A new benchmark built specifically to test how well AI models understand aging biology found that a compact, purpose-built model beat every frontier system thrown at it, including the large general-purpose models that usually top every leaderboard. The test wasn't about writing fluent sentences about telomeres. It scored models on tasks like predicting how a cell's molecular markers shift as it ages and matching interventions to the biological pathway they actually affect, the kind of structured reasoning where a model has to know the mechanism, not just describe it. The frontier systems, trained on a slice of the entire internet, did worse than a model trained narrowly on biological data.

The reason is not mysterious once you separate what "big" buys you from what a benchmark like this actually tests. A frontier model's size helps it predict plausible-sounding text across every domain at once; it does not force the model to hold a consistent internal model of, say, how NAD+ depletion cascades into mitochondrial dysfunction. A compact model trained only on aging biology datasets can afford to get that one cascade right, because it never has to also be good at writing poetry or debugging code. This is the same lesson protein-folding models taught the field years ago: a narrow model with the right training data beats a general one on the task it was built for, every time. The next gate is whether this compact model's edge holds up against real experimental validation, meaning wet-lab tests of the interventions it flags, not just its score on a written benchmark.