Artificial General Intelligence

The distinguishing capability of artificial general intelligence (AGI) is the ability to learn and apply intelligence across a wide range of tasks, rather than perform only a specialised task. AGI refers to the goal of broadly capable intelligence that can transfer learning between domains and adapt to unfamiliar problems, unlike narrow AI, whose capabilities are restricted to particular tasks or domains.

The core distinction: generality of capability

The difference concerns the breadth and flexibility of intelligence, not simply how accurately or quickly a system performs. A narrow system may exceed human performance at one task without possessing general intelligence.

• Narrow AI is designed or trained for particular functions, such as recognising images, recommending products or playing a specific game.
• AGI would combine learning, reasoning and problem-solving across diverse domains instead of requiring a separate specialised system for each kind of task.
• Transfer of learning means using knowledge or skills acquired in one setting to help solve problems in another.

What would demonstrate general intelligence?

The central test is whether a system can use its capabilities flexibly when confronted with unfamiliar tasks, rather than merely reproduce strong performance within a familiar task.

• It would need to learn new tasks and adapt its knowledge to changing circumstances.
• It would need to connect information across domains and select appropriate approaches to different problems.
• High performance on a single benchmark is insufficient to establish AGI, because that result demonstrates competence only within the benchmark's scope.