Hello! I'm your financial journalist and friend, also an economist. Today, we're not talking about cold, impersonal code, but about a "new species" that's quietly changing the way we live in the future—the World Model.
Lately, the term "World Model" has become incredibly popular in the tech community. Giants like Li Feifei, ByteDance, Tencent, Alibaba, and Ant Group have all jumped into the game. But what might the average person be asking? What exactly is this, and what does it have to do with me? Can I use it right now?
To figure this out, I delved into a detailed review article from "Xi Xiaoyao Technology Talk." This article doesn't overwhelm you with jargon; instead, it tests five of the mainstream World Model products in China as if you were playing a video game yourself.
Now, I'll break down this in simple terms into five key aspects to help you understand where the technology of "generating worlds" has currently arrived.
---
1. Conceptual Clarification: World Models Are Not Videos, They're Interactive Parallel Universes
First, let's clear up a big misconception: World Models ≠ Video Generation Models.
- Video Models (like Sora): You're just a spectator. You can't interact; you see whatever the director has filmed. It's linear and fixed.
- World Models: They're more like playing games like "The Legend of Zelda" or "Minecraft." You can interact in real time:
- You can walk forward, backward, or turn around.
- You can give commands to the world: "Make it rain," "Open the door," "Turn that cat into a superhero."
- Core Logic: The next scene is calculated and generated in real time based on your actions and commands.
Simple analogy:
If a video model shows you a tape of an underwater world, a World Model throws you right into that world. If you want to swim left, the fish scatter; if you throw a stone, water splashes. It's continuously running and provides immediate feedback.
Currently, five companies in China have made their products available for testing:
1. Ant Group LingBot-World 2.0
2. Loopit Zing-0.5
3. Alibaba HappyOyster 1.0
4. Aishi R1
5. Tencent Hunyuan HY-World 2.1
These products fall into three categories:
- Roaming Type: The world changes as you move (Tencent, some Alibaba products).
- Director Type: The world shows you what happens based on your commands (Aishi, some Alibaba products).
- Joint Control Type (most similar to games): You can change the world while you move (Ant Group, Loopit).
---
2. Basic Capability Comparison: Can It Remember Who You Are?
The most frustrating thing in games is when the character model doesn't match what you expect (for example, you suddenly become a random passerby or the house behind you disappears). The article reveals the real levels of "consistency" among these products through two key tests:
1. Input Accuracy (Does the Start Look Right?)
- Test: Given a picture of a desert caravan, which product can accurately recreate the start of the scene?
- Results: Zing, HappyOyster, LingBot performed the best, closely replicating the original image.
- Failures:
- Aishi R1: You provided a backview, but it showed a close-up of the face, completely changing the feel of the scene.
- Tencent HY-World: It refused to accept images of people, only recognizing landscapes, indicating it's more focused on scene generation rather than character interaction.
2. Consistency After Movement (Does the World Stay Stable?)
This is a critical test of the model's stability. If you move forward, the world should not change, disappear, or randomly add objects.
- Consistency of Characters/Objects: Zing & HappyOyster: Very stable; the car and the person remain the same.
- LingBot: Had a small issue; four large black tires appeared under the car, and it couldn't revert to the original state.
- Aishi R1: Disastrous performance: The character's face changed three times while driving in the desert—from a backview to a frontal view, then to a face with sunglasses, and finally to a female face. This shows it lacks long-term memory of the character.
- Consistency of the Environment: HappyOyster: The champion; the stars stayed in place whether you looked up or down at the village.
- Zing: Good, but the stars slightly distorted.
- LingBot: Failed; a second round revealed an extra moon in the sky, turning the Van Gogh-style stars into blue stripes.
💡 Journalist's Note: Alibaba HappyOyster performs best at maintaining a stable world, making it ideal for immersive experiences; Loopit Zing is close behind. Ant LingBot and Aishi R1 struggle with long-term interaction and need significant improvement.
---
3. Physics and Logic: Does It Understand Common Sense?
The world model needs to make sense logically. For example, a tsunami shouldn't flow down from the top of a ship.
The article tested two physical scenarios:
1. Tsunami Attack
- Command: "A tsunami is coming."
- Zing: The waves rose from the sea and hit the pier, following physical laws.
- LingBot: The water flowed down from the ship's top like a waterfall, while the sea around remained calm. This would be unrealistic in real life.
