This keynote will show how AI is transforming power electronic converters from statically tuned devices into self-learning components that understand their own health and flexibility. It will start from model predictive control (MPC) for converters and drives, and explain how AI was first used as an offline assistant to automate MPC design and support reliability-aware design under mission profiles. It will then move to offline-trained AI running online in the loop, where neural networks emulate MPC and support grid-impedance identification and adaptive tuning of grid-connected converters.Building on this, the talk will present recent work on online learning inside converters through self-commissioned, edge-based diagnostics and, finally, online learning of latent variables such as process flexibility and health margins. These ideas will be illustrated through the DFF MAGIC project and the ERC Consolidator Grant ARTEFACT, which aim to quantify flexibility directly in motor drives and use it for reliable grid services. The keynote will conclude with lessons learned from industrial collaboration and the PHLIT spin-out on how these concepts connect to real-world converter-dominated energy systems.