“"We know this algorithm is dumb and needs major improvements, but at least you can see us struggling to improve it." — Elon Musk (source: basenor.com, fetched 2026-05-18)”
You know that feeling when a social feed surfaces content you reported last week, and the company's only explanation is 'our algorithm thought you'd like it'? Until this repo, every major social platform kept its ranking logic proprietary. If you wanted to build a recommendation system, you worked from academic papers and behavioral experiments — no ground truth. This codebase gives you the first public look at a production social feed pipeline: a four-component system whose engagement weight table, filter stages, and transformer attention mask are now all readable.
Every time you open X, Home Mixer fires a request that pulls candidates from two parallel sources: Thunder retrieves posts from accounts you follow using a Kafka-fed in-memory store with sub-millisecond lookups, while Phoenix retrieval uses a two-tower embedding model to find out-of-network posts that match your interest profile via dot-product similarity. Both batches get enriched with metadata, then filtered for duplicates, old posts, and blocked authors. Surviving candidates go into Phoenix's ranking transformer — a Grok-1-derived architecture that scores each post in isolation: no candidate attends to other candidates in the batch, only to your history and context. The model simultaneously predicts 19 engagement types and combines them with asymmetric weights (a 'report' costs −369 points; a 'like' earns +0.5). The ranked list exits via gRPC.
If you're an ML engineer designing a recommendation pipeline, this gives you a concrete reference for candidate isolation, multi-task engagement scoring, and the Rust/Python serving split — all non-obvious design choices that are well-documented here. If you're a backend engineer curious how high-throughput ranking pipelines are structured in Rust, the composable trait-based candidate pipeline is worth a read. This is not useful yet if you need a production-ready recommendation library: the released weights are a mini checkpoint trained on a demo corpus, training infrastructure is absent, and...
Worth reading if you're designing a recommendation pipeline and want a concrete production reference for candidate isolation, multi-task scoring, and the Rust/Python serving boundary. The May 15 update makes it experimentally runnable for the first time, lowering the barrier from architecture reading to hands-on study. Not worth adopting in production: the checkpoint is a 4-layer, 128-dim mini model against a demo corpus, production weights are absent, the single CI-agent contributor signals no external contribution path, and there are no versioned releases.
Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.