Contingency (Smolin Reading)
Contingency is the principle that the specific forms produced by evolutionary processes depend on chains of accidents that cannot be re-run. Gould famously argued that if the tape of life were replayed from the Cambrian, the specific species that eventually emerged would be different. The general direction — toward increasing complexity — might hold, because the physics favors it. But the specific configurations — the particular organisms, the particular ecosystems, the particular events — depend on contingent sequences that no law determines. Smolin's framework extends this principle to cosmology and to technology. The arrow of complexity is a feature of the physics; the specific forms the arrow produces are contingent on the unrepeatable sequences of events that actualize one possibility rather than another.
Convergence Under Selection Pressure — Contrarian ^ Opus
There is a parallel reading that begins not from Gould's Burgess Shale but from Conway Morris's counter-argument: evolutionary outcomes converge far more than contingency implies. Eyes evolved independently dozens of times. Flight evolved independently in insects, birds, bats, and pterosaurs. The camera eye structure appears in organisms separated by hundreds of millions of years. What looks contingent at the level of individual lineages reveals profound constraint at the level of functional solutions—the state space of viable adaptations is far smaller than contingency assumes.
For AI, this convergence argument carries weight the Smolin reading underestimates. Transformer architectures did not emerge from a single contingent decision—they represent a convergent solution to the computational problem of capturing long-range dependencies in sequence data. Attention mechanisms were being independently explored across multiple research groups before "Attention Is All You Need" crystallized the approach. The dominance of large language models reflects not corporate contingency but the discovery that scale + self-supervision reliably produces capabilities—a functional attractor the economics of compute inevitably finds. The specific companies leading may be contingent, but the concentration dynamics are not: the capital requirements, the talent concentration, the infrastructure demands all push toward oligopoly. What appears as contingent choice often masks deeper structural constraints. The physics may not determine the path, but the economics and the mathematics constrain it far more than the contingency frame acknowledges.
In the AI Story
Gould developed the contingency argument in Wonderful Life (1989), using the fossils of the Burgess Shale as his primary case. The argument was directed against two positions: the adaptationist view that natural selection produces predictable outcomes given environmental conditions, and the progressivist view that evolution tends toward particular endpoints (typically toward humans or something human-like). Gould's point was that both views overstate what selection determines. Selection operates on variation, but the variation available at any moment depends on which organisms happen to exist, which in turn depends on the chains of accidents that produced the prior state.
Smolin's framework is compatible with and extends Gould's. The arrow of complexity is a feature of universes whose constants were selected for complexity-production. The constants produce conditions in which self-organizing systems can form and sustain themselves. But the specific self-organizing systems that actually form, and the specific trajectories they follow, depend on contingent sequences in the thick present. The physics determines the general direction; contingency determines the specific path.
For the AI discourse, the principle has immediate implications. The emergence of AI is not an accident — the arrow of complexity favors the development of systems capable of sophisticated information processing, and human civilization reached the technological threshold at which such systems could be built. But the specific forms AI takes are contingent. The dominance of transformer architectures. The centrality of large language models trained on internet text. The specific companies that dominate. The specific regulatory frameworks that emerge. All of these are contingent features — they depend on sequences of choices that could have gone differently, and that will shape what AI becomes in ways the physics does not determine.
This matters because it complicates both the triumphalist and the catastrophist readings of the AI transition. The triumphalist treats the current trajectory as inevitable and the specific outcomes as predetermined consequences of the technology itself. The catastrophist treats the same trajectory as inevitable and focuses on the harms it will produce. Both share the assumption that the trajectory is fixed. Contingency denies this. The trajectory is real, but it is being constructed, not discovered. The specific forms AI takes are being shaped by the specific precedents being established now — and those precedents are contingent choices that could have been different and could still be different.
Origin
The contingency principle is most associated with Stephen Jay Gould's work in evolutionary biology, particularly Wonderful Life (1989). Smolin's extension to cosmology is developed implicitly throughout his work and explicitly in his collaborations with Kauffman. Its application to technology — the argument that the specific forms of AI are contingent rather than predetermined — is developed here for the first time with this degree of specificity.
Key Ideas
Direction without destination. Physics determines the general direction toward complexity; contingency determines the specific forms that complexity takes.
Unrepeatable sequences. Specific outcomes depend on chains of accidents that cannot be re-run, even if the underlying dynamics are understood.
Against determinism. Contingency denies that the specific forms produced by evolutionary processes are predetermined consequences of the initial conditions.
Applies to AI. The specific forms AI takes — architectures, dominant companies, regulatory frameworks — are contingent rather than predetermined.
Enables responsibility. If the specific path is contingent, then choices about which path to build matter in ways that pure determinism would deny.
Appears in the Orange Pill Cycle
Scale-Dependent Contingency Gradients — Arbitrator ^ Opus
The right weighting depends on the temporal and structural scale being examined. At the level of fundamental capabilities—whether AI develops sophisticated pattern recognition, whether it achieves fluid language generation—convergence dominates (80/20 toward Conway Morris). The mathematics of learning and the economics of compute create powerful attractors. Multiple paths lead to similar functional outcomes because the constraint space is tight. At this scale, Smolin's physics-determines-direction holds strongly, but so does the convergence critique.
At the level of specific implementations and institutional arrangements, contingency reasserts itself (60/40 toward Gould). That OpenAI rather than DeepMind released the product that catalyzed mass adoption was contingent. That regulatory frameworks emerged in Europe before America reflects contingent political sequences. That transformer architectures became dominant before alternative approaches were fully explored represents a contingent path-dependency—early success created momentum that foreclosed exploration of the adjacent possible. These outcomes matter enormously for who benefits, who is harmed, and what forms of AI become infrastructural.
The synthetic frame is scale-dependent contingency: tight constraint at the level of fundamental capabilities, widening degrees of freedom as you move toward implementation details and institutional forms. This matters because it locates agency precisely. You cannot choose whether AI develops language capabilities—that attractor is too strong. But you can shape which architectures receive investment, which companies dominate, which regulatory frameworks constrain deployment. Contingency does not mean everything is open; it means the right things are open at the right scales. Responsibility operates in the widening space between what the physics determines and what human choices construct.
Further reading
- Stephen Jay Gould, Wonderful Life (Norton, 1989)
- Simon Conway Morris, Life's Solution (Cambridge, 2003)
- Lee Smolin, Time Reborn (Houghton Mifflin Harcourt, 2013)