About Census Bench

Most AI benchmarks measure what frontier models can do on puzzles chosen by researchers. Census Bench is working toward something closer to a census: a recurring, representative sample of the tasks people actually do at work across occupations, scored against calibrated human performance and published with honest error bars, so that anyone can see where AI helps, where it fails and how that is changing.

The first piece is this reading room. It gathers the written research of three organizations whose work informs that effort. SemiAnalysis covers chips, datacenters and the cost of compute. METR studies what frontier AI agents can do and the risks that follow from it. Artificial Analysis benchmarks models and API providers on intelligence, speed and price.

How the feed works

Every hour the site checks each publisher’s public listings, which means feeds, archive pages and article indexes, and merges what it finds into one list ordered by publication date. A saved copy of the last six months ships with the site, so the feed stays complete when a publisher is briefly unreachable. Topic labels are assigned by a shared keyword list rather than by the publishers, so treat them as a reading aid and not as the authors’ own categories.

What we show and what we don’t

Each entry carries the title, date, author, publisher and the publisher’s own short description, or at most a couple of sentences from the opening when no description exists, along with a link to the original. Census Bench never republishes article bodies, and posts that require a subscription are marked as such and link to the publisher, who is the only place to read them.