AI learns the “dark art” of RFIC design

	AI learns the “dark art” of RFIC design(spectrum.ieee.org)
	270 points by Brajeshwar 58 days ago \| 174 comments
	tl;dr: Princeton researchers are using reinforcement learning, inverse design, and diffusion models to automate RFIC design—a notoriously artisanal field where chips for 5G, radar, and satellite comms have traditionally been hand-crafted over years. Their AI-generated power amplifiers, which often look like QR codes rather than symmetric layouts, have achieved record bandwidth and efficiency while cutting design time from months to minutes. The main bottleneck now is training data, most of which sits locked behind corporate NDAs, prompting calls for open chip-design datasets akin to ImageNet.
	HN Discussion: ~This phenomenon of algorithm-designed uninterpretable circuits is decades old, not novel ~The article overstates novelty; this is similar to genetic algorithms and brute force search ↓Questions robustness of AI-generated designs and whether conventional subblocks carry the weight ↓Conflating LLMs with traditional ML techniques muddies the discussion unfairly •Philosophical musing on whether nature's truths may be ugly messes only machines can grasp