Investigating how intelligence is formed and operates, where it lives, what it can do, and when to use it.
-
The question is deliberately about location. Most current discussion treats intelligence as a property of a model, and measures it accordingly, against benchmarks. But the performance that matters to an organisation is produced by an arrangement: an analyst, a set of tools, a body of institutional knowledge, a set of records, and a workflow that determines what any of them get to see.
Benchmark performance is a poor predictor of performance in that arrangement. This is not a controversial observation, but very little follows from it in practice, because there is no established way to measure what a particular arrangement can do in a particular context.
Related programmes: Neurosymbolic AI, Control Interfaces.
-
Benchmark performance does not transfer reliably to situated performance.
Capability is distributed. The relevant unit of analysis is usually larger than the model and larger than the individual.
Improving a component does not reliably improve the system containing it, and sometimes does the opposite.
-
Whether "intelligence" is a useful unit of analysis at all, as against capability in a stated context.
Where accountability sits when capability is genuinely distributed, and whether it can be assigned without distorting how the work is done.
Whether the introduction of capable tooling moves the locus, or merely changes its composition.
-
How to measure situated capability well enough to make a procurement or deployment decision.
Whether a shift in the locus is detectable from the record, or only visible in hindsight.
What actually improves the intelligence of an arrangement, as distinct from the performance of its parts.