The summary.
IN ROBOTICS, NO SINGLE SENSOR IS ADEQUATE
Sensor fusion exists because each fails differently. And safety must come from independent mechanisms, not from the main system behaving correctly.
TRAIN AND TEST COMPUTER VISION ON DEPLOYMENT CONDITIONS
Models trained on clean images fail on real ones. And watch for backgrounds correlating with classes — the model learns the background.
IN VIRTUAL AND AUGMENTED REALITY, PERFORMANCE IS A COMFORT REQUIREMENT
Dropped frames cause immediate physical discomfort, not merely visual roughness.
ASK OF ANY DISTRIBUTED LEDGER PROPOSAL WHETHER IT REQUIRES THE ABSENCE OF A TRUSTED PARTY
If a trusted party exists and is acceptable, a database is simpler, faster and cheaper.
A CORRECTLY IMPLEMENTED WRONG MODEL IS CONFIDENTLY WRONG
Which is why validation and verification are different activities, and why assumptions and uncertainty must be stated alongside results.
MOST OPTIMISATION DIFFICULTY COMES FROM MISSTATING THE PROBLEM
Constraints nobody mentioned emerge only when a solution is proposed.
NEVER COMPARE FLOATING POINT VALUES FOR EXACT EQUALITY
And use established numerical libraries — naive implementations of standard algorithms are frequently unstable.
BUILD THE MATHEMATICS FIRST, THEN BUILD SOMETHING, CONTINUOUSLY
The gap between understanding a paper and implementing it is where the learning actually happens.
