Exploration vs Exploitation in ML. It seems you explore a lot to find whats 'best', but it keeps you from implementing (in this case Exploiting) a set of ideas. There is an optimum between the two. You could read up on this.
My 2 cents: Learn by need. Figure out your needs (e.g. required feature) and go get just that knowledge from the internet, and put it to practice.
I think this depends. You can (and should) learn things when you need them, but I have found that knowing about things before you need to use them greatly helps you when the rubber hit the roads.
Ideally I think it's a mix, where you always set some time aside (10-15%) to play and experiment with new / unknown things to sort of try to figure what they are/how they work. After that you'll form an opinion and be able to better pick thing to use when building something new and/or ramp up quickly in the learning by doing box because you've had some previous exposure.
I did try to do the way as you said in your 2 cents but the more I solve problems like that I tend to feel empty. It feels that I don't know anything and I just know some exceptions or just some solutions to some problems. I don't feel I have a fundamental knowledge base that lets me think from my own neural networks.
Break it up by time. As an example split... say you wanted to build a dashboard in a week. Use three days for research and four days for executing. Look up immediate blocking problems during your execution phase, but stay at a granular level.
My 2 cents: Learn by need. Figure out your needs (e.g. required feature) and go get just that knowledge from the internet, and put it to practice.