Jagged Intelligence
JAG-id in-TEL-ih-jence
Definition
Jagged intelligence describes the inconsistent capability profile of current AI systems - where models can perform at PhD level on some tasks while failing high-school level challenges in other domains.
Why It Matters
The term, popularized by Demis Hassabis, captures a core barrier to AGI: lack of consistency. An AGI system should be reliable across all domains, but current models show dramatic capability gaps.
Examples
- GPT-4 can win gold at International Math Olympiad but fails simple logic puzzles
- Models can analyze complex philosophy but struggle with consistent chess play
- Claude can write sophisticated code but might miss obvious bugs
Key Insight
"You would expect from an AGI system that it would be consistent across the board." — Demis Hassabis
The jagged profile suggests current architectures have fundamental limitations - they're not uniformly intelligent, but rather have peaks and valleys of capability.
Implications
For researchers: Benchmarks that test one domain don't predict performance in others For users: Don't assume capability in one area transfers to another For AGI: Solving jagged intelligence may require architectural changes, not just scaling
Related Terms
- World Models - One proposed solution for more uniform capability
- Generalization - The underlying problem