Tumblr’s adult-content ban takes effect Monday, announced two weeks ago (proximate cause: a child-safety app-store removal; structural cause: Verizon-owned Tumblr’s ads business needing brand safety), and executed via an ML content classifier whose false positives have become a genre unto themselves: flagged sand dunes, flagged classical sculpture, flagged Tumblr’s own announcement post. Two lessons for the file. One: content moderation at scale is classifier deployment at scale, and shipping a high-stakes model with that precision profile is the ML equivalent of the TSB cutover; the confusion matrix is the product. Two, the bigger one: communities are load-bearing (the Reddit lesson) and this is its corollary: a platform purging its core community’s content is a migration event for that community, and the diaspora (Twitter, bespoke sites, Discord) is already routing around the damage. Platforms are temporary; archives are personal. Export your things.
Also this fortnight, two flights of note: Virgin Galactic’s SpaceShipTwo reached space (by the 50-mile definition) Thursday, the first crewed US spaceflight from US soil since the Shuttle retired, and a genuine redemption arc four years after their fatal 2014 crash. And Google+ announced it’s dying faster, a second API leak (52M this time) moved the shutdown up four months. Even the graveyard has SLAs.
Smash Bros. Ultimate is out and the office bracket is seeded; I main King K. Rool out of respect for legacy systems that finally got their port.
TIL: precision-recall trade-offs have politics. Tumblr chose recall (catch everything, flag the sand dunes) because the cost function was written by lawyers, not users. Every threshold is a values statement wearing a decimal point.