News · 2026-06-27
AI is learning a 'dark art' that even expert engineers struggle with
Inside every phone, router, and wireless earbud sits a kind of chip that engineers speak about with unusual reverence and a little dread: the radio-frequency integrated circuit, the part that sends and receives wireless signals. Designing one is so reliant on hard-won intuition, so resistant to neat rules, that practitioners openly call it a "dark art." Now, according to reporting from IEEE Spectrum that drew real attention among engineers this week, AI is starting to learn that art - and it says something about where automation is heading next.
To see why this is hard, you have to understand what makes RF design different from ordinary chip design. Most digital circuits are, at heart, clean logic - ones and zeros, rules you can write down. RF circuits live in the analog world, where the physics is messy and unforgiving. At the frequencies radios operate, tiny details matter enormously: the exact length of a wire, the spacing between components, the way a signal at one spot leaks into a signal somewhere else. Two layouts that look nearly identical can behave completely differently. There's no clean formula that takes a specification and hands back a working design. Instead, expert engineers rely on years of accumulated feel - patterns they've internalized from thousands of designs, much of which they couldn't fully put into words if asked. That tacit, unwritten knowledge is exactly what makes it a dark art, and exactly what has made it so hard to automate.
Here's an analogy. Automating ordinary digital design is like teaching a computer to follow a recipe - the steps are written down, so a machine can execute them. RF design is more like teaching a computer to cook the way a grandmother does, by taste and instinct, adjusting on the fly with knowledge she never wrote down and perhaps couldn't. For decades, that kind of know-how was assumed to be safe from automation precisely because it lives in human intuition rather than in any manual. The news here is that AI is beginning to absorb it anyway - learning the feel from data, the way it has learned other skills that resist explicit rules.
Why this matters goes well beyond one corner of chip engineering. Most of the public conversation about AI automation has been about knowledge work that's already fairly explicit - writing, summarizing, coding, drafting. RF design is a different category: deep, specialized, physical engineering that experts themselves describe as more art than science. If AI can make real headway there, it suggests the automation frontier isn't stopping at tasks we can spell out. It's moving toward tacit expertise - the accumulated craft that takes a human a career to build. Engineers discussing the story extended exactly this worry and this hope: the same shift that could displace hard-won specialist jobs could also unlock designs and speed that human intuition alone never reached, and put scarce expertise within reach of teams that don't have a veteran RF guru on staff.
How would an AI even learn something nobody wrote down? The same way it learns most things it isn't explicitly taught: from examples. Feed a model enough real designs - the layouts, the measured results, the tweaks that worked and the ones that didn't - and it can begin to internalize the statistical regularities that experts feel but can't articulate. The machine never reads a rulebook because there isn't one; it absorbs the patterns directly from the record of what good designs look like, the way a chess engine learns winning positions without being handed a theory of chess. Pair that with the ability to rapidly simulate and test thousands of candidate layouts - far more than a human could try by hand - and the AI can search the messy design space in ways that complement, rather than copy, human intuition. That combination, learned feel plus brute-force search, is what makes a problem long thought too tacit to automate suddenly look approachable.
The honest caveat is to keep the scale in proportion. "AI learns the dark art" does not mean AI has replaced RF engineers or matched the best of them across the board. Learning to contribute to a hard design problem and learning to own it end-to-end are very different milestones, and the gap between an impressive assist and a trustworthy autonomous designer is wide - especially in a domain where a subtle physical mistake can sink an entire chip. What's genuinely notable is the direction of travel: AI reaching into a field long considered too intuition-soaked to touch. Whether RF design proves to be a one-off or the first of many "dark arts" to fall is the thing worth watching. For the bigger pattern of AI systems that plan and act in specialist domains, see AI agents.