DIAMOND, short for Diffusion for World Modelling, shows how far AI game simulation has moved and how much remains unsolved. Researchers used it to simulate a rough version of Counter-Strike: Global Offensive (CS:GO) inside a neural network, running at 10 frames per second on a single Nvidia RTX 3090 graphics card.
The result is not a polished replacement for the game. It is a working demonstration of a world model that can respond to player input, imitate parts of a familiar shooter, and expose the limits of learning a game world from recorded play.
What DIAMOND actually does
DIAMOND is an AI model designed to simulate game environments rather than simply analyze them. In the CS:GO demonstration, players interact with the generated environment using a keyboard and mouse, and the model produces what happens next inside the simulated world.
The source describes the result as a janky version of Counter-Strike: Global Offensive, but the technical point is still significant. The model is not only drawing still images. It is attempting to carry forward a playable environment with player interactions, weapon mechanics, and environmental physics.
That makes the demo more demanding than a passive video prediction task. A player can move, aim, and interact, and the model must respond in a way that remains close enough to the game for the illusion to hold.
DIAMOND had already been shown for Atari games. For CS:GO, the same general idea is being pushed into a more complex setting, where movement, combat, objects, and maps create many more ways for the model to lose track of reality.
Why the training data matters
The team trained DIAMOND on just 87 hours of CS:GO gameplay data. According to the source, that is only 0.5% of the data used for similar projects like GameNGen.
That small dataset is important for interpreting the demo. The model is not being presented as a finished game engine. It is evidence that a neural network can learn enough from a limited slice of gameplay to produce an interactive simulation that resembles CS:GO in recognizable ways.
The approach uses a Transformer-based method that treats player movements as "tokens," similar to words in a sentence. By predicting these tokens, the model learns to anticipate the next movement based on earlier actions.
In plain terms, DIAMOND learns patterns from gameplay and uses those patterns to generate what should happen next. If the player behaves in ways the model has seen often enough, the simulation has a better chance of staying coherent. If the player goes somewhere unusual or acts outside the patterns in the data, the system can break down.
The demo works, but the cracks are visible
Researcher Eloi Alonso demonstrated the model's capabilities on Twitter on October 11, 2024. The videos showed players controlling the simulated CS:GO environment with standard keyboard-and-mouse input.
The demonstration is notable because it includes complex elements that are hard for an AI simulation to keep consistent. Player interactions, weapon mechanics, and environmental physics all require the model to maintain a changing world state while responding to user input.
But the limitations are not minor. The model has serious bugs. One example is that players can jump an infinite number of times because DIAMOND does not take into account the Source Engine's gravity or collision detection.
Another failure mode appears when players leave the common paths represented in the training data. Departing from frequently used routes can lead to a complete collapse of the simulation. That weakness shows the difference between imitating likely gameplay and understanding all the rules that make the original environment work.
The researchers themselves acknowledge that the model has many limitations. They also expect the world model to improve by scaling up data and compute, especially because the dataset only amounts to 87h of gameplay.
How it compares with GameNGen
DIAMOND for CS:GO was inspired by GameNGen, an AI system developed by Google Research, Google DeepMind, and Tel Aviv University. GameNGen can fully simulate parts of the classic game DOOM at over 20 frames per second on a single Google TPU chip.
The comparison gives useful context. GameNGen shows that AI simulation can reach playable speeds in a classic game setting. DIAMOND applies a related ambition to Counter-Strike: Global Offensive, a more complex environment with different demands on interaction and consistency.
The CS:GO result is slower, running at 10 frames per second, and it is visibly imperfect. Still, it points toward the same broader research direction: AI systems that learn to generate interactive worlds from gameplay data.
The researchers see potential beyond games. The source says they believe scaling up data and computing power could help develop AI models that navigate complex real-world environments. That claim does not make DIAMOND a real-world simulator today, but it explains why a rough CS:GO demo is being treated as more than a novelty.
Why this AI Counter-Strike model matters
The value of DIAMOND is not that it perfectly reproduces Counter-Strike: Global Offensive. It does not. The value is that it shows an AI model producing an interactive approximation of a complex game world on a single Nvidia RTX 3090 graphics card.
Its failures are just as informative as its successes. Infinite jumping and simulation collapse reveal where learned patterns fall short of engine-level rules such as gravity and collision detection. Those problems matter because an interactive world must stay stable even when a user does something unexpected.
For now, DIAMOND is best understood as a research milestone. It runs, it responds, and it demonstrates how a world model can absorb gameplay structure from a small dataset. At the same time, its glitches make clear that scaling data and compute is only part of the challenge.
The DIAMOND model is available on GitHub for people who want to explore it further. As a research demo, it offers a direct look at both the promise and fragility of AI-generated game simulation.