Empirical

PAC learning and VC dimension's bounds predict that an overparameterized network cannot generalize, and it generalizes anyway.

Essential vs accidental complexity, and why adding people to a late project makes it later.

The constant factors that dominate asymptotics in practice, and cache-oblivious algorithms as the theoretical response.

SAT is NP-complete (The Cook–Levin theorem) yet solvers handle millions of variables.

Systems mirror the communication structure of the organizations that build them. Empirical, not proven — and the inverse maneuver as its practical use.

Amdahl's rebuttal, and then the law that adds a negative term for coordination: past a point, more nodes make it slower.

Every observable behavior of your API is depended upon by someone; and the robustness principle that made that worse, plus the modern case against it.