The frontier of intelligence has repeatedly learned from games.
Games helped prove reinforcement learning, one of the central paradigms behind today's most capable AI. They gave agents a tight loop of action, consequence, and verifiable reward—and gave researchers environments that could be copied, accelerated, reset, instrumented, and measured.
In the current moment, ARC-AGI-3 uses novel game environments to test something closer to general intelligence: whether an agent can explore an unfamiliar world, infer its rules and goals, build a useful model, and act efficiently without instructions. Humans solved every environment in its technical report; frontier systems scored below one percent.
Looking forward, games are becoming training grounds for world and action models. Labs are learning from gameplay at scale, generating interactive worlds, and placing embodied agents inside them. The first applications are already close to games because games bring perception, action, causality, and other agents together in one controllable medium.
Games may be the best instruments we have created for porting human intelligence and interaction into virtual worlds. They preserve more than answers: they capture intent, action, consequence, cooperation, competition, creativity, and judgment as processes unfolding over time.
The frontier of intelligence still has a great deal to learn from games. Games are not AGI, but they may be one of the most productive environments for building and evaluating it.
Games made learning from consequences scalable
Reinforcement learning predates deep learning; games made it scalable and verifiable. Atari DQN learned policies from pixels and reward, AlphaZero made self-play the data engine, and OpenAI Five carried the loop into a partially observed, long-horizon, multi-agent world.
The durable result was not a sequence of game-playing systems but a training paradigm: act, observe the consequence, update, and repeat. Games made that loop cheap enough to run at scale and precise enough to tell whether learning was real.
Self-play turned the learning loop into a curriculum
In self-play, the current policy generates the opponent and therefore the next useful data. Each improvement creates stronger counter-strategies, keeping the training distribution near the learner's capability frontier instead of freezing it in a dataset.
Multi-agent games add cooperation, communication, deception, and emergent tool use; OpenAI's hide-and-seek is the canonical example. The important property is automatic task generation: self-play does not replace good objectives or environment design, but it can continuously answer where the next difficult experience comes from.
ARC-AGI tests whether that learning transfers to unseen worlds
There is an obvious objection to this thesis: a chess engine can be superhuman at chess and useless everywhere else. That objection is correct. Mastery is not generality.
The more important test is what happens when an agent enters a world it has never seen. Can it explore deliberately? Infer the rules? Identify the goal? Build a useful model from a handful of interactions? Change its mind when the evidence changes?
ARC-AGI-3 turns those questions into a benchmark. It places agents in novel, interactive game environments without instructions. Success requires exploration, world-model construction, goal identification, planning, and execution—not the retrieval of a memorized playbook. The benchmark measures how efficiently an agent can generalize to an unseen world. In the March 2026 technical report, the environments were 100 percent solvable by human testers while frontier AI systems scored below one percent. (ARC-AGI-3 technical report)
This is a profound change in what games measure. Chess asked whether a machine could search better than us inside a known world. ARC-AGI asks whether it can reason its way into a new one.
World models turn game data into learned dynamics
World models learn environment dynamics well enough to predict and plan. Games are a scalable source of action-conditioned trajectories: rich enough to require spatial and temporal reasoning, but structured enough to reset, accelerate, and evaluate.
The pattern is already visible across the frontier. General Intuition's MIRA trains on 10,000 hours of Rocket League trajectories; DeepMind's Genie generates controllable worlds from video; and GameNGen learns to simulate Doom autoregressively. Different architectures, same substrate: game data turns passive video into models of what actions do.
Paired with embodied agents such as SIMA, the direction is straightforward: generate worlds, place agents inside them, and use interaction to improve both the model and the policy. Games are the fastest practical version of that loop.
Game generation tests whether models can construct coherent worlds
It is no surprise that making games is one of the first things people want to do with AI. Games turn a model's capabilities into something immediate and legible: not a benchmark score or a polished answer, but a world you can enter, test, share, and judge for yourself.
As models reach the frontier, game generation is becoming a kind of public evaluation. The question is no longer only whether a model can write code that runs. It is whether it can make the game a person actually asked for—and whether the result is coherent, surprising, impressive, and fun enough that anyone wants to keep playing.
Language models became useful programmers by producing code that could be checked against a compiler, a test suite, or a known answer. Games offer verification too, but the generation problem is far more multidimensional.
A working game must hold many systems together at once:
- Rules and state: the objectives must be coherent, every action must update the world correctly, and edge cases cannot corrupt the run.
- Space and time: movement, physics, collisions, cameras, animation, and pacing all have to agree frame after frame.
- Presentation: art, sound, interface, feedback, and controls must make the underlying system legible.
- Other people: multiplayer adds networking, synchronization, lobbies, fairness, and the unpredictability of human behavior.
- Performance: the whole experience has to respond immediately on real hardware, not merely produce a plausible result eventually.
A model can get any one of these right and still produce a broken game. It can also get all of them technically right and produce a boring one.
That makes game generation an unusually dense test of intelligence. The model must translate ambiguous human intent into executable rules, maintain a causal system over time, compose many creative and technical disciplines, notice failures through play, and revise the result. It is not simply generating code. It is designing a world and then being held accountable by everyone who enters it.
Human play supplies the reward benchmarks miss
Fun is difficult to specify because it does not live in a single metric. It emerges from the relationship between a player and a system: the timing of a jump, the tension before a reveal, the fairness of a loss, the pleasure of mastery, the surprise of another person doing something clever.
These qualities are contextual and often contradictory. A game can be unfair in a way that creates laughter, difficult in a way that creates trust, or simple in a way that produces endless strategic depth. The same mechanic can delight one group and lose another.
This is why human play matters. Players supply a reward signal richer than winning. They retry, improvise, laugh, quit, teach one another, break the rules, and invent goals the designer never anticipated. They reveal not only whether a world functions, but whether it is worth spending time inside.
This is also the longer-term thesis behind Instaplay. Letting anyone turn an idea into a playable world expands the distribution of mechanics, goals, aesthetics, social dynamics, and definitions of fun. The result is not one simulation built by one lab. It is a growing universe shaped by millions of human preferences.
Creation and play form a compounding training loop
- Humans and models invent new games.
- People play them, revealing what works, what breaks, and what is fun.
- Agents learn from those trajectories and explore further through self-play.
- Their failures expose missing capabilities.
- New games are created to train and test those capabilities.
The game generator becomes a teacher. The agent becomes a playtester. Humans keep pulling the curriculum toward worlds that are surprising, meaningful, and worth exploring.
The frontier of intelligence will keep learning from games
Games are not the real world. They simplify consequences, compress values into rules, and allow failures to be reset. No final high score will flip a switch labeled “AGI.”
But games give intelligence somewhere to practice being an agent. They demand perception and action, reasoning and adaptation, competition and cooperation, invention and taste. They can generate nearly unlimited experience while keeping outcomes observable enough to learn from.
The path so far has moved from search to reinforcement learning, from human examples to self-play, from one fixed game to unseen worlds, and from hand-authored environments to worlds generated on demand. Each step has made games look less like a benchmark at the edge of AI research and more like the engine underneath it.
If games bring us to AGI, then the world looks promising.
Hopefully AGI is fun.