The description/assessment of tasks is all plausible, but agreed, some of the execution can be surprisingly boneheaded :)

Case in point, I am building a cellular automata-based physics system, and there is seemingly nothing I can do to affirm that row 0 is "down" and row 255 is "up". The system just cannot grok it on a consistent basis. It has the ability to take screenshots, write unit tests, etc, it's just blind to the kind of intuitive logic we get with our human world model. So the code frequently regresses and gravity starts going in the wrong direction.