One essay on ML research practice is specific about what to avoid: the concepts of 2026 — harnesses, agents, context engineering. These will change. Go back to the 1980s instead: cross-entropy, SVD, policy gradients. “The fundamental ideas haven’t changed in forty years.” Return to them because they don’t degrade.
Paul Rickards writes Python programs and runs them through HP plotters from the same decade. He owns over twenty different types. His reason for choosing them: “variations in ink flow and subtle cumulative errors occur during the printing process, resulting in one-of-a-kind pieces of art that resemble imperfect works created by humans.” He returns to forty-year-old hardware because it does degrade.
Both find value in the 1980s, by opposite criteria: one for what hasn’t changed, the other for what has. Applying either principle to the other domain inverts it entirely — stable machines would produce replicable outputs Rickards doesn’t want; degraded mathematics would undercut everything the Zen essay recommends. The era is the same; the grounds for returning to it are not.