After half a day with Jev, a developer wants to rebuild their job crawler
The author frames Jev as a model for multiple-choice decisions and plans to rebuild a crawler with it to judge whether links are job posts and whether listings have expired.
After half a day with Jev, my mental model is simple: LLMs are great writers. Jev is great at multiple-choice questions. LLMs generate, explain, and reason through problems. Jev takes a situation and a list of possible answers, picks one, and tells you how confident it is. Your software can act on that answer. That makes Jev a good fit for automation built around lots of small decisions. An agent playing a game has to keep choosing whether to jump, dodge, attack, or move forward. A web crawler does the same kind of work: follow this link, skip that page, classify this result. I once built a crawler to collect job listings from companies' official websites. It had to decide whether each link led to a job, whether the listing had expired, and which category it belonged to. I want to rebuild it with Jev and see how it holds up.