Why the term 'AGI' is now useless for describing AI progress

According to a recent essay by Noosphere89 on LessWrong, the concept of 'AGI' has become nearly useless for meaningful discussion about AI capabilities. The author points out that AI progress has turned out to be highly 'jagged'—meaning that systems can exceed some definitions of general intelligence while completely failing others. For example, today's best models can perform real economic work (the author pegs this milestone at November 2025), yet they are not nearly as sample-efficient as humans, lack recurrence, and rely on chain-of-thought reasoning. As a result, different people use 'AGI' to mean vastly different things, from a system that can automate any job to one that achieves human-level learning efficiency. The fuzziness that was once harmless now creates confusion, with some claiming AGI has already arrived and others insisting we are far from it.

The essay traces the historical utility of the term, which was coined in the early 2000s to contrast with narrow AI systems. Back then, gesturing toward a general-purpose future AI was useful because we were far from that threshold. However, as capabilities improved, the smooth progression assumed by earlier definitions (e.g., mouse-to-chimp-to-human) never materialized. Instead, AIs excel at language but are sample-hungry, and reasoning is achieved through explicit chain-of-thought rather than intuitive neuralese. The author concludes that we now need more specific terminology to discuss phase changes and capability ceilings—for instance, distinguishing between 'systems that can automate most remote work' and 'systems that learn as efficiently as humans.' Until we adopt finer-grained terms, debates about AGI will remain unproductive.