AGI

Today’s systems are narrow in a specific sense: each is strong within the distribution of its training and unreliable outside it. A model that writes excellent code can fail at a puzzle a child solves, not because the puzzle is harder but because it is unlike anything in the data. AGI names the absence of that boundary.

Why nobody can tell you when

Because there is no agreed test, and the disagreement is not a detail — it is the whole argument.

Some define AGI by benchmarks, in which case parts of it arrived years ago. Some define it economically: a system that can do most remote work. Some require the ability to learn a genuinely new skill from few examples, the way a person picks up an unfamiliar tool. Some insist on understanding rather than performance, which raises the problem of testing for something we cannot define in ourselves.

A prediction of “AGI by 2030” is therefore not really a prediction until you know which definition is meant. The same forecaster with two definitions gives two answers a decade apart.

What to listen for

When a company announces progress toward AGI, the useful question is what changed and how it was measured. Benchmark scores rise steadily and say little about the boundary that matters. What would be genuinely new is reliability on tasks the system was not built for — and that is the thing least often demonstrated, because it is hardest to show.

The term also does commercial work. “On the path to AGI” is a fundraising sentence as much as a technical one, and treating it as a claim about capability rather than positioning is usually a mistake.

Worth separating from superintelligence, which means capability beyond human level rather than equal to it, and from AI agent, which describes how a system acts rather than how broadly capable it is.