Over more than a decade in the industry, I’ve witnessed iterations of PC games, mobile games, and open-world experiences. Recently, while studying materials related to world models, I suddenly made a new discovery: the games we play and build every day are not only entertainment products—they can also be a top-tier training ground for an AI physical world.
Eighty years ago, the Scottish psychologist Craik proposed that the brain completes perception, prediction, and action in a closed loop by constructing small internal models of reality. This long-forgotten idea is now being realized through game data and neural networks—turning it into world models.
In 2018, the (World Models) paper officially brought the concept of “AI dream training” to the spotlight:
V: visual compression + M: memory prediction + C: decision controller
This three-piece set enables AI to train strategies in the virtual fantasy track before deploying in real environments
Later, there were successive breakthroughs. The most震撼 one was that Google used a pure neural network to fully replace the traditional DOOM engine—no rendering pipeline, no handwritten collision code; everything is generated in real time by the model learning the rules of the game world
Why do we prefer games to train world models? I’d like to share my understanding from an industry practitioner’s perspective:
1. Naturally labeled data built in
Player operations every frame + automatic screen pairing; for example, https://t.co/Lt616IQbKd. In one year, 2 billion game screen-record segments, complete with full button and joystick action logs. No need to hire teams to label at high cost like in autonomous driving. DeepMind Genie can train interactive 3D worlds with 200,000 hours of game videos
2. A controllable, simplified physics classroom
Games have built-in gravity, collision, and fluid logic. With rules simplified and clear, they perfectly match the logic of curriculum learning: AI first masters virtual physics, then migrates to the real world and robot simulation
3. Massive, effectively infinite trial-and-error at zero cost
Hundreds of thousands of games and daily matches at the billion scale. In games, AI repeatedly tries without loss—that data scale is something physical robots and real-world road testing simulation can’t reach
4. Mature, general-purpose action-interaction system
For decades, the keyboard/mouse and controllers have iterated toward standardized intent-compression interfaces. With one joystick-and-button logic, you can control humanoid characters, but also drive cars and fly vehicles. This action space is a treasure for training embodied intelligence
In the past, we made games limited to “providing entertainment experiences.” Now, with the wave of world models, the value of games is being redefined:
Reforming game industrialization from within: eliminate rigid NPC behavior trees and handwritten physics scripts; use AI to autonomously understand scene generation and dynamic interactions, enabling infinite generation of original, playable virtual worlds
Crossing over externally to become foundational infrastructure: game data, virtual simulation—these become the core training base for robot, autonomous driving, and aviation defense AI
The most cognition-shattering point: human operations recorded by game controllers, in essence, are “intent-to-action mapping data” liberated from humanoid and species constraints. Once the model learns to achieve goals in virtual dynamic systems, this logic can be reused on any physical entity robot
Eighty years ago, Craik couldn’t witness his theory taking shape in the real world. Now, silicon-based AI is recreating the modeling logic of the biological brain—and the testing ground for this transformation is, above all, the game industry.
In the next decade, games will no longer be just an entertainment track, but the core entry point for general AI with physical perception. All game practitioners stand at the turning point of a paradigm shift
Huge opportunities for the 3D game industry in the AI era!!!