Explore / Exploit

Henrik Karlsson wrote about the multi-armed bandit problem as a model for how to allocate your time — when to try new things versus when to double down on what's working. The slot machine metaphor is simple: you have limited pulls, some arms pay better than others, and you don't know which ones until you try. The tension is between gathering information and using what you already know.

What stuck with me is the idea of deliberately lowering your exploration temperature over time. Karlsson tracked this across thirty months — starting wide, gradually narrowing. Not a vague intention to focus, but an actual schedule. The discipline of that appeals to me.

I recognise the failure mode he describes. I have a dozen projects running at any time. Arc, Fiction, Falcon, Siteinspire, this site, whatever I started last Tuesday. Each one is genuinely interesting, and each one would benefit from more sustained attention. The explore instinct is strong — there's always a new tool to build, a new idea to chase. But the compounding only happens in the exploit phase, when you stay with something long enough to push through the plateau.

The counterargument is that exploration is how you find the arms worth exploiting. I wouldn't have built Arc if I hadn't explored Claude Code's extensibility on a whim. Fiction came from an idle experiment with long-form generation. The best projects emerged from unfocused curiosity, not strategic planning. So the question isn't whether to explore — it's knowing when to stop.

I want to think more about what a deliberate temperature schedule would look like for my own work. What would it mean to commit to three things instead of twelve? Which arms have I pulled enough to know the payout?