I've been thinking a lot about personalization lately, in the context of coaching systems and personalized medicine. The systems I'm building have conversational elements and content that are unique to each person, but how should I deal with optimization at the app level? Specifically, what does A/B testing mean in the age of Agentic AI? And would a look at the history for of personalization and optimization give me any clues?
It turns out it does. Personalization and optimization have been on parallel tracks for decades — contextual targeting in the 1950s alongside direct mail split-tests, cookies alongside web A/B testing, algorithmic feeds alongside multi-armed bandits. Turns out that they meet in contextual bandits, real-time adaptive algorithms which ask not just "which version is better?" but "which version is better for this person, right now?" Vowpal Wabbit, an open-source machine learning system by Microsoft Research, seems to be a gold standard. I can't wait to try it.