2. Dawn Process
- Command: "Dawn is breaking."
- Zing: The sky and sea gradually lit up, with a natural transition of light and shadow, like a real sunrise.
- LingBot: The changes were subtle and seemed fake.
💡 Journalist's Note: Zing leads in terms of physical logic. Its world feels like a well-functioning system, not just a collection of visual effects. This is crucial for future games and simulations; if the physics aren't correct, users won't believe it.
---
4. Core Gameplay: Is It For Watching or Playing?
This distinguishes advanced videos from true World Models. The article categorizes capabilities into Director Mode and Joint Control Mode:
1. Director Mode (AI Generates on Its Own)
- Aishi R1: Like a wildly creative writer: The visuals are stunning, but the story often gets confusing (for example, a cat walking suddenly develops butterfly wings or iron chains appear out of nowhere).
- Alibaba HappyOyster: Like a professional director: It uses camera techniques well, knowing when to zoom in and out. For example, it shows the cat's paws hitting the ground, then the flowers blooming, and then pulls back to a panoramic view. The narrative is clear, providing the best immersive experience.
2. Joint Control Mode (You Control and Change the World)
This is the ability closest to future games:
- Scenario: You control the character and change the environment simultaneously.
- Zing: Performances are excellent; it keeps track of your position and the dragon's, making it clear what to do next.
- LingBot: The effects are too distracting; blue halos and purple flames obscure your view, making it hard to focus on the action.
💡 Journalist's Note:
- Choose HappyOyster if you want to watch a story.
- Choose Zing (Loopit) if you want to play and interact.
- LingBot (Ant Group) has stunning effects but can be overwhelming and interfere with your gameplay.
---
5. Commercialization: Is It Fast and Affordable?
No matter how good the technology, if it's laggy or too expensive, it won't be accessible to most people.
1. Response Speed (Is It Responsive?**
- Zing: Responds quickly to environmental changes (like thunderclouds); the world changes immediately.
- LingBot: Responds quickly to character actions.
- Aishi R1: Slower; you need to wait a few seconds for the scene to update.
- HappyOyster: Responsive for roaming, but you can't issue multiple commands at once.
Conclusion: Zing offers the best immediate feedback, closest to the smoothness of real games.
2. Cost (Is It Affordable?**
This is a practical economic consideration. The article compares the cost for 10 minutes of use:
| Product | Cost for 10 Minutes | Notes |
| :--- | :--- | :--- |
| Loopit Zing: < 1 RMB | This is the inference cost (server fees + computing power), not the user's bill. If priced accordingly, it would be very competitive. |
| Ant Group LingBot: ~$2 (~15 RMB)** | Publicly quoted price for 10 minutes of use. |
| Alibaba HappyOyster: ~$10.4 (~75 RMB)** | Includes the cost of creating the world and real-time session fees. The most expensive. |
| Tencent/Aishi: Data missing or pricing unclear | Tencent charges based on the outcome; Aishi's pricing is unknown. |
💡 Journalist's Note:
- Zing has a significant cost advantage. If it can reduce the inference cost to below 1 RMB and price it accordingly, it could become widely popular.
- HappyOyster, due to its high cost of creating the world, is more like a high-end custom service for professionals, not for general users.
---
📝 Summary: Where Has the World Model Reached?
From an economic perspective, the World Model is currently at the GPT-3 stage:
- It's not like ChatGPT: It's not yet fully mature and ready for everyone to use out of the box.
- It's like GPT-3: The technical path is established, and its capabilities are proven. You can feel that this direction is correct, but the experience still needs refinement, and there are many bugs.
Advice for the Average Person:
1. Don't expect to play complex games right now: The products are still in the research and preview phase, with many bugs (character model issues, inconsistent visuals).
2. Focus on Joint Control: This is the key difference between World Models and video models. The product that can change the world smoothly and affordably will be the winner.
3. Watch Out for Alibaba and Loopit:
- Alibaba HappyOyster excels in storytelling and stability, suitable for films and interactive dramas.
- Loopit Zing stands out for its interactivity and low cost, ideal for game prototyping and real-time interactions.
In conclusion, the World Model is no longer just science fiction; it's becoming a new creation tool.* Although it's still somewhat random, it has already shown us the potential of real-time world generation.* Next, we'll focus on whether it can keep us engaged and interested enough to keep playing in that world.