Short answer: Yes—this kind of interpretability result is exactly what we’d want if we’re trying to get mechanistic intuition about reptile phylogeny (why crocs are closer to birds than to lizards) without relying on naive “percent identity” diffs. A long‑context DNA model like Evo 2 appears to learn a geometry of life in its activations; distances along that manifold can reflect evolutionary relatedness even when morphology misleads. That geometry is plausibly driven by higher‑order signals—codon usage, regulatory grammar, repeat landscapes, synteny, and co‑evolving motifs—rather than only raw sequence similarity. Interpreting those signals lets you talk about “complexity multipliers”: small regulatory changes that propagate through gene‑regulatory networks and yield outsized morphological effects (think limb enhancers in snakes or modular enhancers in fish). (Goodfire AI)

Why birds + crocs, despite the look‑alike lizards?

• Phylogeny: Crocodilians and birds are the two surviving archosaur lineages; genomics strongly nests crocs with birds rather than with squamates (lizards/snakes). (PMC)
• Rates and modes: Crocodilian genomes have evolved exceptionally slowly (low substitution, TE turnover, synteny breakage), a pattern consistent with their morphological “stasis”; birds show lineage‑specific acceleration—an autapomorphy—relative to that slow archosaur background. This rate asymmetry helps explain why birds look so different while crocs still look “reptilian.” (PMC)
• Regulatory drivers: Many avian innovations are tied to regulatory DNA rather than coding changes. For example, thousands of avian‑specific highly conserved elements are almost entirely non‑coding and enriched for regulatory roles near developmental genes. That’s exactly the substrate where small edits can have large, system‑level consequences. (Nature)

This is the setting where a model’s internal geometry can outperform naive “diffs”: the model can pick up which bits of the genome carry phylogenetically informative, function‑shaping signals.

What Evo 2 is learning that helps (beyond raw similarity)

• A phylogeny‑shaped manifold: When you average activations over many random windows per species and measure geodesic distances along the induced manifold, those distances track branch length on ground‑truth trees—even after controlling for direct sequence similarity. (Goodfire’s case study demonstrates this clearly for bacteria; the same recipe should transfer to eukaryotes.) (Goodfire AI)
• Long‑range dependency modeling: Evo 2 ingests up to ~1 M nucleotides, so it can integrate motifs that are far apart—promoters, enhancers, architectural elements—picking up the grammar of regulation and genome organization, not just local k‑mer counts. (arcinstitute.org)
• Phylogenetic signals that aren’t simple diffs: Codon‑usage “signatures,” dinucleotide/codon‑pairing, and other compositional features carry deep phylogenetic signal across life and are detectable without alignments; LMs reliably pick these up. (PubMed, Oxford Academic, PMC)

Takeaway: The model’s notion of “closeness” can align crocs with birds because it aggregates many weak but functionally relevant cues (regulatory grammar, compositional signatures, synteny patterns), rather than just counting mismatches.

“Complexity multipliers” in biology (and what a model could see)

Your intuition maps well to regulatory hubs, modular enhancers, pleiotropy, and epistasis in gene‑regulatory networks (GRNs):

• Small changes, big effects: Deleting or altering a single enhancer can rewire a developmental program—e.g., repeated loss of pelvic spines in sticklebacks via deletion of a Pitx1 pelvic enhancer; identical proteins, different expression programs. (PubMed)
• Reptile‑specific example: Limb loss in snakes is associated with accumulated changes in the ZRS long‑range enhancer of Shh; swapping snake ZRS into mice truncates limbs—classic “multiplier” behavior via a regulatory switch. (PMC)
• Network architecture matters: Heritability and selection concentrate in tissue‑specific regulatory hubs and modules—exactly the places where small sequence changes can have amplified, coordinated phenotypic effects. (PMC)
• Regulatory conservation despite sequence drift: Even when enhancer sequence diverges across vertebrates, positional/syntenic conservation can preserve function—explaining why “raw diffs” miss what models (and assays) can infer from broader context. (Nature)