my conclusion is to never go depth first (either specialist or generalist) unless you're feeling clear benefits right away.
balancing
Even when machine learning solutions have unambiguous value propositions and direct paths to cost-effective implementations based on known techniques and resource requirements, the projects are subject to extreme risk of internal political shutdowns, arbitrary gatekeeper constraints on data limitations or limitations of what technique can be used so that specialists in technique X can build up longterm control of a profit stream (I’ve seen this a lot in investment firms) and eat up all the promotion opportunities by convincing leadership that their version of a research platform is the only allowable one, and more.
The article “Who Cares About Forecast Accuracy,” by Robin Hanson is a good discussion of this.
https://www.cato-unbound.org/2011/07/13/robin-hanson/who-car...