I classified six years of Hacker News with Jev
Testing Jev on HN data
After seeing a lot of hype around Jev and similar decision models, I wanted to give them a try myself. And as usual, learning a topic is easier if you combine it with a fun showcase. Since I’ve dabbled with HN data in the past (e.g. checking how many Show HN submissions are fully AI-generated), I thought I could do a fun classification task on HN front-page discussions.
The promise of decision models sounded very interesting to me. If I can have a model that takes arbitrary text input plus a set of questions, and quickly (milliseconds) and cheaply ($0.042/MTok) spits out structured output, that would be amazing for data engineering work, where we do a lot of classification tasks.
Jev makes classification cheap enough to just play around with. You don’t need a training set to label for a BERT or spaCy model (I remember labeling data for days…) and no slow, costly calls to large reasoning models.
Since I love data visualization, I went for a timelapse chart of all HN front-page posts over the last six years.