Why This CEO Thinks Video Games Make Better Training Data Than the Internet

Background and Context

The pursuit of Artificial General Intelligence (AGI) has entered a critical phase where the limitations of current Large Language Models (LLMs) are becoming increasingly apparent. While systems like ChatGPT and Claude have demonstrated remarkable proficiency in text generation, logical reasoning, and knowledge retrieval, they exhibit significant deficiencies in understanding the fundamental laws of the physical world. Specifically, these models struggle with spatial-temporal reasoning and causal interaction, often appearing clumsy and unreliable when tasked with comprehending how objects move, interact, and change over time. This core contradiction has prompted the industry to re-evaluate the sources of training data that drive AI development.

A prominent CEO of a leading AI company recently articulated a compelling argument during a TechCrunch AI channel discussion: video game environments may offer superior training data compared to the massive, noisy, and unstructured text scraped from the internet. This perspective does not dismiss the value of textual data but highlights a structural flaw in the current training paradigm. The internet, while vast, is filled with noise, bias, and abstract symbols that lack physical anchors. In contrast, the virtual worlds constructed by game engines are closed, controlled systems with explicit physical rules, offering a pathway for AI to transition from merely "reading" the world to actively "experiencing" it.

Deep Analysis

From a technical and commercial standpoint, this viewpoint addresses the fundamental pain points of current AI architectures. The existing Transformer architecture is essentially a next-token prediction model based on statistical probability. It excels at capturing semantic associations within language but lacks an intrinsic understanding of the physical constraints of the real world. For instance, an LLM can generate a perfect description of an apple falling, yet it does not truly comprehend gravity, mass, or collision mechanics unless these concepts are repeatedly emphasized in its training data. In video game environments, however, the dynamics are fundamentally different.

Game engines such as Unity or Unreal Engine enforce strict physical laws on all objects. When an AI agent attempts to move, grab, or jump within these environments, it must process sensor inputs in real-time, predict the consequences of its actions, and adjust its strategies based on environmental feedback. This trial-and-error mechanism is the core of Reinforcement Learning. In game scenarios, cause and effect are immediate and unambiguous: push a block, and it falls; miss a ledge, and the character drops. This high-density causal feedback loop allows AI to learn object permanence, spatial relationships, and physical interaction rules with high efficiency—skills that pure text data struggles to provide. Furthermore, game environments allow for the generation of infinite variations of scenes and tasks, solving the high cost and difficulty of labeling real-world data, thereby providing an ideal sandbox for training generalized embodied agents.

Industry Impact

This technological trajectory is reshaping the competitive landscape and the future of AI applications. Firstly, it intensifies the arms race among tech giants in simulation infrastructure. Companies with strong game engine technologies or those developing high-fidelity simulation platforms, such as NVIDIA with its Omniverse platform, Unity, and major cloud service providers, are transitioning from content providers to key infrastructure providers for AI training. Secondly, this trend directly benefits the fields of Embodied AI and robotics.

Startups like Figure AI and Tesla with its Optimus robot face the core challenge of enabling robots to operate flexibly in complex environments. Large-scale pre-training in games can significantly reduce the training time and cost required for robots in the real world, facilitating a smooth "Sim-to-Real" transfer. For end-users, this implies that future AI assistants will evolve beyond chatbots into embodied intelligences capable of understanding physical commands, controlling smart home devices, driving cars, or operating robotic arms. The competitive focus is shifting from a mere比拼 of model parameters to the fidelity of simulation environments, the richness of interaction data, and the efficiency of simulation-to-reality transfer. Enterprises that can construct high-quality, diverse gamified training environments will gain a strategic advantage in the AGI race.

Outlook

Looking ahead, the transformation of AI training paradigms will not happen overnight, but several key signals warrant attention. First, it is crucial to observe whether major AI laboratories will integrate "gamified training" into their core R&D roadmaps rather than treating it as a peripheral experiment. Projects such as DeepMind's Alpha series and Google's RT-2 have already demonstrated the potential of vision-language-action models in game or simulation environments. In the future, we may see more large-scale pre-trained models based on general game engines emerge. Second, the prevalence of open-source game datasets and simulation platforms will determine the speed of innovation in this field.

If high-quality interaction data can be widely shared like open-source code, it will accelerate innovation among small teams and academia. Finally, changes in evaluation standards will be a critical indicator. Traditional NLP benchmarks like MMLU will gradually give way to comprehensive evaluation systems that include physical reasoning, task planning, and multimodal interaction. While completely replacing text data is unrealistic, a multimodal hybrid training paradigm of "text plus interaction" is highly likely to become the mainstream architecture for AGI. This shift is not just about technical efficiency but concerns whether AI can truly understand and integrate into the physical world we live in, evolving from a tool into a genuine intelligent partner.