Learning pharmacology made me better at AI systems

I noticed the overlap late: pharmacology had been teaching me AI system design all along. The models rhyme so well I now debug AI features with pharmacy reflexes.

The transfers

Dose-response curves. Drugs don't improve linearly — effect rises, plateaus, then toxicity. AI capability follows the same curve: more model, more context, more autonomy — up to a point. Past the plateau, extra spend buys risk, not quality. Find the plateau; build there.

Half-lives. Every drug has a decay rate; dosing schedules exist to stay above the effective threshold. Knowledge in a system decays too — cached answers, embeddings, documentation. Systems need "redosing" schedules: refresh cycles, staleness limits, re-indexing. My RAG corpus has a half-life. So does my prompt library.

Interactions. Two safe drugs can be dangerous together. Same in systems: feature flags + caching + AI generation can interact in ways none of the three does alone. The interaction is where production bugs live, and checklists exist to catch exactly those combinations.

Therapeutic index. Drugs are dosed between effective and toxic. Features have the same window — the margin between useful and harmful. Autonomy features have narrow indexes; I widen them with human checkpoints the way pharmacists widen safety with monitoring.

The meta-lesson

Disciplines rhyme. The person who only knows one misses the rhymes — and the rhymes are where intuition comes from.