What Personalisation Leaves Out
A briefing shaped around known interests can deepen understanding while making unrelated but valuable ideas less likely to appear.
Developed from a conversation between Pete Winn and Andy David
Andy designed Pliny to turn useful material from Twitter and other sources into personalised educational briefings. Instead of reproducing a noisy feed, the agent identifies subjects that match his interests, researches them and builds on what he has already read. Feedback about which briefings matter can improve that relevance further, but it can also make the system increasingly attentive to familiar territory.
A social feed has a quality that a tightly filtered briefing struggles to reproduce. Its mixture of unrelated posts creates chance encounters as attention moves between distant subjects. Andy might deliberately build a cluster around AI agents, yet still find an unexpected detail about the last line of Roman emperors compelling. Nothing in his established AI interests would necessarily lead a recommendation system to that second subject.
Pliny could search just beyond the edges of Andy’s current interests, using a language model to suggest nearby topics, sources or podcasts. That would broaden each cluster without restoring a genuinely random walk, because every suggestion would still begin from what the system already knows he likes. The better Pliny becomes at following the AI cluster, the easier it is for Roman history to remain invisible.