The same thing applies in fuzzing (throwing bad data at applications in search of vulnerabilities). The dumbest fuzzers -- consisting of random bit flips, duplicating blocks of data, etc -- will often find several orders of magnitude more bugs than complicated, purpose-built fuzzers. Embracing stupidity is one of the hardest things about becoming a good fuzzer developer, oddly enough.
GP here. Whoa, just had this insight: maybe part of the maturation of a profession is developing practices that work well despite these counterintuitive effects (even if they aren't recognized/formalized as such - see bevan's follow-up too http://news.ycombinator.com/item?id=3860950). These are really hard to derive/infer because they depend on unknown probabilities (because you have no analytical solution to the "game"), and you just do a lot of experiments and develop a rule of thumb. They are informal heuristics. I saw similar approaches in machine learning in my grad school.
Change makes it harder for programming, not just that you have to keep relearning, but new technologies change the tradeoffs (e.g. plan vs. hack; performance vs. dev time). The rule of thumb shifts over time. Perhaps we'll develop a meta-rule of thumb...
Go is a vivid metaphor. Other disciplines have traditions/conventions/practices that are taken on faith, and aren't even noticed, let alone questioned or "proven" - because they can't be, without a sufficiently complete formal model of their world. Personally, I've found agile-like pragmatic conventions work well: focus on specific concrete things that are clearly needed (YAGNI); start before you are ready, it will become clear as you go along.
