What turned out to matter

Papers published twelve to twenty-four months ago, ranked by how often the field has since cited them. Unlike the reading list, this is a measurement rather than a recommendation — but citation counts say a work was built on, which is not the same as a work being right. Counts come from OpenAlex.

Why this window: papers younger than about six months have essentially no citations at all, so ranking them would rank noise. Closing it at twenty-four keeps this a retrospective instead of an all-time chart that never changes. And the pool is what we happened to be collecting one to two years ago — broader and noisier than today, which is why a real-time object detector can outrank every language model here. That is a fact about our old intake, not a claim about the field.