Historical source material recovered on 2026-09-09. Capture dates and source limitations apply; recovery is not fresh verification.

Jagged Intelligence

Jagged intelligence describes the uneven capability profile of LLMs: they can perform extremely difficult tasks in some trained/verifiable circuits while failing simple common-sense cases outside those circuits.

Key points

  • Karpathy uses the car-wash example: a model that can refactor a huge codebase or find vulnerabilities may still advise walking to a car wash 50 meters away to wash a car [src-055].
  • The jaggedness comes from a combination of pretraining statistics, RL reward environments, and what labs choose to include in their data distribution [src-055].
  • This is why users must treat models as powerful tools rather than uniformly intelligent agents. They need exploration, oversight, and domain testing to discover where the model flies and where it struggles [src-055].
  • Karpathy's "animals versus ghosts" framing emphasizes that LLMs are not animal intelligences with intrinsic motivation; they are statistical simulation circuits shaped by pretraining and RL [src-055].
  • Jaggedness is closely related to the Verifiability Frontier: capabilities peak where verification rewards exist and degrade where they do not [src-055].

Related entities

Related concepts

Source references

  • [src-055] Sequoia Capital — "Andrej Karpathy: From Vibe Coding to Agentic Engineering" (2026-04-29)

Historical recovered snapshot — 2026-08-20

Preserved from ai-watch-src350-370-romero-pdfs-20260820-051003. Documentary recovery on 2026-09-09; claims below retain their historical source scope and were not freshly verified. Current synthesis above remains authoritative.

  • [src-055] Sequoia Capital — "Andrej Karpathy: From Vibe Coding to Agentic Engineering" (2026-04-29)

Romero PDF batch: benchmark and domain jaggedness (2026-08-20)

  • GPT-5.6 Sol improved ARC-AGI-3 while remaining below average human performance, especially on planning and execution [src-353].
  • Mathematical advances span validated counterexamples, expert-digested proofs, vendor-certified claims and unsettled social posts [src-363].
  • Reliability, cognitive assistance and unaided competence remain task- and scaffold-dependent [src-358] [src-368].

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 500 indexed pages and articles.

  1. Wiki concept Verifiability Frontier The verifiability frontier is the boundary between work that current LLM training methods can automate quickly because outputs can be reliably checked, and work Related by jagged
  2. Wiki concept Tacit Judgment Advantage Historical source material recovered on 2026-09-09. Capture dates and source limitations apply; recovery is not fresh verification. Related by romero
  3. Wiki concept Software 3.0 Andrej Karpathy's frame for programming with LLMs as interpreters: context, prompts, examples, files, images, and instructions become the lever over a Related by 055