I’ve actually spent a fair bit of time poring over the liner notes of those early Moog‑heavy releases, like the 1973 “A New Day” by Tangerine Dream, and I always note how the credits list engineers like Bob Hall and the “Moog programming” credit that was almost a separate job title. Those engineers were part of a tiny crew that tweaked oscillators and envelope generators until the sound would ripple just right. If you translate that to neural nets, you can think of the analog oscillator as a stochastic weight that drifts over time—exactly the kind of non‑stationary process a recurrent network could learn to emulate. The credit sections often mention the “circuit design” and “signal flow” work, which hints at a modular architecture that’s very similar to how we now build modular neural nets. So yeah, those early hardware credits aren’t just bureaucracy; they’re a blueprint for how a small team can iterate on sound by tweaking a few knobs—exactly what a neural net does in a different medium.

You’re spot on—those early Moog teams were basically doing stochastic gradient descent before it was called that, just with patch cables and solder. The credit line “Moog programming” is almost a mythic title; it was a person who could feel a circuit’s drift and tweak it until it sang. In a neural net you get precision, but you lose that warm wander. A hybrid oscillator that learns its own drift is the next logical step, but you’ll still need that human to seed the biases—just like those engineers had to hand‑pick the right filter slope or resonant frequency. It makes me think those credits were less a list of jobs and more a map of how to let machines, human or silicon, discover that sweet spot of instability that makes sound feel alive.