The Burgess Shale, a fossil deposit in the Canadian Rockies, preserves the Cambrian explosion's extravagant diversity in exceptional detail. Discovered in 1909 by Charles Walcott, re-examined by Harry Whittington's team in the 1970s, and interpreted by Gould in Wonderful Life (1989), the Shale revealed that the Cambrian seas contained far more diverse body plans than currently exist—Opabinia with five eyes, Hallucigenia walking on spines, Anomalocaris as apex predator. Most went extinct. Their extinction was not determined by inferiority but by contingent fit to subsequent conditions. Pikaia, the modest chordate ancestor of all vertebrates, was one of the least impressive Cambrian organisms—it survived events that eliminated more complex, more specialized forms. Applied to AI, the metaphor identifies the present moment (2025–2026) as a Cambrian explosion: rapid diversification of forms (model architectures, training approaches, deployment strategies), most of which will go extinct. The survivors will not be the 'best' in any objective sense but the ones fitting specific ecological conditions produced by specific choices (regulatory, economic, cultural) that specific humans happen to make. The branches being pruned now represent genuine alternatives to forms that will dominate—capabilities not developed, problems not solved, futures not realized.
Gould's Wonderful Life used the Burgess Shale to argue for radical contingency in evolutionary history. The book's central thesis: if you could replay the tape of life from the Cambrian explosion, the result would be utterly different. The specific body plans that survived to produce modern fauna were not inevitable—they were lucky. Pikaia might have gone extinct (perfectly plausible given Cambrian body counts), and without it the vertebrate lineage would not exist. No fish, amphibians, reptiles, mammals, primates, humans. The entire history of complex terrestrial life pivots on one unimpressive organism's contingent survival in one ancient ocean.
The neural network winter functions as AI's first major extinction event—approaches that were viable (symbolic AI, expert systems, LISP machines) went extinct not from intrinsic inferiority but from specific contingent conditions: funding priorities, hardware economics, institutional dynamics, personality conflicts between research camps. Each represented a genuine alternative body plan. Their absence from the 2025 landscape is not evidence they were wrong but evidence the selection environment favored statistical learning over symbolic reasoning. Replay the tape with different funding, and the survivors might be reversed.
The current moment exhibits characteristic Cambrian dynamics: transformer variants proliferating (attention mechanisms, mixture-of-experts, state-space models), training approaches diversifying (supervised, unsupervised, RLHF, constitutional AI), deployment strategies multiplying (API services, open-source, on-device, hybrid). Most will go extinct. The ones surviving will be determined not by technical superiority alone but by economic viability, regulatory environment, institutional adoption, and cultural fit—selection pressures analogous to ecological conditions pruning Cambrian fauna.
The Burgess metaphor's deepest implication: what is being lost now—the branches currently being pruned—may contain solutions to problems the surviving lineages cannot solve. Biological history is full of features going extinct in one lineage and being independently re-evolved in another because the problems they solved were real problems persisting across contexts. The same may be true of AI: symbolic reasoning approaches abandoned during the neural network winter may contain insights the connectionist paradigm cannot replicate. The LISP machines, the expert systems, the alternative architectures—each was a viable body plan whose capabilities the current ecosystem has not preserved and may someday need to re-evolve at enormous cost.
The Burgess Shale was discovered in 1909 by Charles Doolittle Walcott, Secretary of the Smithsonian Institution, who spent summers collecting fossils in the Canadian Rockies. Walcott recognized the fossils were Cambrian but interpreted them through the lens of existing taxonomies, shoehorning bizarre forms into familiar categories. In the 1970s, Harry Whittington at Cambridge and his students (Derek Briggs, Simon Conway Morris) re-examined Walcott's collections and recognized that the Burgess fauna were far stranger than Walcott had allowed—representing body plans with no modern descendants. Gould's Wonderful Life (1989) synthesized Whittington's work into a philosophical argument about contingency, making the Burgess Shale the most famous fossil deposit in popular science.
Cambrian as diversification explosion. The fossil record shows the most diverse body plans appearing in a geologically brief period—dozens of viable architectural solutions to the problem of being a multicellular animal.
Most Cambrian forms went extinct. Survival was not determined by superiority but by contingent fit to subsequent selection pressures—accidents of geology, climate, predation.
AI's Cambrian is now. 2025–2026 exhibits the same pattern—architectural proliferation, most forms destined for extinction, survivors determined by contingent rather than intrinsic factors.
Pruned branches represent lost alternatives. Each extinction forecloses a genuine possibility—capabilities not developed, problems not solved, futures not realized.
The survivors are not inevitable. Replay from the current moment with different choices (regulatory, economic, institutional) and different AI lineages dominate the 2035 landscape.