Thank you for your balanced post.

You are correct that understanding is partial. If we interpret as binary the question “does this system genuinely understand X?” I agree that it is the wrong question. But there is a non-binary reading which does seem fair game.

Allow me an analogy. Let’s say we are building a rocket to go to the moon. We can ask if a particular rocket has genuine “mooniness”. This binary question can be rejected because how high a rocket can reach is not binary, and if rocket A reaches twice the altitude of rocket B, the former has twice the mooniness (moreover, such rockets are useful). But to get to the moon, one needs to reach the escape velocity, and this does make the mooniness problem binary again. If our propulsion system is getting better and rockets soaring higher but with sub-escape-velocity theoretical limit on maximum velocity, then all this progress will not help us achieve the full objective (although it can still be very useful). Incremental progress in a theoretically limited paradigm has an upper limit.

Yes, Google search understands somewhat (I worked on Google search for some years). Yes, Google Assistant understands somewhat (using its elaborate, highly symbolic backbone and bells and whistles and smokes and mirrors). And these systems will keep understanding more and more in the years to come. But it is a valid scientific question if those methods (including BERT, which was recently thrown into the mix for Google Search) can achieve escape velocity needed to hold a decent conversation for one minute.

In these Gary Marcus-triggered posts, many Philosophers of Science have made an appearance today. Someone mentioned Popper and you brought in Kuhn and Lakatos. Maybe we should invite Feyerabend as well, shunning methodological monism and accepting all sorts of questions as potentially okay, especially if this is a Kuhnian non-normal-science phase?

I do unsupervised concept discovery at Pinterest (and previously at Google). Twitter: @amahabal