# Levels of AGI: Operationalizing Progress on the Path to AGI

**Authors:** Meredith Ringel Morris, Jascha Sohl-dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, Shane Legg  
**arXiv:** [2311.02462](https://arxiv.org/abs/2311.02462)  
**Date:** 2023-11-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.02462v1/

## Key points

- Authors propose a framework for classifying AGI models and their precursors
- Framework introduces levels of AGI performance, generality, and autonomy
- Six principles for an ontology for AGI are identified
- Levels of AGI based on depth and breadth of capabilities are proposed
- Current systems are analyzed within this ontology
- Future benchmarks needed to quantify behavior and capabilities against these levels
- Interaction between levels of AGI and deployment considerations discussed
- Importance of selecting Human-AI Interaction paradigms emphasized
- Comprehensive framework for classifying AGI models is presented

## Summary

In their paper titled "Levels of AGI: Operationalizing Progress on the Path to AGI," authors Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet and Shane Legg propose a framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models and their precursors. The framework introduces levels of AGI performance, generality and autonomy with the aim of providing a common language to compare models, assess risks and measure progress towards AGI. To develop this framework, the authors analyze existing definitions of AGI and distill six principles that an 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 instead of focusing solely on the endpoint. Based on these principles, the authors propose "Levels of AGI" that are based on both depth (performance) and breadth (generality) of capabilities. They also reflect on how current systems fit into this ontology. Additionally, they discuss the challenging requirements for future benchmarks that can quantify the behavior and capabilities of AGI models against these levels. Furthermore, the paper explores how these levels of AGI interact with deployment considerations such as autonomy and risk. The authors emphasize the importance of carefully selecting Human-AI Interaction paradigms for responsible and safe deployment of highly capable AI systems. Overall, this paper presents a comprehensive framework for classifying AGI models based on their capabilities and behavior. It provides a roadmap for assessing progress towards AGI while considering important factors such as generality, performance, autonomy, deployment considerations and human-AI interaction paradigms.