Is the scoring model actually weighting engagement actions differently or is it treating likes/replies/reposts the same? The real algorithm weights those pretty aggressively differently
It's real embeddings (MiniLM, truncated to 128-dim), so the cosine similarities between tweets are genuine, but it's obviously not X's actual trained model. The relative distances are meaningful, though.
How close do the in-browser embeddings actually get to what the real pipeline produces? Or is it more about showing the general flow than matching real outputs?
They're weighted differently. Currently it's set to: reposts 2.0x, replies 1.5x, bookmarks 1.2x, likes 1.0x, and clicks 0.5x. These are guestimates based on public info.