The Nature of Technology: What It Is and How It Evolves
Source: https://www.simonandschuster.com/books/The-Nature-of-Technology/W-Brian-Arthur/9781416544067 ↗
Technologies are not invented from scratch; they evolve by combining with one another.
Every technology is an assemblage of earlier technologies, and innovations arise from recombinations, not isolated inspirations.
Arthur offers a framework for understanding why AI is not a one-off invention but a layer that combines with everything else — and why its effects are unpredictable and emergent, not plannable.
Read after Coase and before trying to forecast anything about AI's impact on your organisation.
Central argument
Arthur argues that technology is not a collection of discrete inventions but a self-referential system that evolves through combination: every technology is built from existing technologies, and new technologies emerge when existing ones are recruited and assembled in novel configurations. He calls this process 'combinatorial evolution,' and grounds it in a theory where technologies exploit phenomena — natural effects — as their core mechanisms. The deeper claim is that the economy itself is an expression of its technologies, meaning that as technologies recombine, they continuously restructure the economic activities and institutions built on top of them.
Critique
Arthur's framework is compelling as a descriptive account of how technologies evolve in retrospect, but it offers limited predictive traction: knowing that innovation is combinatorial does not tell you which combinations will prove generative, when, or why some recombinations succeed while structurally similar ones fail. The theory also underweights the role of social, regulatory, and institutional forces in shaping which combinations get adopted — treating the evolution of technology as somewhat endogenous and self-propelling risks underestimating how profoundly non-technical factors determine outcomes.
Why it matters for product
For a CPO, Arthur's combinatorial lens reframes the product strategy question: instead of asking 'what should we build,' it prompts asking 'what existing capabilities, data layers, and integrations can we recruit and assemble in a configuration competitors haven't tried.' It also explains why AI integration resists clean roadmapping — if AI is a general-purpose layer that recombines with everything, its effects on your product surface will emerge iteratively from use rather than from planning, which has direct implications for how you structure discovery and how skeptically you should treat any long-horizon AI impact forecast.