Game of Life as a Path to AGI: A New View of the Nature of Mind

The question of how to create true artificial general intelligence remains the industry's biggest unsolved mystery. While giants like OpenAI, Google DeepMind, and Anthropic scale up language models, hoping quantity will translate into quality, independent researchers are searching for fundamentally different paths. One such attempt appeared on Habr—and despite its ambition, deserves attention if only because it asks the right questions.

The article's author conducted an intriguing experiment: he created a neural network tasked with recovering the rules of Conway's Game of Life cellular automaton. But with a fundamental constraint: the network observes the system's dynamics from inside, without access to the complete picture. It doesn't see all cells simultaneously and doesn't know the rules in advance. Instead, it must build a model of the world based on limited observations—much like how a biological organism perceives reality through sensory organs, without direct access to the "source code" of physical laws.

This approach fundamentally differs from standard machine learning, where models typically receive data "from a bird's-eye view." A neural network trained to classify images sees every pixel. A language model processes entire texts. Here, the agent is immersed in an environment and forced to approximate its patterns based on local, incomplete information. This, the author claims, is how the human brain works—and this principle might be key to creating AGI.

Based on his experiment, the researcher builds a broader theoretical framework. He introduces his own terminology system—central to which are "infures"—and describes the mathematical-informational nature of mind within the context of the computable universe hypothesis. This hypothesis, traced back to Konrad Zuse's ideas and later developed by Stephen Wolfram, suggests that all physical reality can be described as a computational process. If this is true, then consciousness isn't a mystical property of biological matter, but a natural result of a particular type of information processing that could in principle be reproduced.

To be honest: the Habr article is more of a manifesto by an independent thinker than peer-reviewed research. Original terminology without connection to established scientific frameworks, broad generalizations based on one experiment, and straightforward parallels between cellular automata and the human brain naturally invite skepticism. Neurobiology and cognitive science have accumulated enough data showing that biological intelligence's architecture is far more complex than any model describable through observing a grid of black and white cells.

Yet behind the ambitious wrapper lies an idea that cannot simply be dismissed. The question of embodied cognition's role in forming intelligence has been discussed in academia for decades. François Chollet, creator of Keras and author of the ARC test for measuring general intelligence, has repeatedly emphasized that scaling transformers alone won't lead to AGI, because these models lack precisely the interaction with environment. Similar arguments come from Yann LeCun at Meta, who champions world models—internal models of the world that an agent builds through experience rather than absorbing texts.

In this context, the Game of Life experiment can be seen as a simplified but illustrative demonstration of how an agent can derive structural patterns from an observer's position. This isn't AGI, and not even a direct path to it. But it's an attempt to formalize the intuition that real intelligence doesn't start with data, but with experience. Not with answers, but with questions the agent learns to ask.

The artificial intelligence industry in 2026 is at a point where scaling shows diminishing returns, and no major architectural breakthroughs are yet visible. In such a situation, any fresh ideas are valuable—even if they come not from a Google laboratory but from Habr. The path to AGI, if it exists at all, is unlikely to be straight. And it's quite possible that one of its turns will indeed pass through cellular automata, embodied experience, and the question of how a system can know the world from within it.