For years, the AI conversation around artificial general intelligence has been dominated by language: models that can summarize, answer questions, write code, and hold surprisingly coherent conversations. But the more researchers and builders push toward systems that can operate reliably in the messy, physical world, the more a different limitation comes into focus—one that doesn’t show up as clearly in text benchmarks.
According to the CEO behind General Intuition, the missing ingredient may not be another clever architecture or a bigger dataset of words. It may be something far more fundamental: training signal that teaches an agent how the world behaves as it changes over space and time. And in a twist that will sound counterintuitive to anyone who thinks “the internet is the biggest dataset on Earth,” the company’s bet is that video games could be a better teacher than the web for this specific capability.
The argument starts with a distinction that many people intuitively understand but rarely quantify: language models are excellent at patterns in language, yet they are not naturally grounded in the physics of action. They can describe motion, explain cause-and-effect in abstract terms, and even simulate reasoning steps in text. But when you ask them to internalize how things move, collide, accelerate, occlude, and recover—how outcomes depend on what an agent does next—the learning problem becomes less like “reading” and more like “experiencing.”
That’s where General Intuition’s framing lands. The CEO’s claim is essentially that if you want intelligence that generalizes beyond language—intelligence that can plan, act, and adapt—you need training that forces the model to build a world model. Not just a model of words about the world, but a model of the world itself: a representation that captures how observations relate to hidden state, how actions change that state, and how the future unfolds from the present.
And while the internet contains enormous amounts of information about the world, it doesn’t provide the one thing that matters most for learning dynamics: consistent, interactive feedback loops.
The web is full of descriptions. Video games are full of consequences.
Why text-heavy training hits a ceiling
To understand why the internet might be insufficient for this particular goal, it helps to look at what language training actually optimizes. Large language models learn statistical relationships between tokens. Even when they’re trained with instruction-following or reinforcement learning from human feedback, the core signal still comes from language outputs and preference judgments—not from direct interaction with a system governed by physical rules.
That means the model can become extremely good at producing plausible explanations without necessarily developing robust internal representations of how the world behaves under intervention. In other words, it can learn to talk about causality without truly mastering causality as a mechanism.
This gap becomes obvious when you move from “What happens if…?” questions to “Do it.” Real-world intelligence requires an agent to choose actions, observe results, update its beliefs, and continue. It requires temporal coherence: the ability to keep track of what changed, what stayed the same, and what must be true for the next observation to make sense.
Text datasets can include stories, diagrams, and step-by-step instructions, but they don’t enforce the same kind of consistency that an environment enforces. A story can say that a character jumps and then lands somewhere else; the model can learn the pattern of narrative structure. But it doesn’t have to learn the underlying constraints that make the landing possible. The “physics” is implicit in the author’s description, not explicit in the environment’s rules.
Video games, by contrast, are rule-bound. They are interactive simulations where the world responds deterministically (or stochastically, but consistently) to actions. That makes them a natural training ground for learning dynamics.
The case for gaming data: action, observation, and adaptation
General Intuition’s thesis is not that video games are magical or that they replace real-world data. It’s narrower and more technical: video games may provide a richer training signal for building world models than the internet does.
In a game environment, an agent repeatedly experiences a loop:
1) observe the current state (visual input, sometimes audio or other signals),
2) decide on an action,
3) see the outcome,
4) infer what must have been true about the hidden state,
5) repeat.
This loop creates a continuous stream of supervision. The agent learns not only what happens, but how what happens depends on what it did. Over time, it can develop internal representations that support planning and prediction.
Even when the game is stylized, the learning pressure is similar to what robotics faces: the agent must handle partial observability, deal with changing scenes, and maintain temporal consistency. It must also learn to generalize across variations—different levels, different enemy behaviors, different layouts, different lighting conditions, different strategies.
That’s a key point. The internet can be diverse, but it’s not interactive. Diversity without interaction often becomes diversity of descriptions rather than diversity of causal mechanisms. Games force the model to confront the causal structure of the environment because the environment is the source of truth.
