Protein engineering has enormous academic and industrial potential. However, it is limited by the lack of experimental assays that are consistent with the design goal and sufficiently high-throughput to find rare, enhanced variants. Here we introduce a machine learning-guided paradigm that can use as few as 24 functionally assayed mutant sequences to build an accurate virtual fitness landscape and screen ten million sequences via in silico directed evolution. As demonstrated in two highly dissimilar proteins, avGFP and TEM-1 β-lactamase, top candidates from a single round are diverse and as active as engineered mutants obtained from previous multi-year, high-throughput efforts. Because it distills information from both global and local sequence landscapes, our model approximates protein function even before receiving experimental data, and generalizes from only single mutations to propose high-functioning epistatically non-trivial designs. With reproducible >500% improvements in activity from a single assay in a 96-well plate, we demonstrate the strongest generalization observed in machine-learning guided protein design to date. Taken together, our approach enables efficient use of resource intensive high-fidelity assays without sacrificing throughput. By encouraging alignment with endpoint objectives, low-N design will accelerate engineered proteins into the fermenter, field, and clinic.