The holy grail of our field is to enable the complete computational design of any arbitrarily chosen biomolecular activity. To enable such template-free design of function, we still need to learn a lot more about how function to encode functions in proteins. We are therefore developing methods to design not a handful of binders or enzymes as in all current protein design methods, but vast repertoires comprising millions of substantially different variants. We use high-throughput screening methods to isolate the functional designs and deep sequencing analysis to fully characterise these designs. Next, advanced machine-learning methods are trained to find molecular features that discriminate the best designs from the rest, and these features are then used to improve the design algorithms, leading to a continuous, unbiased and systematic approach to learn the rules for designing new biomolecular activities. We are applying this strategy to the design of new hydrolytic enzymes, single-domain camelid antibodies, small-molecule binders, and a membrane transporter.