AGI (Artificial General Intelligence)
A hypothetical system with broad human-level capability across domains — undefined enough that people can argue about whether it's arrived.
When not to use it
- In technical planning. It's not a specification and it doesn't inform any decision you'll make this year.
- As a reason to ignore present harms. Systems deployed today affect people today, whatever arrives later.
- As a claim about a product. "A step toward AGI" is unfalsifiable and usually means the benchmarks moved.
Reach for something else instead
- Specific capability claims — "it does X at Y accuracy on Z" is testable and therefore means something.
- Task-level evaluation for anything you're building. Your problem is your problem.
- Levels or continuum framings if you must discuss general capability — at least they're operationalisable.
Sources & further reading
- Morris et al. (2023), Levels of AGI: Operationalizing Progress on the Path to AGI — an attempt to make the term measurable, and a fair account of why it's hard.
- Bubeck et al. (2023), Sparks of Artificial General Intelligence — the most cited argument that something changed, and worth reading with its critics.
- Chollet (2019), On the Measure of Intelligence — the case that current benchmarks measure skill, not intelligence, and a proposed alternative.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reasoning from impressive performance on one task to expected performance on another. Human abilities correlate; a model's don't.
- Treating disagreement about timelines as a factual dispute. It's largely a definitional one.
- Assuming the definitional vagueness is accidental. Plenty of people have reasons to define it where it suits them.
At a glance
FieldFoundations
Core ideabroad human-level capability
Statushypothetical, undefined
Track recordseventy years of optimistic predictions
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
AGI vs. narrow AI — hypothetical breadth vs. the systems that actually do things.