Source: A Definition of AGI — Hendrycks et al., arXiv 2510.18212
Technical AI papers can look intimidating from the outside. Lots of architecture names, benchmark tables, and acronyms behaving like they pay rent. But the useful question is usually simple: what bottleneck is this trying to remove?
The AGI debate has always been stuck on the same problem: nobody agrees on what it means. Every time AI masters a task — chess, bar exam, image recognition — the definition shifts upward. It’s an endlessly moving goalpost, and it makes productive conversation nearly impossible. A paper from the Center for AI Safety and collaborators across MIT, Stanford, Oxford, and others takes a serious crack at fixing this. They ground their definition in Cattell-Horn-Carroll CHC theory — the most empirically validated framework for human cognition, built over a century of psychometric research. AGI, by their definition, means matching the cognitive versatility and proficiency of a well-educated adult across ten core domains: reasoning, memory, language, perception, and six others. The results are humbling in both directions. That’s extraordinary progress — more than doubling in one model generation. But the profile is “jagged”: the models score very high on knowledge retrieval and language tasks, and very low on long-term memory storage and certain kinds of grounded perception. The long-term memory finding is the one I keep returning to. It’s not a small gap — it’s a fundamental architectural limitation. Current models have no persistent memory across contexts unless explicitly engineered.
In plain English, that is why the result matters beyond the chart. It changes where people should look, what they should question, and which comfortable assumption probably needs to be retired.
My takeaway: architecture matters because it decides what kind of intelligence is affordable, reliable, and usable outside the lab. The benchmark is the beginning of the story, not the ending.