Levels of AGI: Operationalizing Progress on the Path to AGI
paperAuthors
Credibility Rating
Good quality. Reputable source with community review or editorial standards, but less rigorous than peer-reviewed venues.
Rating inherited from publication venue: arXiv
A Google DeepMind paper offering a structured framework for categorizing AGI progress; widely referenced in discussions about defining and measuring AGI milestones and their safety implications.
Paper Details
Metadata
Abstract
We propose a framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models and their precursors. This framework introduces levels of AGI performance, generality, and autonomy, providing a common language to compare models, assess risks, and measure progress along the path to AGI. To develop our framework, we analyze existing definitions of AGI, and distill six principles that a useful ontology for AGI should satisfy. With these principles in mind, we propose "Levels of AGI" based on depth (performance) and breadth (generality) of capabilities, and reflect on how current systems fit into this ontology. We discuss the challenging requirements for future benchmarks that quantify the behavior and capabilities of AGI models against these levels. Finally, we discuss how these levels of AGI interact with deployment considerations such as autonomy and risk, and emphasize the importance of carefully selecting Human-AI Interaction paradigms for responsible and safe deployment of highly capable AI systems.
Summary
This paper proposes a taxonomy of AGI progress organized along dimensions of performance depth, task breadth, and autonomy, introducing tiered 'Levels of AGI' analogous to autonomous driving levels. The framework aims to establish a common language for benchmarking AI systems, assessing risks, and guiding safe deployment decisions as capabilities advance.
Key Points
- •Proposes a 2D taxonomy: 'depth' (task performance relative to humans) and 'breadth' (generality across domains), plus autonomy as a third dimension.
- •Defines six guiding principles, including focusing on capabilities over mechanisms and establishing meaningful intermediate milestones rather than only a binary AGI endpoint.
- •Introduces five levels of AGI (Emerging, Competent, Expert, Virtuoso, Superhuman) applicable to both narrow and general systems.
- •Argues that autonomy level should be calibrated to capability and context, emphasizing human oversight as a key safety mechanism for high-capability systems.
- •Aims to provide a shared vocabulary for researchers, policymakers, and developers to compare systems and measure progress toward AGI responsibly.
Cited by 2 pages
| Page | Type | Quality |
|---|---|---|
| AI Timelines | Concept | 95.0 |
| Controlled Vocabulary for Longtermist Analysis | -- | 55.0 |
Cached Content Preview
\\correspondingauthor
merrie@google.com
# Levels of AGI: Operationalizing Progress on the Path to AGI
Meredith Ringel Morris
Google DeepMind
Jascha Sohl-dickstein
Google DeepMind
Noah Fiedel
Google DeepMind
Tris Warkentin
Google DeepMind
Allan Dafoe
Google DeepMind
Aleksandra Faust
Google DeepMind
Clement Farabet
Google DeepMind
Shane Legg
Google DeepMind
###### Abstract
We propose a framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models and their precursors. This framework introduces levels of AGI performance, generality, and autonomy. It is our hope that this framework will be useful in an analogous way to the levels of autonomous driving, by providing a common language to compare models, assess risks, and measure progress along the path to AGI. To develop our framework, we analyze existing definitions of AGI, and distill six principles that a useful ontology for AGI should satisfy.
These principles include focusing on capabilities rather than mechanisms; separately evaluating generality and performance; and defining stages along the path toward AGI, rather than focusing on the endpoint. With these principles in mind, we propose “Levels of AGI” based on depth (performance) and breadth (generality) of capabilities, and reflect on how current systems fit into this ontology. We discuss the challenging requirements for future benchmarks that quantify the behavior and capabilities of AGI models against these levels. Finally, we discuss how these levels of AGI interact with deployment considerations such as autonomy and risk, and emphasize the importance of carefully selecting Human-AI Interaction paradigms for responsible and safe deployment of highly capable AI systems.
###### keywords:
AI, AGI, Artificial General Intelligence, General AI, Human-Level AI, HLAI, ASI, frontier models, benchmarking, metrics, AI safety, AI risk, autonomous systems, Human-AI Interaction
## 1 Introduction
Artificial General Intelligence (AGI)111There is controversy over use of the term “AGI.”
Some communities favor “General AI” or “Human-Level AI” (Gruetzemacher and Paradice, [2019](https://ar5iv.labs.arxiv.org/html/2311.02462#bib.bib30 "")) as alternatives, or even simply “AI” as a term that now effectively encompasses AGI (or soon will, under optimistic predictions). However, AGI is a term of art used by both technologists and the general public, and is thus useful for clear communication.
Similarly, for clarity we use commonly understood terms such as “Artificial Intelligence” and “Machine Learning,” although we are sympathetic to critiques (Bigham, [2019](https://ar5iv.labs.arxiv.org/html/2311.02462#bib.bib10 "")) that these terms anthropomorphize computing systems. is an important and sometimes controversial concept in computing research, used to describe an AI system that is at least as capable as a human at most tasks. Given the rapid advancement of Machine Learning (ML) models, the concept of AGI has pass
... (truncated, 86 KB total)c2393fe3db65d3ef | Stable ID: MWIyMjEzOT