---
date: '2024-02-07'
description: artificial general intelligence, measured by breadth of tasks and performance, with autonomy and self-improvement as separate questions.
id: AGI
modified: 2026-06-05 15:08:27 GMT-04:00
tags:
  - seed
title: AGI
created: '2024-02-07'
published: '2024-02-07'
pageLayout: default
slug: thoughts/AGI
permalink: https://aarnphm.xyz/thoughts/AGI.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Artificial general intelligence (AGI) concerns how broadly an AI system can perform intellectual tasks and learn new skills. Definitions differ over the tasks, human reference group, and required level of performance.

[Morris et al.’s _Levels of AGI_](https://proceedings.mlr.press/v235/morris24b.html) proposes two capability dimensions:

- **Performance**: how well the system does a task relative to people with the relevant skill.
- **Generality**: the range of tasks on which it reaches that level, including learning new skills.

They treat **autonomy**, how much a system acts without human involvement, as a separate deployment question. A capable model can be used with a person reviewing every action.

For [[thoughts/LLMs]], success on a [[thoughts/NLP|natural-language]] task provides evidence about that task. An AGI claim needs a stated threshold and evidence across tasks. Fluent explanations alone leave the reliability of the answers untested.

## superintelligence

In this framework, artificial superintelligence means exceeding human performance across a broad range of tasks. AGI definitions can use lower performance thresholds. Learning a new skill and improving the system’s own design are also different capabilities.

Recursive self-improvement adds a causal claim: better machine design produces a system better at machine design, repeating the process ([Good, §2](https://www.sciencedirect.com/science/article/pii/S0065245808604180)). That feedback needs its own evidence.

![[singularity.canvas]]