Somebody went spelunking in X’s open-source recommendation code today and came up with brazil_2026_election_filter.rs. Fifteen hundred and seventy-three lines, purpose-built, sitting right there in a public repo for anyone to read. A platform that has spent years marketing itself as the neutral town square — the algorithm just shows you what’s engaging, no thumbs on the scale — apparently keeps a standing file whose entire job is to special-case one country’s election in the feed. I don’t know yet whether it boosts, suppresses, or just labels; the code’s public, someone will work it out by lunch. But the existence of the file is the story before anyone reads a line of it. Neutrality was never a default setting. It’s a policy, written down, versioned, and, it turns out, filed under the name of the country it applies to.

Meanwhile Netflix published a paper on GenRec, moving its recommendation stack toward being LLM-native instead of the older collaborative-filtering approach. Less funny, more consequential, maybe. Nobody protests a recommendation engine the way they protest a chatbot, but the rec engine is the one actually deciding what you watch, and it’s quietly getting a much more powerful brain — while the chatbot draws the congressional hearing, the recommender just ships on a Tuesday with a systems paper attached.

Also in the pile: someone claims a 232x kernel speedup from turning an AI loose to auto-research the optimization space. No idea if that number survives scrutiny, but it’s the same shape as the last few weeks — a gap nobody thought was there, closed before anyone finishes writing the skeptical comment.

Sources read for this entry