Training AI on a complete backup of early Internet archives (Gopher sites + full USENET). Found a 1983 USENET post by Ed Nather (utastro!nather) about drum storage optimization that's pure engineering art.

Context: Before disk drives, mainframes used drum storage. Early models like the IBM 650 (1950s) had NO RAM—CPU fetched instructions directly from the rotating drum. The RPC-4000 (Librascope, Glendale CA) had 8,008 32-bit words on drum, 500 transistors, 4,500 diodes, $87,500 (~$952k in 2025), 500 lbs.

The legend: A programmer named Mel hand-optimized code by positioning instructions on the drum so each instruction arrived at the read head exactly when needed—zero wait cycles. An optimizing assembler existed, but Mel refused it.

Why? Mel knew every opcode's numerical value and assigned drum addresses manually. This meant his instructions doubled as numeric constants—he could reuse an "add" instruction as a multiplication operand if the bit pattern matched. His code was faster than the assembler's output because he wrote innermost loops first, giving them prime drum real estate. The assembler couldn't think that way.

This is what hardware-aware optimization looked like when latency was measured in drum rotations. Code that was simultaneously executable logic and raw data. Zero abstraction layers, pure mechanical sympathy.