What would make an AI “general”?
Breadth, transfer and self-direction, on most accounts. A general system would handle unfamiliar tasks without retraining, carry learning across domains the way people do, reason causally, and improve from its own experience; IBM’s overview is representative in calling AGI a hypothetical stage at which systems match or exceed human cognitive abilities across tasks (IBM, “What is artificial general intelligence?”, read 26 October 2026). Today’s artificial intelligence, including every frontier model, is jagged instead: remarkable on dense-data tasks, brittle one step sideways. Generality is the missing property, not raw scores, and no benchmark yet measures it to anyone’s satisfaction.
Why does everyone disagree about AGI?
Because the word does economic work. There is no consensus definition in the research community, and the loudest definitions are interested ones: labs raising capital frame AGI as automation of most economically valuable work, skeptics raise the bar toward human-like understanding, and at least one famous partnership tied the term to commercial milestones. Adjacent research camps, world models, embodied systems, reasoning models, each cast their path as the road to it. When someone says AGI, ask what test they would accept as failure; an unfalsifiable AGI claim is a slogan.
What should a business do with AGI talk?
Discount it to capabilities and dates. Procurement decisions are about what a system does this quarter, under your constraints, at your risk tolerance; AGI timelines, long or short, change none of that arithmetic. The term matters to a buyer in exactly two places: strategy, where betting on capability curves continuing beats betting on a discontinuity, and risk, where the governance and AI alignment questions raised for future systems, who controls it, what it optimizes, already apply in miniature to the agents being deployed now. Plan for better tools arriving steadily, and govern today’s systems as if autonomy were the risk, because it already is.