I haven’t seen anyone build it yet, but I’m just waiting to see when we first get a fully integrated hardware device where the software was driven entirely by a single learned agent doing predictive control, all the way down to the firmware and controllers.

You’d obviously need a hierarchical approach, some kind of differentiable neural program synthesis that let you have super low latency components generated from the core. But that’s doable already.

What I’m really curious about is how much of the operating system can go too. If you only need to run a single program, with a fixed size and number of layers, why have any intermediate components?

We build hardware today for recursive hierarchical programs today. But what should hardware for corecursive, non-terminating, adaptive software look like?