What AGI Means: One Design, Infinite Tasks

In this explanatory post intended for a broad audience, the author defines AGI by drawing a sharp distinction from how AI systems work today. The current approach to AI development requires that researchers and engineers conduct R&D, gather more training data, build new training environments, and change or scale up algorithms in order to improve performance on any given task—then repeat this cycle entirely for the next task. The author illustrates this with a concrete example: if you want an AI to drive a car or control a computer using a mouse, it is a huge project involving dozens of experts working for years to build a new AI system. By contrast, the author describes AGI as a single design that, once created, would not require such repetitive engineering. Instead, many copies of this one AGI design could autonomously learn to do everything in the global economy—just as many copies of one human brain design, barely changed since the African savannah, have built the global economy from scratch. This analogy anchors the definition: human cognition, with a single core design refined over evolutionary time, has proven capable of mastering an enormous range of domains and tasks without needing a separate "brain variant" for each domain. The author frames AGI as a system that would share this property of general-purpose, autonomous learning. The post concludes by acknowledging that this definition will strike different readers differently—obvious to some, obviously wrong to others—reflecting the genuine disagreement in the research community about what AGI is or should be.

The post establishes a conceptual boundary between narrow AI and AGI by focusing on scalability and generalization rather than raw capability. Today's AI systems are specialized tools—each domain (autonomous driving, computer control, etc.) demands its own R&D cycle, expert teams, and domain-specific training. The author argues AGI represents a categorical difference: a single underlying design that learns autonomously across domains, much as human cognition transfers knowledge from one task to another without retraining from first principles. This framing emphasizes that AGI is not merely "better AI" but a fundamentally different kind of system—one that generalizes across the entire scope of economic and intellectual work. The author explicitly notes that this definition invites disagreement, acknowledging that researchers and observers hold varying theories of what constitutes AGI.