How the Substack feed is learning to understand your reading journey (mrkcohen.substack.com)

Substack announced it has rebuilt its feed recommendation system, migrating core retrieval tasks from classic two-tower models to sequential architectures that treat reading as a dynamic journey rather than a static profile. The change — retrieval cutover already complete, ranking models to follow — moves the feed from a rules-driven/averaged-embedding approach to sequence-aware models (Transformers/RNNs/LSTMs with attention) that incorporate hundreds of recent interactions (≈10x more signal than before). The goal is session-aware discovery: the feed can immediately react to a click or a new subscription, surface content that logically continues your current session, and infer higher-level states like “exploring” versus a stable long-term taste. Technically, the two-tower baseline encoded users and items into precomputed vectors for fast nearest-neighbor retrieval but was limited by static user embeddings, coarse pooling, and tight feature caps. Sequential models instead maintain a dynamic representation updated in-session, using attention to weigh different interactions (recent spikes vs persistent subscriptions) and predict the next engagement more precisely. This mirrors the trend of applying language-model architectures to recommendation tasks and has practical implications for retrieval/ranking pipelines, training/compute trade-offs, and richer personalization signals across content types (posts, notes, publications, creators). For ML practitioners, Substack’s move is a concrete production example of scaling sequence models for large, diverse recommendation inventories.