Natural products built most of the pharmacopoeia, yet the discovery engine that produced penicillin, the statins and the majority of clinically used antibiotics largely stalled by the late 1990s. The reason was not that microbes had run out of chemistry. Bioactivity-guided fractionation kept returning the same abundant molecules, and the enormous biosynthetic capacity of uncultured organisms stayed out of reach. A compendium of roughly 170,000 bacterial genomes estimates that only about three percent of that biosynthetic potential has ever been characterised. This review argues that the field is being rebuilt around that gap by three technologies that matured separately and are now converging: genome mining, which reads biosynthetic gene clusters directly from sequence; synthetic biology, which expresses those clusters in tractable hosts; and machine learning, which prioritises both the clusters worth expressing and the molecules worth testing. Each is examined for what it has actually delivered, from antiSMASH-driven cluster detection and evolution-guided antibiotic discovery to the heterologous production of artemisinic acid and the deep-learning identification of halicin and a new structural class active against methicillin-resistant Staphylococcus aureus. The central claim is that the rate-limiting step has shifted from finding biosynthetic novelty to expressing and prioritising it, and that near-term returns depend on integrating the three approaches rather than advancing any one. The appraisal is deliberately uneven, because the evidence is: genome mining is mature, heterologous expression remains the chronic bottleneck, and artificial intelligence is powerful for property prediction but constrained by scarce, biased training data and a wide gap between an in silico hit and a validated drug.