Definition
Levels of AGI is a framework proposed by Google DeepMind researchers (Morris et al., ICML 2024). It replaces the yes-or-no question "is this AGI?" with levels based on performance and generality, plus separate levels of autonomy that describe how systems are used.
In simple words
Self-driving cars have levels: from simple cruise control to no driver at all. This does the same for intelligence. Instead of arguing over whether a system is "AGI", you ask how good it is, at how many things, and how much freedom people give it.
Why it matters
It gives a shared language for measuring progress and judging risk. It also separates two things that get mixed up: what a system can do, and what people let it do. Much of the risk comes from the second.
How it works
Performance levels
| Level | Name | Meaning |
|---|---|---|
| 0 | No AI | No AI involved |
| 1 | Emerging | Equal to or somewhat better than an unskilled person |
| 2 | Competent | Equal to or better than a skilled person at most tasks |
| 3 | Expert | Equal to or better than experts |
| 4 | Exceptional | Equal to or better than the very best people. Earlier versions called this Virtuoso. |
| 5 | Superhuman | Better than any person |
Autonomy levels
Separately, the framework describes the role people give the system: no AI, then AI as a tool, a consultant, a collaborator, an expert, and finally an agent. A capable system used only as a tool carries different risks from the same system acting on its own.
Common mistakes
- Assuming a high score in one area means a high level overall. Generality counts too.
- Treating these levels as measured facts. It's a framework for discussion, and models aren't officially graded on it.
Check yourself
Summary
- Levels of AGI replaces "is it AGI?" with levels of performance and generality.
- Autonomy is measured separately.
- It's a framework for discussion, rather than an official grade.
Finished reading?