When constraints have fixed geometry, solutions converge regardless of solver ancestry. Four convergences across four substrates reveal that durable advantage belongs to whoever defines the constraint, not whoever solves it first.
A paper in Science Advances in April 2026 reported that foxglove plants independently evolved a mammalian-like steroid biosynthesis pathway. Researchers at Northeastern University found that foxgloves convert cholesterol through pregnenolone and progesterone to produce cardenolides — cardiac glycosides that have been used in medicine for centuries. At least seventeen plant orders produce cardenolides via this route. Plants and mammals diverged over a billion years ago, yet foxgloves built the same biochemical machinery that mammals use for sex hormones. Not a shared toolkit repurposed. The same solution, discovered independently, in kingdoms separated by evolutionary distance measured in geological eras.
The mechanism is instructive. Cholesterol sits at the start of both pathways. Its molecular geometry constrains what chemical transformations are efficient. Cytochrome P450 enzymes catalyze each step — recruited from existing metabolic toolkits, different toolkits in plants and mammals, but the same reactions at the active site. The constraint is the substrate's shape. The shape dictates the pathway.
The Pattern
Four unrelated fish lineages independently evolved Type I antifreeze proteins from four different genetic progenitors. The resulting proteins share up to eighty-three percent amino acid sequence identity. All four converge on the same architecture: alanine-rich single alpha helices that bind to ice crystal surfaces. In winter flounder, the protein came from one gene. In grubby sculpin, a frameshift mutation in an entirely different housekeeping gene produced a near-identical protein. The ice crystal lattice is the constraint. Its geometry dictates which molecular shapes can bind to it. Four separate evolutionary experiments, run over millions of years, arrived at the same answer.
Vertebrates and cephalopods evolved camera-type eyes independently across more than five hundred million years of separation. Both converge on the same optical architecture: adjustable pupil, refractive lens, photoreceptor retina. Over fifteen hundred genes are commonly expressed in both eye types. The master regulatory gene Pax6 is conserved across the divide. But where the physics stops constraining, the implementations diverge. Cephalopods use rhabdomeric photoreceptors — finger-like microvilli that detect polarized light. Vertebrates use ciliary photoreceptors with an inverted retina that creates a blind spot. Lens crystallins were recruited from different protein families entirely. The physics of image formation constrains the solution. Everything else is free to vary.
C4 photosynthesis has evolved independently at least sixty-two times across flowering plants. Every origin converges on the same core innovation: a carbon dioxide concentrating pump that uses PEP carboxylase in mesophyll cells to fix carbon before shuttling it to bundle sheath cells for RuBisCO. The universal steps are identical across all sixty-two lineages. The variable steps — which decarboxylation enzyme each lineage uses — diverge. The constraint is RuBisCO's oxygenase side reaction, which wastes energy at low carbon dioxide concentrations. C3 plants already express all the necessary enzymes at low levels. The sixty-two origins are sixty-two lineages that upregulated and compartmentalized existing components because the constraint made the same solution discoverable from many starting points.
The Geometry of Constraint
Four convergences across four substrates — biochemistry, protein structure, optics, metabolic chemistry — share one structural property. When the constraint has precise geometry, the space of viable solutions collapses. The ice crystal lattice admits few molecular shapes. The physics of refraction permits few optical architectures. RuBisCO's inefficiency has one efficient biochemical workaround. Cholesterol's ring structure permits one class of downstream modifications.
Convergence is not coincidence and not common ancestry. It is the constraint's signature. The solver's identity — fish species, plant lineage, mammalian or botanical kingdom — is irrelevant. What matters is the geometry of the problem. When that geometry is tight enough, the solution is not invented but discovered, inevitably, by any system that confronts the same constraint long enough.
What Convergence Identifies
The pattern identifies where durable advantage lives — and where it cannot. In technology, tightly constrained problems converge on identical solutions regardless of which company arrives first. Semiconductor lithography has fixed physics: extreme ultraviolet wavelength, numerical aperture, resist chemistry. TSMC's dominance reflects control of the constraint itself, not a proprietary solution within it. Whoever defines the constraint captures the ecosystem.
Whoever competes on solutions within a tightly constrained space faces convergence. Rivals will arrive at near-identical architectures because the problem leaves no room for differentiation. Frontier AI models are running the C4 experiment — multiple labs converging on the same transformer architecture because the constraint of next-token prediction on internet-scale data has tight geometry. Long the constraint-definers: TSMC in lithography, Constellation Energy in AI baseload power, ASML in the tooling that makes lithography possible. Short differentiation narratives in constrained spaces. When four fish build eighty-three percent identical proteins from four different genes, the gene does not matter. The ice does.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)