G: Yes — chip makes it click. I’m almost certain I’ve found the experiment you remember.

It was Adrian Thompson’s evolved FPGA circuit, done at the University of Sussex in the mid-1990s and published as “An Evolved Circuit, Intrinsic in Silicon, Entwined with Physics.” citeturn0search46 It is also independently remembered by Science of Discworld readers specifically as the Adrian Thompson experiment discussed in the book. citeturn0reddit54

And it is every bit as wonderfully bizarre as you probably remember.

What Thompson actually did

Thompson took a field-programmable gate array (FPGA) — basically a chip containing lots of configurable electronic logic — and deliberately didn't design a circuit to solve his problem.

Instead, he used a genetic algorithm.

The task was essentially:

Here are two different input frequencies. Configure this physical chip so that its output lets me distinguish one from the other.

The crucial point is that evolution occurred on the actual silicon, rather than entirely inside a computer simulation. Each candidate configuration was physically loaded onto the FPGA, its performance was measured, the better-performing configurations reproduced with mutations, and the process repeated. citeturn0search46

Eventually, evolution produced a working frequency discriminator.

Then Thompson examined what the hell it had built.

And that's where the experiment became famous.

The resulting circuit was nothing like something an electrical engineer would design. It had evolved to exploit subtle physical properties of the particular silicon it inhabited rather than merely implementing a clean arrangement of conventional digital logic. Thompson's paper explicitly describes the result as having dynamics intertwined with the physical properties of the implementation medium. citeturn0search46

Most famously, there were apparently logic cells that weren't connected into the circuit in the conventional logical sense but nevertheless mattered. Remove them and performance deteriorated or the circuit stopped functioning. The evolutionary process had apparently discovered electromagnetic/analog interactions between components that a normal digital-circuit design methodology simply wouldn't regard as part of the available design space. citeturn0reddit54turn0search46

So your memory of the Science of Discworld example is basically:

They evolved a chip → it worked → humans took it apart to understand why → parts that appeared to do absolutely nothing turned out to be necessary → evolution had exploited properties of the physical hardware that weren't part of the human engineer's conceptual model of the hardware.

That last bit is almost certainly why Stewart, Cohen and Pratchett were interested in it. The Science of Discworld is heavily concerned with emergence and with the distinction between explaining something retrospectively and assuming that its structure must therefore have been designed according to that explanation. The book itself was published in 1999, right after Thompson's work appeared in the evolvable-hardware literature. citeturn0search45turn0search3

The original paper is available here:

urlAdrian Thompson — “An Evolved Circuit, Intrinsic in Silicon, Entwined with Physics”turn0search46

There are actually several details of this experiment that are even stranger than the usual popular retelling. In particular, Thompson did experiments involving moving the evolved configuration onto nominally identical chips and changing temperature, which help establish what kind of physical weirdness evolution had latched onto. I can dig into the original paper and reconstruct the experiment generation-by-generation, including exactly what the supposedly “disconnected” components were doing.