I was impressed by Jev, please explain why I shouldn't be

That is the pitch: define the topic in natural language at runtime and get a bounded probability back, with no training step. The pushback is that a fine-tuned BERT-class model has always done this, it is just not turnkey and does need your own labels.

Yes, and it did not embarrass it. The original poster downloaded FLAN-T5 for a similar task, got similar results, and pointed out the local model also keeps the data on their own machine instead of sending it to a hosted endpoint.

No. It is one practitioner's anecdote on their own participant entries, with no labeled set, no codebook agreement check and no precision or recall. The entry's guidance holds: check accuracy against your own codebook before trusting it.

One comparison is that it is rule-based programming with a better pattern matcher, pointing at work from the eighties. Another asks whether you could always get probabilities over a fixed token set without training anything new — the thread's real dispute is convenience versus novelty.

Offering non-technical researchers a way to classify thousands of participant entries by arbitrary themes without prior training, at near-instant speed and negligible cost, such as flagging entries mentioning Italian food even when no exact dish is named.

BERT, BART and FLAN-T5 for text classification, all roughly half a decade old, transformer-based and small enough to run locally. One commenter downloaded FLAN and reported similar results, with the benefit of local serving and privacy.

One commenter compared it to the old Hacker News comment dismissing Dropbox because you could just set up an FTP server: fine-tuning a classifier is possible but not turnkey, and Jev is high quality at what it is built for.