What problem does the two-tower retrieval architecture solve that simple subreddit-subscription filtering doesn't?
Subscription-based retrieval only captures interests a user has explicitly declared by subscribing — it misses evolving interests and can't distinguish which posts within a subreddit are actually relevant to this specific user. A two-tower model learns separate embeddings for users and posts from historical click behavior, enabling a nearest-neighbor search that surfaces relevant content beyond explicit subscriptions, computed cheaply because post embeddings are precomputed ahead of time.
Answered in
Reddit's Hardest Notification Problem Isn't Ranking, It's RestraintReddit's notification pipeline spends as much effort deciding when to stay silent as deciding what to send — budgeting runs before ranking.
Read the full analysisOther questions this article answers
- Why does Reddit's notification system decide a daily budget before it even looks at candidate posts?
- Why use causal modeling instead of simple correlation to set notification budgets?
- Why does the ranking model predict multiple outcomes (click, upvote, comment) instead of just one?
- Why does the system still need a reranking stage after the ranking model has already scored everything?
More how ai actually works questions
- How many new researchers can receive the next OpenAI grant?
- When do applications open?
- Does the classifier make a final decision about an image?
- Can any developer call the classifier?
- How can Cursor answer questions about an entire codebase without storing the code?
- If embeddings are stored instead of source code, can the original code be reconstructed from them?
- Why does Cursor use a Merkle tree to keep the index synchronized instead of just re-scanning changed files?
- What happens when the server needs the actual source code, not just its embedding?
Every answer on Crashtech is written by the editor of the article it comes from — never auto-summarised. Browse all answers or the How AI Actually Works beat.