Implicatio Research

Applied complexity science, building on Rashevsky, Rosen and Louie.

Structural control for AI systems that model themselves.

AI systems increasingly carry models of themselves and their surroundings, and act on what those models predict. We ask what can be shown about such a system from the outside, by its structure rather than its training: that its parts do the jobs they are named for, that every route to a harmful action passes through a stop, and that it holds an aim it cannot rewrite.

The mathematics for this exists. Robert Rosen worked out the theory of anticipatory systems, systems that act on a model of themselves, in 1985, building on Nicolas Rashevsky's relational biology; A. H. Louie has carried it furthest. We apply it to engineered learning systems, test it on small purpose-built systems, and publish the checks with their results, negative ones included.

How results are shared

Work that would make a system better at improving itself, without a check shown to hold it, is not published.

People

Mark Hampton

Founder and scientific director. Author of Seeing Anticipatory Systems (SpringerBriefs in Philosophy, forthcoming).

Scientific advisory board

Roberto Poli

UNESCO Chair in Anticipatory Systems, University of Trento.