# on exploration vs exploitation
In reinforcement learning an agent has to choose between the action it already knows pays off and an action it has not tried yet. Pull the known arm and you collect a reliable reward. Pull an unknown one and you might find something better, or waste the pull.
I think the same trade-off runs through a career, and through a life. Most of my working hours go to exploitation: shipping agent infrastructure that has to be correct today. A deliberate fraction goes to exploration: reinforcement learning, robotics, ideas with no obvious payoff. The fraction is the interesting decision. Too little and you plateau. Too much and nothing compounds.
The cleanest version of this is the secretary problem. You see n candidates one at a time, must accept or reject each on the spot, and want the best. The optimal strategy is to reject the first n/e of them, about 37%, no matter how good they look, purely to learn what good looks like, then take the first one who beats everyone so far. That first 37% is exploration you have already decided to pay for. It shows up in apartment hunting, hiring, and job offers with deadlines: anywhere options arrive in sequence and cannot be revisited.
The honest caveat is that the 37% assumes no recall and that only the very best counts. Relax either and the exploration window shrinks. Knowing which assumptions your situation actually meets is most of the skill.
This site is built the same way. Flip the switch or move the slider and you change the policy. Exploit shows the work I am confident in. Explore shows the arms I am still pulling.