Algorithmic Intelligence

Algorithmic Intelligence grounds AI in what it actually is instead of a fake attempt at intelligence. What is the algorithmic part? A well-architected pipeline of data structures, loss functions, linear algebra, and conditional state machines. The useful contrast is not artificial against real, but algorithmic against biological.

Regardless of its biological or algorithmic source, an intelligence has to make verifiable statements against reality. A biological one does this continuously and cheaply: you judge that the branch will hold your weight, you step onto it, and either you fall with pain or you sit on the branch.

For algorithmic intelligence this is much harder, because the training happens over data that was created or curated by people or other machine pipelines. It lacks a hard connection to the physical consequences that matter. A model that performs well on the data it was shown has demonstrated prediction on that dataset; it has not proven anything against the physical reality. What is needed instead is provable method tested against a directly sensed reality, and models composed from the relationships between objects rather than fitted whole.

That is what moved me from building these systems to studying the mathematics beneath them: discrete mathematics and network theory, and from there to complex adaptive systems, swarms, and criticality. Systems of many interacting parts are not described by the average behavior of the whole. They have thresholds, cascades, and regimes — and those are exactly the properties an intelligent system has to be held to, if it is going to be trusted to act.

Writing on this is collected in The Confluence.