A system is general-purpose to the degree that the set of tasks it can be directed at exceeds the set it was optimised for. The term is preferred here over AGI because it names a measurable property rather than a threshold, and because the threshold framing tends to collapse into an argument about definitions.

Three axes worth separating

Generality is usually discussed as one quantity. It is at least three, and systems can be strong on one while weak on another:

• Task breadth — how wide the set of addressable tasks is at a fixed set of weights.
• Specification cost — how much effort it takes to direct the system at a new task. Retraining is expensive; a natural-language instruction is nearly free.
• Accumulation — whether competence acquired during operation persists. A system that solves a novel task and retains nothing is general in breadth but not in time.

Current large models are strong on the first two and close to absent on the third, which is why continual learning and catastrophic forgetting recur throughout this archive.

What it is not

Not a claim about consciousness, agency, or resemblance to human cognition. Not a synonym for large. A very large model narrowly specialised is not general-purpose; a small model that can be steered across many domains partly is.

Regulatory use of the term

“GPAI” also appears as a legal category in AI regulation, where it carries a specific and narrower technical definition tied to training compute and capability thresholds. That usage is related but not identical to the one above; where an entry means the legal sense, it says so.