The summary.
LEARN EVALUATION EARLY, BEFORE FRAMEWORKS
It is where beginners most often go wrong, and the errors are invisible. You can run a model without understanding it, and then cannot diagnose it.
Work on real, messy data — curated teaching datasets omit every difficulty that dominates real work.
MOST COMMERCIAL PROBLEMS ARE SOLVED BY CLASSICAL METHODS ON TABULAR DATA
Which is why that order — fundamentals, classical, then deep learning — reflects what the work actually requires.
READ PAPERS IN THE ORDER THAT LETS YOU ABANDON THEM QUICKLY
Abstract, figures, conclusion, then method. Be sceptical of weak baselines and missing ablations, and reproduce anything promising on your own data before adopting it.
A DEPLOYED, WORKING SYSTEM DEMONSTRATES MORE THAN ANY NUMBER OF NOTEBOOKS
It proves you can get past the point where most projects die. Write up what failed too — failures demonstrate judgement that successes alone do not.
BREADTH ACROSS THE LIFECYCLE BEATS DEPTH IN MODELLING
Most organisations need someone who can take a problem from framing to production and keep it working.
FUNDAMENTALS LEARNED WELL OUTLAST EVERY FRAMEWORK
Note the promising thing, and revisit in a few months. Most do not survive that long.
