---
schema: 1
type: term
title: Jagged intelligence
description: >-
  The unevenness of AI model capability: models may excel at hard tasks (such as
  proving theorems) while failing at simpler, common-sense business tasks, so
  their strengths and weaknesses arrive in spikes rather than on a smooth curve.
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/jagged-intelligence-179a5d74e5b5
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/jagged-intelligence-179a5d74e5b5.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# Jagged intelligence

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/jagged-intelligence-179a5d74e5b5

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

The unevenness of AI model capability: models may excel at hard tasks (such as proving theorems) while failing at simpler, common-sense business tasks, so their strengths and weaknesses arrive in spikes rather than on a smooth curve.

## Jagged intelligence in this material The term comes from **Silvio Savarese, Chief AI Scientist at Salesforce**, in his Dreamforce 2025 interview with Dan Shipper. He uses it to name the uneven, spiky profile of model capability: the same system that can prove theorems can stumble on ordinary, common-sense business tasks. [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@220]  It matters for enterprise AI because capability is not a smooth curve you can scale blindly. Savarese pairs the observation with his argument that agents are useless without **connected data and context** — and with his practice of using **simulated organizations with synthetic data** to stress-test agents against the real-world gap that jagged intelligence leaves open. [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@342]   - **Jagged intelligence** describes uneven model capability — brilliant in some domains, weak in others. - Savarese cites the pattern: models that can prove theorems yet fail at common-sense business tasks. - It motivates simulation-based stress-testing of agents before customers ever touch them.