2023 LevelsofAGIOperationalizingProg
- (Morris et al., 2023) ⇒ Meredith Ringel Morris, Jascha Sohl-dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. (2023). “Levels of AGI: Operationalizing Progress on the Path to AGI.” doi:10.48550/arXiv.2311.02462
Subject Headings: Artificial General Intelligence (AGI), AGI Autonomy Categorization, Autonomous Vehicle Categorization.
Notes
- The paper proposes a framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models and their precursors, introducing levels of AGI performance, AGI generality, and AGI autonomy to compare models, assess risks, and measure progress.
- The paper identifies six principles for a useful ontology for AGI: focusing on AGI capabilities, AGI generality, AGI performance, cognitive and metacognitive tasks, AGI potential rather than deployment, ecological validity for benchmarking, and the path to AGI rather than a single endpoint.
- The paper introduces a matrixed leveling system based on performance (depth of capabilities) and generality (breadth of capabilities), specifying minimum performance levels for different AGI ratings, such as Emerging AGI, Competent AGI, Expert AGI, Virtuoso AGI, and Superhuman AGI.
- The paper discusses the implications of these principles for developing benchmarks that quantify AGI behavior and capabilities, emphasizing the importance of choosing tasks that align with real-world values and focusing on ecological validity.
- The paper presents a leveled ontology of AGI, proposing that current frontier language models, such as ChatGPT and Bard, exhibit Competent performance for some tasks but are still at Emerging levels for others.
- The paper suggests documenting the performance levels of frontier AI models in model cards to help stakeholders understand the likely uneven performance of systems progressing toward AGI.
- The paper highlights the importance of metacognitive capabilities, such as learning new tasks and knowing when to ask for help, as key prerequisites for achieving generality in AGI systems.
- The paper proposes a taxonomy that rates systems according to their performance and generality, noting that generality requires a broad suite of cognitive and metacognitive tasks, and emphasizing the importance of including a wide range of perspectives in benchmark development.
- The paper introduces Levels of Autonomy, correlated with Levels of AGI, to characterize human-AI interaction paradigms and provide nuanced insights into risks associated with AI systems.
- The paper emphasizes the need for a living AGI benchmark that includes diverse and challenging tasks, and suggests that the ability to perform a requisite set of tasks at a given performance level should be sufficient for declaring a system as AGI.
- The paper concludes by discussing how the proposed framework can reshape discussions around the risks associated with AGI, highlighting the importance of investing in human-AI interaction research alongside model improvements【8:0†source】.
Cited By
Quotes
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.
References
;
| Author | volume | Date Value | title | type | journal | titleUrl | doi | note | year | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2023 LevelsofAGIOperationalizingProg | Shane Legg Meredith Ringel Morris Allan Dafoe Noah Fiedel Jascha Sohl-dickstein Tris Warkentin Aleksandra Faust Clement Farabet | Levels of AGI: Operationalizing Progress on the Path to AGI | 10.48550/arXiv.2311.02462 | 2023 |