Combining foundation model outputs with weak supervision yields faster model development and requires fewer ground truth labels.