Agents can't infer your preferences from a few questions — users must teach the model

jamesbrandecon · x · 2026-09-25

James Brandecon argues there are obvious hard limits to how well any agent can infer a principal's preferences from just a few questions, with no way around it. Users will therefore have to invest real time teaching a model — not only specific preference rankings, but how to model their preferences so it can extrapolate to new situations. The observation highlights a core tension in agent personalization: passive questioning isn't enough, and user-side teaching effort is unavoidable.

Original post →

More from AGI Musings

AGI Musings channel →