Diverse Adjacency

In the AI Story

The significance of diverse adjacency becomes clear through contrast with its absence. If every position on a network were adjacent to the same set of alternative phenotypes, then dispersal would produce no advantage: one organism in one place would have access to the same innovation menu as a million organisms dispersed across the network. But because different positions are adjacent to different subsets of the phenotype space, the collective coverage of the population scales with its dispersal — more dispersal means more distinct adjacencies, more adjacencies means more accessible innovations.

Wagner demonstrated diverse adjacency empirically across multiple biological systems. In metabolic networks, organisms that accumulated neutral mutations found themselves adjacent to different novel metabolic capabilities at each position. In protein sequence space, different sequences that fold into the same three-dimensional shape had access to different alternative folds through single amino acid changes. The same functional phenotype, viewed from different genotypic positions, sees different possibility neighborhoods.

The computational analog is structural. Neural networks in different regions of flat minima that produce equivalent performance on training data are adjacent to different capabilities through parameter perturbations. Two models with the same benchmark scores may have different sensitivities to specific inputs, different failure modes, different patterns of generalization — precisely because they occupy different positions on the network of computationally equivalent configurations. The diversity of adjacency is why two ostensibly 'equivalent' models can produce qualitatively different outputs under novel conditions.

The cultural analog extends the principle. A developer in Lagos, an engineer in Mountain View, and a researcher in Shenzhen occupy different positions in the space of technological possibilities — different problems, different constraints, different cultural contexts — and are adjacent to different innovations. The democratization of AI tools is significant in Wagner's framework not only because it increases the size of the exploring population but because it increases the diversity of positions occupied, which translates directly into diversity of accessible innovations.

Origin

Wagner's empirical demonstration of diverse adjacency across biological systems was developed in parallel with the broader genotype network framework, culminating in the synthesis presented in The Origins of Evolutionary Innovations (2011) and Arrival of the Fittest (2014). The conceptual precursor appears in Stuart Kauffman's notion of the 'adjacent possible,' though Wagner provided the mathematical specificity and empirical validation that Kauffman's more metaphorical formulation lacked.

Key Ideas

Adjacency is position-specific. Different nodes on a genotype network are adjacent to different alternative phenotypes — the 'one step away' menu varies across the network.

Dispersal expands the menu. As a population spreads through the network, its members collectively occupy more diverse positions with more diverse adjacencies, expanding access to innovation.

Equivalent configurations are not interchangeable. Two models or organisms with the same current performance may have profoundly different innovation neighborhoods.

Diversity of explorers matters beyond numbers. The variety of positions occupied, not only the count of explorers, determines the diversity of accessible innovations.

Culture inherits the logic. Different cultural positions — geographic, disciplinary, demographic — provide different adjacencies; diversity of exploration is adaptive advantage.