This is a great example of how we "collectively" [1] conflate phenomena, skillsets, ... into one topic: machine learning, AI.
1) There is the general phenomena or collective project, where hardware, algorithms and human insights are improved to approach the situation of man-made intelligent machines.
2) There are the people who are designing algorithms, using mathematical intuition and knowledge, analogies with physics, etc... Most people would agree these people are doing optimization / machine learning "proper".
3) There are the people working on improving hardware for machine learning / optimization purpouses, by looking at the most performant algorithms, breaking them down into primitive operations and requirements for hardware, there are also people working on the algorithms themselves and finding computational shortcuts (which can end up in software or hardware, can end up as proprietary knowledge or common knowledge, ...). The distinction between hard and software is somewhat blurry, since hardware designers can optimize or implement a section of software into hardware. A lot of this can still be considered ML "proper".
4) Then there are the people who apply the ML frameworks and their exposed choices and settings to a specific problem domain. Many of them don't need to understand the internals if they don't need state of the art results. Many would nevertheless benefit from understanding the internals, and the requisite math. What I propose is to stop calling their activity as Machine Learning, and instead call it Machine Teaching. They are teachers, and just like elite schools they can choose which specific type of available student they will teach, and they can tweak (or filter from a large family of students) which student they select to teach the task at hand. There are bound to be many advantages of having actual human teachers get involved in machine teaching. These people will not be proficient in designing novel families of students unless they also know the requisite math, and identify those ML papers that are ML "proper" instead of ML "teacher". When trying to find important foundational insights in ML "proper" one is typically overwhelmed by a large surplus of ML "teacher" type papers. These are important datapoints, and necessary to advance human insight into ML "proper", but they are data, not knowledge. There are actual ML "proper" knowledge papers out there that explain why a certain phenomena is such and so, and they get very little attention because they necessarily lag the breakthrough ML datapoint paper, and most ML "teachers" don't have the math background to understand them. So the probability that a given ML "proper" researcher fundamentally improves the state of the art is much higher than the probability that a given ML "teacher" will fundamentally improve the state of the art. At the same time the probability that a given fundamental breakthrough was achieved by an ML "teacher" is higher than the probability that a given fundamental breakthrough was achieved by an ML "proper" researcher:
P( Breakthrough | Proper ) > P ( Breakthrough | teacher)
while
P ( Teacher | Breakthrough ) > P ( Proper | Breakthrough )
Since most people don't have the broad math / physics / ... knowledge to draw on, the number of ML "teachers" is much higher than ML "proper" researchers.
[1] well, really, some actors have vested interests in conflating those together...
EDIT: just to be clear, I am not complaining about ML Teachers, we need the ML Teachers, and their breakthrough datapoints. What I am complaining about, is conflating both activities of ML Proper and ML Teaching. This makes it harder for the few ML Proper researchers to find each other's insights.