous vehicle to make unsafe decisions, who bears responsibility? If a generative model produces convincing but false simulations of industrial processes, how do regulators catch it? These questions remain largely unanswered as the industry races forward.

The world model gold rush reflects genuine technological progress and legitimate commercial opportunity. But beneath the optimistic headlines lies a fragmented industry speaking different languages, pursuing incompatible goals, and racing ahead of the safety and regulatory frameworks needed to deploy these systems responsibly in the physical world.

" }