This paper proposes a new framework for understanding and classifying Artificial General Intelligence (AGI) by introducing levels of AGI based on performance and generality. The authors analyze existing AGI definitions, establishing six key principles for a useful ontology, emphasizing capabilities over processes and the importance of ecological validity in benchmarks. Their leveled system aims to provide a common language for comparing AI models, assessing risks, and measuring progress towards AGI, also considering the interplay between these levels and autonomy in deployment scenarios. Ultimately, the work advocates for a more nuanced and operationalizable approach to defining and discussing the path to AGI.

Sources: https://arxiv.org/abs/2311.02462