The more I dig into Goodfire Research, the more I realize how the branch of interpretability is quietly turning into one of the most interesting frontiers in AI (and especially AI for Science) They published this research last year with core idea of how a DNA foundation model internally organizes species in its embedding space in a way that mirrors the real evolutionary tree of life. or basically how the model rediscovered phylogeny purely from DNA sequences. they studied Evo 2 (DNA model developed by EvolutionaryScale) and discovered that: + each species’ genome gets mapped to a vector embedding inside the model. + these embeddings form a curved geometric structure (a manifold). + distances along this manifold correspond to actual evolutionary distance between species. so inside the model: similar species → close embeddings distant species → far embeddings and the structure that emerges is essentially the tree of life. This might proves something groundbreaking how foundation models may rediscover scientific structures automatically.