Gould's contingency thesis, articulated most fully in Wonderful Life (1989), proposes that the specific outcomes of evolutionary history are not inevitable expressions of natural law but products of contingent sequences of events. His famous thought experiment: replay the tape of life from the Burgess Shale (530 million years ago), and the result would be utterly different—humans would almost certainly not evolve, mammals might not, vertebrates might not. The specific lineages dominating the modern Earth are products of specific accidents: the survival of Pikaia (modest chordate ancestor), the K-T asteroid clearing ecological space for mammals, Pliocene climate favoring bipedalism in African apes. Remove any contingency and downstream consequences cascade through subsequent history. Applied to technology, the thesis reveals that the transformer architecture, large language models, the December 2025 threshold—the entire AI landscape of 2025—is contingent on specific decisions (Vaswani team), institutions (Google Brain), economic conditions (GPU cost curves), and timing (when computation, data, architecture converged). Different contingencies would produce different AI with different capabilities, limitations, and social consequences.
The Burgess Shale fauna—preserved in exceptional detail in Canadian Rockies deposits—revealed the Cambrian explosion's extravagant diversity of body plans, most of which went extinct. Gould's analysis emphasized that survival was not determined by intrinsic superiority but by contingent fit to subsequent selection pressures. Pikaia, the unimpressive worm-like chordate, survived events that eliminated dozens of more complex, more specialized contemporaries. Had it gone extinct—a perfectly plausible outcome—the vertebrate lineage would not exist. The entire history of complex terrestrial life pivots on one organism's contingent survival.
The neural network winter provides AI's clearest punctuated equilibrium case and its deepest contingency demonstration. The mathematics of backpropagation was ready by 1970 (Linnainmaa), refined by 1974 (Werbos), published prominently by 1986 (Rumelhart et al.). The hardware was not ready, and funding—shaped by DARPA priorities, Cold War politics, and the misreading of Minsky and Papert's Perceptrons—dried up. Twenty years of potential development was lost. Had the winter not occurred, or had it ended a decade earlier, the AI landscape of 2025 would be unrecognizably different—three decades more mature, with fundamentally different capabilities.
The transformer architecture's contingency is equally sharp. The 2017 'Attention Is All You Need' paper by Vaswani and colleagues at Google Brain introduced self-attention mechanisms that replaced recurrent networks. But the specific path was paved with accidents: the team's composition, their institutional resources, the decision to dispense with recurrence entirely (not foregone even within the team), the availability of internet-scale text corpora, GPU cost curves shaped by the video game industry. Replay the tape—imagine the team disbands in 2016, or attention is combined with recurrence, or computational economics differ modestly—and the architecture dominating 2025 does not exist.
Gould's contingency thesis does not deny patterns exist—rivers flow downhill, evolution produces complexity (as statistical artifact), technology expands capability. But specific channels, body plans, architectures are determined by contingencies, not by general tendencies. The river provides momentum; the landscape shapes the channel. The specific AI we have is not the only AI we could have had. The future it produces depends on choices that inevitability myths tell us don't matter but that contingency demonstrates matter more than anything.
Contingency became Gould's organizing theme across his mature work, most fully articulated in Wonderful Life: The Burgess Shale and the Nature of History (1989). The book combined paleontological analysis with philosophical argument: the Burgess fauna demonstrated that the Cambrian explosion produced far more diverse body plans than currently exist, most went extinct for contingent rather than adaptive reasons, and replay would produce different survivors. The thesis was controversial—Simon Conway Morris and others argued convergence constrains outcomes more than Gould allowed—but the argument reshaped how the history of life is understood, emphasizing the role of accident over inevitability.
Replay the tape, get a different world. The specific outcome of any branching contingent process depends on unrepeatable event sequences—alternative sequences produce alternative outcomes.
Survival is not superiority. Organisms dominating current distribution are there not because the process aimed at them but because they survived specific contingent selection events.
The transformer's contingent ancestry. Every feature of current AI—architecture, capabilities, limitations—traces to specific decisions that might have been otherwise.
The myth of inevitability suppresses agency. If trajectory is determined, choices are decorative; if contingent, choices are constitutive of the future that emerges.
Uncertainty is signature of genuine transition. Not knowing which future will materialize is the epistemological condition of standing inside a branching event where multiple outcomes are possible.