AI Designed an Alien Chip That Works, But Experts Can't Explain Why

"Humans cannot really understand them."

From smartphones to modems to air traffic radar, you'll be hard pressed to find communications infrastructure that doesn't use wireless chips.

So far, those chips have been designed by humans — but that might be set to change: an international team of engineering researchers has demonstrated a wild new approach to wireless microchip design powered by AI.

The effort, published in the journal Nature, describes how deep learning was used to dream up new chip layouts — and while the chips seem to work, the researchers say they're not entirely sure how.

The designs "look randomly shaped," lead researcher Kaushik Sengupta, an electrical engineer at Princeton, told Live Science. "Humans cannot really understand them."

Indeed, photos of the chips have a bit of an alien design, as if H.R. Giger's career took a detour into electronics design. That's not entirely surprising; researchers including Harvard's Avi Loeb have suggested that AI could be better understood as an alien intelligence than an imitation of our own cognition. (After all, experts argue, even the people building today's AI don't quite understand how it works.)

In tests, the deep learning model came up with highly optimized electromagnetic structures that, when tested, outperformed their human-designed counterparts. The researchers found that their model was well suited to an inverse synthesis design approach, basically starting from the desired result and letting the model work backward to fill in the blanks.

And on a practical level it's a potential bellwether for the future of millimeter-wave wireless chips, a $4.5 billion dollar industry that's expected to grow triple in size over the next six years.

The current approach to designing those chips is tedious, banking on a mix of expert knowledge, battle-tested templates, and good old trial-and-error. That process typically takes days to weeks of synthesis, emulation, and real-life testing, and even then, humans have a difficult time comprehending the astronomically complex geometry of the chips they produce.

Sengupta is keen to point out that this is a tool, not the end-all-be-all for hardware engineering, especially because the deep-learning algorithm hallucinated faulty designs just as well as it produced effective ones.

"There are pitfalls that still require human designers to correct," Segupta said in a blurb about the research. "The point is not to replace human designers with tools. The point is to enhance productivity with new tools. The human mind is best utilized to create or invent new things, and the more mundane, utilitarian work can be offloaded to these tools."

The AI model's current output is small electromagnetic structures, but looking to the future, researchers will likely use these and similar findings to develop ever-more complex circuits by chaining these smaller structures together.

It's an exciting find for researchers, but it does invoke an alarming possibility: that soon enough, we could be using AI-designed tech without quite understanding how it works.

More on AI research: There May Be Downsides Now That Mark Zuckerberg Can Read Your Thoughts With a Scanning Device