Can we produce a definition of AGI that is constitutive rather than behavioural?
Status: open, 11 May 2026
Problem statement:
Can we produce a definition of AGI that is constitutive rather than behavioural?
What is AGI, such that we'd know whether we'd built one? The current frame treats AGI as a benchmark target — pass enough tests, demonstrate enough capabilities, and the label applies. This is the wrong shape of answer because it makes AGI a moving threshold rather than a category. Every time systems pass an old benchmark, the benchmark gets revised, which means the benchmarks weren't tracking what AGI actually is; they were tracking what we hadn't built yet. The problem is that the field is operating without a constitutive definition, only with provisional behavioural proxies, and the proxies keep failing in the same direction.
Conjecture: AGI is a Universal Explainer