Two complementary strategies a recommender system can use:
Pure exploitation over-serves already-popular content to already-engaged niches and starves new/small creators of reach, since the system never tests them against users outside their existing footprint. Weighting some ranking capacity toward exploration — e.g., showing a niche genre to a fan of mainstream artists, as TikTok does — is a deliberate lever to give small creators a chance at new audiences, at some cost to short-term engagement optimality.
Related to Illegible Embeddings Becoming Legible via LLMs (the representation exploration/exploitation operates over) and Frequency-Ladder Growth Metric.
Adam Mosseri says that, for the first time in his tenure, Instagram believes it has 'line of sight' to matching or beating TikTok on some recommendation dimensions. He credits TikTok/ByteDance's exploration-heavy ranking — surfacing content a user doesn't yet know they'd like, e.g. showing Afrobeats to someone who mostly engages with Bieber — as the inspiration for Instagram's own increased investment in originality and breakout-content ranking.