Everyone (including me) was recently convinced that world models are hard to use for games, but now we've been proven wrong—
LinearGame, a company based in Silicon Valley and Singapore, has used a world model to create a game development platform called Yoroll.
They claim that Yoroll proposes a new paradigm of "engine-free games."
After looking into it, let's just say that even if Yoroll doesn't ultimately succeed, the solution approach it proposes is inspiring enough for many people.
Successful AI startups today are all like Yoroll: when a new technology emerges, they think about what new demands it can create.
The key is to see who can find the leverage to unlock new demands.
Simply put, currently, the only public information about Yoroll comes from a third-party report on Business Wire and an introductory video on the LinearGame website.
They each provide two pieces of information: 1. How Yoroll is built; 2. How to use Yoroll.
Together, they form a complete picture of the project: how LinearGame uses world models to create new demand.
First, the information from Business Wire, which likely comes from the team's sharing, is about Yoroll's product concept.
Surprisingly simple.
The big background is that the consensus in the gaming industry about world models is that they essentially rely on memory, like dreaming, and definitely cannot be used for projects.
Because they are unstable, not persistent, and the opposite of the "controllability" the gaming industry seeks.
But Yoroll's understanding is that the first principle of games is not stable persistence, but feedback.
This can be seen from the product structure. According to reports, Yoroll contains a three-layer technical structure: a Presentation Layer, a Judgment Layer, and a State Layer.
The most fundamental is the Presentation Layer, which is also where the world model resides.
The world model, like what Genie and others demonstrate, generates high-quality visuals + physics & causality rules + some basic interactions.
Yoroll hasn't changed the quality of the world model; the only difference is that they increased the types of interaction, such as movement, and also clicks, swipes, voice, etc.
That's it, just that much.
On top of the underlying technology, Yoroll adds two engineering systems—this is also the key to why a world model can make games.
The first layer is the Judgment Layer, a vision-language large model. It is responsible for monitoring the visuals generated by the world model and judging what is happening in the scene.
The world model only extrapolates what will happen and presents it visually, but the visual itself has no behavioral data.
The Judgment Layer's job is to summarize key behaviors from the visuals into quantifiable data.
After an action is performed, does it increase favorability, or decrease health? These are definitive pieces of data.
At the second layer, the State Layer, the data goes into traditional data storage, corresponding to values like health points, status effects, etc.
Then, the State Layer feeds this data back, using it as prompts to give to the world model, influencing its subsequent changes.
Thus, a technical foundation plus two engineering insights form a simple but effective loop:
"Generate - Collect - Store + Feedback" -> influences the next generation cycle.
Why can Yoroll step out of the box of "world models can't make games"? Because Yoroll and traditional game systems are completely different logics.
A conventional game is like a precision instrument, with gears meshing one after another: every interaction feedback depends on collision volumes, parameter judgments...
Yoroll is not like that. It is: see what happens, do something, then tell the world model what changes occur next.
The engineering is brute-force simple, but it's enough to pull world models into the realm of games.
However, there's another important factor for why Yoroll can use world models to make games: Interactive Film Games (Interactive Film).
Interactive film games emphasize visuals over feedback and require constant scene cuts.
This genre and world models are a match made in heaven.
Interactive film games precisely do not require world stability, perfectly avoiding the disadvantages of lacking 3D assets and short spatial persistence.
Meanwhile, the world model's advantage of producing AAA-quality visuals is exactly the long-board this genre needs.
In other words, this new technology (world models) emerging, interactive film games are the architecture that allows it to generate new demand.
What kind of demand? This leads to the second source: the video on LinearGame's official website, showing how the market should use the Yoroll platform.
Like the product structure, Yoroll's user-facing approach also follows first principles: for interactive film games, the key is the idea.
Yoroll's creation process is extremely simple, comparable to posting a daily video on TikTok, with only 3 steps—
1. Set the direction, write a prompt -> generate with one click;
2. Finalize assets, review story and characters -> generate with one click;
3. Determine details, adjust camera and interaction, check content衔接-> pack with one click.
If there isn't enough content and you want to add more, it's still a one-click generation.
All the complex filming and editing is omitted. What's left for the user to do is simply "propose ideas."
If you're unsatisfied and want to change something, you only need to "input a Prompt" and "click settings."
It's "boss-style development": don't worry about the implementation details below; just see clear results above and propose ideas with good taste.
Therefore, the project engineering Yoroll presents is extremely simple, with only two core components: Story & Cast, and Storyboard.
Story & Cast: The story is just a synopsis and main characters, no long-form text, making it easy to shorten or lengthen content at will.
Characters will have three-view design sketches with clothing details. If unsatisfied, directly modify via prompt.
Moving to Storyboard, the structure remains simple: just Script and Canvas.
Script manages every individual shot, showing the basic visuals, events, and dialogue for each short, as well as GamePlay and transition methods.
Canvas manages the overall narrative structure, the main storyline route, and branch conditions. Adjustments can be made with simple drag-and-drop.
Finally, Publish: The user just adds a title and cover, and it's packed into a game that can be played directly by clicking.
The packaged game can be run on a PC or played on a mobile phone.
The entire process is so simple there's little to say—it's easier than shooting a video or publishing an article.
The "light production, light publishing, light experience" created by Yoroll, if successful, will give birth to a much lighter new form than the interactive film games it spawned from.
This is also what was mentioned at the beginning: new demands, even a new ecosystem.
Yoroll is currently invite-only and still has a way to go before a full release, but the思路 is already clear enough.
The judgment that "world models can't make games" is based on the viewpoint of how the existing industry can use AI for efficiency gains.
When people see a new technology, the first reaction is always, "Can it help me do my current job better?"
This is a typical "replacement" mindset.
The industry saying the market has misjudged world models isn't wrong; that judgment was also about replacing 3A games.
But when a new technology appears, especially one not born from the industry, its existence isn't for replacement, but to spawn new things.
Therefore, traditional thinking can actually constrain people's imagination of it.
Perhaps Yoroll isn't the optimal solution for world models, nor is it guaranteed to truly make games.
But Yoroll has already, through its own thinking, using the small lever of the "Judgment Layer & State Layer," become one of the first to pry world models toward games.