AGI: specify what generality would mean

Artificial general intelligence (AGI) describes broadly capable AI rather than a system restricted to one narrow task. Definitions differ over the required breadth, performance level and ability to learn new tasks.

Those dimensions should be stated separately. A system can perform at a high level on a fixed test collection while needing extensive new training for a small change in the task.

Test the adaptation, not only the score

“Levels of AGI” separates performance and generality and discusses autonomy as another deployment dimension. “On the Measure of Intelligence” emphasizes how efficiently a system acquires skills given its prior knowledge and experience.

A useful comparison might present unfamiliar game rules, limit demonstrations and computation, and measure how quickly performance improves. That probes something different from replaying a well-practised benchmark.

Consequently, an AGI claim needs an operational definition: which tasks, how much prior exposure, what adaptation budget and which success threshold? Autonomous tool use, fluent language and broad competence are related possibilities, not interchangeable measurements.