Learning a stack by shipping with it
Java, then enterprise banking, then cloud, then platform engineering, now AI platforms. I have never once learned one of these before needing it.
There is a version of a thirteen-year career that reads as a list of technologies. Mine looks like that on a résumé and it is misleading, because the list is a side effect. What actually happened each time is that a problem arrived that the current stack could not answer, and the stack changed underneath the problem.
The most recent one is Python and FastAPI, picked up in delivery on the post-trade platform. Before that, Kubernetes. Before that, AWS serverless. Before that, micro-frontend architecture, which in 2021 I had read about and not run.
I have come to think the reading-first instinct is mostly avoidance. You can spend a long time being nearly ready. Shipping something small and slightly wrong with a new tool teaches you which of its abstractions leak, and that is the only part you cannot get from documentation.
The caveat matters though, and it is where this goes wrong in regulated work: learn on the thing whose blast radius you understand. I have learned new stacks on internal tooling, on the second service rather than the first, and on a replica running in parallel with the system it was replacing. Curiosity is not a reason to put a trading desk at risk.