AI-designed radio chips look like QR codes and beat human ones

Radio-frequency chips — the parts in every phone, car radar, and satellite link that send and receive signals — have stayed a hand-crafted "dark art," mastered over years, even as CPUs and GPUs got designed by algorithm. Writing in IEEE Spectrum, Princeton's Kaushik Sengupta describes teaching AI to design these RFICs from scratch, with no human templates — and the results look nothing like human work.

One AI-designed amplifier's layout "looked more like an arbitrary pattern or perhaps a QR code than the regular symmetrical structures we are used to seeing." Some chips "look more like modern art than circuit layouts."

Yet the prototypes beat the best silicon circuits on bandwidth, power, and efficiency, and the AI conceived them in a fraction of a human's time. Trained to explore the design space by playing against itself, the way AlphaGo Zero learned Go, a diffusion model now spits out a fabrication-ready structure "as if it were a vending machine," in about six minutes.

Previously:

• AlphaGo Zero shows corporate R&D has starved basic research
• Libratus poker AI beats several of the world's best players
• The Secret Life of Circuits book explains electronics for everyone