There’s also a practical advantage: games can generate massive amounts of data cheaply and repeatedly. You can reset the environment instantly, run millions of episodes, and explore counterfactuals by varying actions. You can also control the distribution of tasks and difficulty. This makes it feasible to train models on long-horizon behavior—something that is notoriously hard to obtain from static text corpora.
World models: the bridge between language and physical reasoning
The phrase “world model” gets used a lot in AI circles, but it’s worth unpacking what it means in this context. A world model is a learned representation that allows an agent to predict future observations given current state and planned actions. It’s not just a memory of what happened before; it’s a compact internal structure that supports imagination-like rollouts.
Language models can sometimes approximate this through text-based simulation, but their “simulation” is constrained by the fact that they’re generating text, not interacting with a system. Their internal representations are shaped by language statistics, not by the geometry and dynamics of the environment.
When you train on game interactions, you encourage the model to learn representations that are useful for prediction and control. The model must connect pixels (or other sensory inputs) to latent variables like position, velocity, object identity, and state transitions. It must learn how actions affect those variables. It must also learn to recover when the environment changes unexpectedly.
This is exactly the kind of capability that generalizes beyond a single task. If the model learns a robust representation of dynamics, it can transfer that knowledge to new scenarios within the same environment family—or potentially across families if the training is broad enough.
General Intuition’s bet, as described in the TechCrunch segment, is that gaming data can help fill the gap between text understanding and broader intelligence. The company’s CEO argues that models trained primarily on language may struggle with “how things move through space and time,” and that games provide a structured way to teach that skill.
It’s not just about movement, either. Space-time reasoning includes:
– object permanence (what remains even when not visible),
– occlusion handling (what you can infer when you can’t see),
– collision and contact dynamics (what happens when objects interact),
– delayed effects (actions whose consequences appear later),
– planning under uncertainty (choosing actions that reduce ambiguity).
Games are full of these requirements. They are, in a sense, a curriculum for learning the rules of the world through interaction.
Why “better than the internet” doesn’t mean “replace everything”
The claim that video games could be better training data than the internet can sound like a dismissal of web-scale learning. But the more accurate interpretation is that games address a specific weakness: the lack of interactive, temporally consistent supervision for dynamics.
The internet is still valuable. It contains language, knowledge, and cultural context. It can teach an agent what humans talk about, how concepts are defined, and how tasks are described. It can also provide examples of reasoning in natural language, which is important for communication and instruction following.
But if the goal is AGI-like generalization—systems that can operate in the world—then the training mix needs more than language. It needs grounding. It needs experience. It needs a way to learn that actions have consequences in a structured environment.
Games offer that grounding at scale. They can also be used to create synthetic tasks that would be expensive or dangerous in the real world. For example, you can train agents to navigate complex spaces, manipulate objects, or coordinate with other agents without needing physical hardware.
The unique take here is not simply “use games.” It’s “use games as training signal for the missing component of general intelligence: dynamic world understanding.”
A curriculum of simulated reality
One reason gaming data is compelling is that it can be shaped into a curriculum. Unlike the internet, where the distribution of content is largely uncontrolled, game environments can be designed to cover a spectrum of challenges.
You can start with simpler mechanics and gradually increase complexity:
– from static scenes to moving objects,
– from short-horizon tasks to long-horizon planning,
– from deterministic outcomes to stochastic environments,
– from single-agent control to multi-agent interaction.
This matters because world modeling is not a single skill. It’s a stack of skills that emerge from repeated exposure to structured dynamics. A model that learns only shallow correlations won’t be able to predict future states reliably. A model that learns only immediate rewards may fail at long-term planning. Games allow training to target these failure modes directly.
They also allow evaluation in a way that text corpora cannot. In a game, you can measure whether the agent’s actions lead to success. You can test generalization by changing level layouts, enemy behaviors, or environmental parameters. You can probe whether the agent’s internal predictions match actual outcomes.
That kind of evaluation is crucial for determining whether the model is learning dynamics or merely learning surface-level patterns.
The “space and time” gap: why it matters for AGI
The CEO’s emphasis on space and time is not rhetorical. It points