Thinking about hard problems in terms of lambda calculus. I took a day to read about all that stuff and at some point so many things fell into place. I learned CS theory in school, learned ML, but it's not the same as being confronted by a seemingly monstrous software project and then realizing ... omg, everything is relational and I can just write an interpreter. No longer asking "how do I write this in ruby/java/js ..." but how do all of these things relate to each other and if which language would be best for encoding my invariants. All kinds of ML and AI fields go from "data science, math stuff ... disregard" to a fascinating universe of answers to questions that you already have in your head for how to solve your real world software problems. The usefulness of macros became obvious and custom monads are rewarding achievements. Closures, instead of being occasional conveniences, became a swiss army knife for eliminating explicit data structures.

Not for all projects of course, but when building something enterprisey, simulated, or creative where users can model their own domain I believe that at least being familiar with the theory is not optional. I'm also somewhat convinced now that most "10x programmers" are maybe ~2x programmers who have simply encoded their expertise into a private DSL using JetBrains MPS or something. Allowing themselves to spend more time snowballing their skills and knowledgebase, and less time doing repetitive tasks and managing incidental complexity.

Oh also, I started using linear constraint solvers for automatically solving small to medium sized problems. You'll usually see them used for UI layout constraints, but that's just the beginning! For example, say you're building a small scheduling system, or a lightweight recommendation system that lies somewhere in between an if-statement and a full blown classifier, or you're writing some arbitrary last-minute business rule for a client where several things depend on each other with varying levels of weights/importance. If you know some things are linear, and the problem is unlikely outgrow the solver, then save yourself hours or work and just solve it at runtime!