I was impressed by Jev, please explain why I shouldn't be
HI version is available. Content is displayed in original English for accuracy.
I showed them an innocuous working example of how they could classify thousands of participant entries as "mentioning Italian food" whether it mentioned "pasta", "penne", "rigatone", etc etc. Jev did this task with high confidence, and showed it could differentiate between the inverse, eg had very low confidence when I switch entry to mentioning "hamburgers". I find that a very useful tool to pass on to non-technical colleagues.
But on HN, there are lots of much more qualified people saying Jev is a con, a step backwards, or bonkers that people are impressed. Will one of them tell me plainly what software was doing this prior to Jev, and the like? I don't doubt it existed, but I never came across it. I'm interested to hear from people more expert than myself.
To be precise, software must be able to: - Classify thousands of entries with arbitrary themes or topics, even if entry doesn't mention that theme or topic directly, with high enough confidence to be useful. - Do it with thousands of entries in 0-3seconds. - