Definition
Everyone's talking about AGI. Few agree on what it means, and that's the first thing to know.
Artificial General Intelligence (AGI) refers to a hypothetical AI that can learn and perform nearly any thinking task a person can, across domains, without being built specifically for each one. There's no single accepted definition, and no agreed test.
In simple words
Today's AI is like a collection of specialists, each great at its own thing. AGI would be like a person who, given time, can learn to do your taxes, then learn chemistry, then learn to manage a team, using what they learned before.
Why it matters
AGI drives huge investment, national policy and many headlines. Without knowing that people mean different things by it, it's easy to be misled in either direction, by claims that it has arrived or that it's impossible.
How it works
| Who | How they frame it |
|---|---|
| Legg & Hutter | An agent's ability to achieve goals in a wide range of environments |
| François Chollet | How efficiently a system learns new skills on tasks it hasn't seen |
| Many companies | Systems that can do most economically valuable work |
| Google DeepMind, 2023 | Levels of performance and generality, rather than a yes-or-no line |
Why it's hard to say
- Current systems do impressive things, and can also fail at things a child finds easy.
- Test scores can rise through narrow methods that don't generalise.
- Experts' forecasts vary by decades.
Common mistakes
- Treating AGI as one clear finish line.
- Thinking AGI needs consciousness. Most definitions focus on what it can do, not whether it feels anything.
- Mistaking a persuasive chat for general ability.
Check yourself
Summary
- AGI is a hypothetical AI with human-like generality.
- There's no agreed definition or test.
- Ask which definition and what evidence before believing a claim.
Finished reading?
Sources: Morris et al., Levels of AGI (ICML 2024) · ARC Prize