HOWEVER I think there's more nuance: there is tremendous value in having a baseline level of foundational knowledge in whatever domain you're in.
For instance, if someone hasn't learned math to an undergraduate level, and they try to sink their teeth into a math research problem, then they're probably going to spend all their time flailing around and being confused. Waste of time.
A better use of time would be for them to develop a baseline level of mathematical knowledge through a structured curriculum, and then take the leap once they've got their fundamentals down.
Same thing with machine learning. If you want to work on a research problem and you have your ML fundamentals down, then yeah, go ahead and jump right into the research problem. But if you don't even know how gradient descent works, or what an eigenvalue is, how to work with continuous probability distributions, etc., then you're woefully underprepared and you'd be better off just shoring up your foundations first.