The progressive encapsulation of architectural knowledgewithin technical systems has led to increasing automation of designand decision-making processes. Long limited to the execution ofexplicit rules, this automation is now undergoing a profound transformationwith the rise of machine learning systems, characterised by self-regulation,probabilistic decision-making, and increasing opacity of theirinternal mechanisms. This article analyses this transformation in threestages. First, it revisits Shannon’s rigorous quantification of information,a prerequisite for its machine processing. Next, it examines the genesisof computability and the conjecture formulated at the Dartmouth conference,according to which intelligence and creativity can be simulatedby formal procedures incorporating controlled randomness. Finally, bycomparing Mario Carpo’s imitation thesis with recent results in high-dimensionalgeometry, particularly those of Yann LeCun, the article arguesthat contemporary AI cannot be described adequately as a simple machineof imitation. Under a strict geometric definition of interpolation,contemporary AI systems are better characterised as operating throughextrapolative generalization than through interpolation in the ordinarysense. This leads to a redefinition of the autonomy of the architecturalautomaton, understood less as the capacity to reproduce precedents thanas the tendency, in high-dimensional spaces, to generate outputs beyondthe strict envelope of available data.
Philippe Morel (Fri,) studied this question.