[exelius]

IMO we're kind of headed that way now. I always saw this as the point of Google's combined interest in TensorFlow and Kubernetes.

Containers are the piece of the toolchain we've been missing. Now we actually have some feasible logical methods (deep neural networks + gradient descent) that can be used to structure existing computational tools into deeper, more intelligent systems.

Think of it this way: what's the difference between (an ideal) container and an artificial neuron? Structurally, they are nearly identical: they both have collectors and emitters, and perform some non-linear action in concert with other similarly-structured systems. Containers can also help with some of the "trust" problems: if we're shipping around trained data models (or container images) rather than the actual training data, we can push storage out to the edge of the network and run the models there. Containers provide a common computing language that enables you to do that.

This platform is not a leap forward for the theoretical capabilities of AI; but it is a shortcut that should eventually make it easy for AI researchers to leverage vast existing libraries of software written in any one of hundreds of programming languages.

I actually suspect that this is just the first generation; there are a number of software problems that currently can't be solved easily in a multithreaded manner. However, if you can build a container that does what you need, you can eventually train a model to replicate the container -- it may be less computationally efficient, but with future orchestration platforms it may end up being more time-efficient.