Playgrounds for AGI

Artificial General Intelligence won't grow out of static benchmarks or passive datasets. It demands agents that interpret language, reason over time, adapt under uncertainty, and act in open-ended settings. Models trained only in synthetic or scripted environments soon plateau; real intelligence is forged through interaction, not imitation. Language-driven, AI-native interactive environments—games—offer the ideal training ground: clear rules meet ambiguity, sparse rewards meet symbolic complexity, and long-horizon planning must be expressed in natural language.

Reinforcement learning already produces super-human results whenever an agent receives a faithful, fully specified environment and ample compute. But today’s “faithful” testbeds (math proofs and coding) are short-horizon and perfectly verifiable. Most real-world tasks aren’t. They unfold over many steps, hide important context, and hinge on the unobservable chain of thought a human brings to a problem. Capturing those dynamics demands richer worlds than any static benchmark can provide.

We launched 4Wall in beta as an AI entertainment platform and quickly grew to 100k+ creators who interact with AI characters inside semantically rich virtual worlds. That scale has given us deep insight into how people naturally engage with language agents and where today’s frontier models still stumble. Claude, for instance, can spend hours stuck behind a rock wall in Pokémon Red yet easily describe the workaround in a separate chat; even advanced models like o3 remain susceptible to simple jailbreaks despite using intricate chain-of-thought in their reasoning.

To fill this gap, 4Wall is evolving into a scalable learning substrate for frontier AI. Our worlds are multi-agent and language-native by design: living environments where humans and (eventually) AIs coexist, collaborate, and compete. Every play session becomes a high-signal training trace of compositional reasoning, social behavior, and exploratory planning. As leading labs shift from RLHF toward pure RL on the road to AGI, 4Wall is the first purpose-built platform ready to fuel that transition at scale.