Exploration vs. Exploitation (fffej.substack.com)

The piece frames product and R&D strategy as a multi-armed bandit problem: companies must balance exploitation (doubling down on known winners) against exploration (testing new bets) while minimizing cumulative regret. It argues organizations systematically under-explore because incentives and timelines (quarterly revenue, 2–3 year career cycles, investor pressure, promotion rules) bias decisions toward short-term exploitation. In fast-changing, high-uncertainty domains like AI/ML this bias is especially harmful: markets are non‑stationary, so a persistent, non-zero exploration rate is required to avoid being locked into suboptimal solutions. Practically, the author maps bandit algorithms to organizational levers: ε‑greedy → commit a fixed exploration budget (suggested ε = 15–20%); UCB → staged, small bets that scale if signals or uncertainty justify it. Tactical recommendations include dedicated “innovation” teams measured on learning and options created (not velocity), time‑boxed hack sprints, and bet-style governance with small teams, short timeboxes, early signals, and kill rules. Key metrics should be time‑to‑confidence, kill rate, and how quickly learnings inform plans. The write-up warns of human switching costs and role specialization (pioneers vs. settlers) and emphasizes institutionalizing exploration—funding it and measuring it—rather than exhorting people to “be more innovative.”