Each point is one of 2,395 bacterial species, placed by Evo2-7B — a model trained only to predict the next letter of DNA, never taught biology. Left is the model’s own internal space (layer-24 embeddings, flattened to three dimensions); right is the established GTDB tree of life for the same species. Pick a taxonomic rank and watch the groups the model invented fall into the same arrangement as the real tree it was never shown.

The model’s space — Evo2-7B, layer-24 embeddings (rotatable)

The real tree of life — GTDB, branches colored by class

Source: Evo2-7B (Arc Institute), layer-24 residual-stream embeddings, mean-pooled per species; 2,395 species across 16 phyla; 3D layout via UMAP (n_neighbors 150, min_dist 0.66). Ground-truth tree: GTDB r220. The 3D view is a flattened picture for the eye; the distance correlations in the recipe test the arrangement in the model’s full space.