tl;dr: Princeton researchers used reinforcement learning combined with inverse design (via a CNN-based EM emulator) to generate RFIC designs from scratch, bypassing the human-crafted templates that have long made RF design an "artisanal" bottleneck. The resulting chips—often bizarre, QR-code-like layouts—set records for bandwidth and efficiency in millimeter-wave power amplifiers, and a diffusion-model extension lets designers dial in more human-interpretable structures. The main remaining obstacle is data: progress toward a foundational model for RF/analog design is gated by proprietary simulation data locked behind NDAs.