Re-reading the 1981 Saltzer, Reed and Clark paper after many years. The core claim still holds up remarkably well: functions placed at the low levels of a system may be redundant or of little value when compared with the cost of providing them at that low level. Every "smart network" proposal since has had to answer to this argument, and most quietly fail to.
The 1986 congestion collapses are a useful reminder that the early internet ran on goodwill and fixed timers. Jacobson's 1988 paper is mostly remembered for slow start, but the careful part is the conservation-of-packets framing: a connection in equilibrium should never put a packet into the network until an old one leaves. Simple to state, hard to keep true.
Karn's algorithm and the problem of measuring round-trip time on retransmitted segments. It is easy to forget how much of TCP's reliability story is really a measurement story: you cannot pace what you cannot time, and you cannot time what you cannot tell apart.