Lost Libraries: Thylacine Extinction and Irretrievable Information

I imagine my Library of Babel—containing every possible book arranged in hexagonal galleries extending infinitely in all directions. Each volume differs from its neighbors by a single character, yet finding any specific text proves nearly impossible. The library contains all wisdom and all nonsense in equal measure, rendering infinite information functionally equivalent to none at all. The last thylacine died September 7, 1936, in Hobart Zoo, carrying in its cells a different kind of infinite text: a genome shaped by 100 million years of divergence from placental carnivores, yet converging through natural selection toward the same predatory form as wolves.

The Marginalia We Cannot Recover

Geneticists propose resurrecting the thylacine through CRISPR, editing the numbat genome—its closest living relative—to approximate the extinct species. Museum specimens preserve DNA sequences like ancient manuscripts in climate-controlled vaults. We possess the text. But convergent evolution reveals that identical morphology emerges from vastly different developmental contexts: the thylacine’s wolf-like skull and digitigrade gait arose through marsupial pathways fundamentally distinct from canid evolution. The genome alone cannot specify the complete animal—it requires epigenetic factors, maternal developmental cues, behavioral transmission across generations.

When the last thylacine died, we lost not merely a sequence of base pairs but the lived context of that sequence. The final speaker of a dying language carries idioms and intonations no grammar book captures. Entropy, as Shannon defined it, measures irreversible information loss. A book burned to ash might theoretically be reconstructed from molecular fragments, but the reader’s marginal notes, the binding’s wear patterns, the page corners folded by previous owners—these vanish absolutely. The thylacine carried similar marginalia: how it raised young in its backward-facing pouch, which prey species it preferred in Tasmania’s remote forests, the social structures governing its hunts. The genome is our printed text; the phenotype was its interpretation, now erased.

Transformations That Destroy Meaning

Neural networks transform inputs through learned geometric representations—mapping raw coordinates through increasingly abstract spaces until patterns become linearly separable. But these transformations are lossy. A representation space, once folded through ReLU activation functions, cannot perfectly invert to recover the original state. Information compresses, discards dimensions, emphasizes task-relevant features while abandoning others. Gradient descent optimizes differentiable surfaces but cannot navigate discontinuous jumps; evolution explores diverse forms but inefficiently, often converging to local optima rather than preserving variety.

The thylacine’s extinction represents a similar lossy transformation. The ecosystem has moved on—Tasmania’s ecological niche shifted in the decades since 1936, making uncertain whether any resurrected thylacine-like creature could reoccupy its ancestral role. We might reconstruct the genome, perhaps even generate a breathing animal, but would this be the thylacine or merely a thylacine-shaped numbat? The species is not solely its DNA but the entire system: genome plus development plus behavior plus ecological context. Some information cannot be recovered because the substrate that preserved it—the living population—has undergone thermodynamic irreversibility.

Extinction resembles burning the only copy of a unique volume in an infinite library: theoretically one among uncountable texts, but practically irreplaceable once lost. Convergent evolution hints that natural selection might recreate similar forms given sufficient time and similar pressures. But time flows forward. The labyrinth has no exit that leads backward to what we have already lost.