Oh man, I wonder if there is some weird self-selection bias when it comes to HN audience. It is a little like my dad telling me repairing cars is easy: you only have to do x,y and z without considering that I don't have years of experience of doing just that.
But my point is imagine if the person that did that simple TD lite model pre-training did something less... ad-related in his life.
ML when connected to business applications is one of those things that sound complicated but is in practice just a few lines of code usually. I'd argue that anything frontend is way harder to code up than a pretrained TF lite model, despite the general idea being the opposite.
You are ignoring the point I am making so I will make it a little more explicit for the sake of ensuring that it is not just my inability to make my point across:
What is the societal opportunity cost of a person not doing $this_thing_you_mentioned and applying it to just about anything else.
You still need to learn enough to convince yourself you can do it, and then do it. A lot of people just shy away from this stuff because they think they'll never get it.