Jagged Intelligence

Overview

LLM performance is “jagged”: excels in specific tasks (e.g. coding certain patterns) but fails unexpectedly in adjacent areas. Unlike smooth human learning curves.

Key Facts

  • From The Kaparthy LLM Wiki: Peaks/valleys in capabilities; expect surprises.
  • Implications: Narrow specialization over general intelligence.
EntityDescription
Andrej KarpathyCoined/observed in LLM Wiki.
ConceptLink
Vibe CodingLeverages peaks.

Sources/References

Connections

  • Complements LLM-Wiki: Tailor ingests to peaks.
  • Vs Agentic AI: Agents smooth jaggedness via tools.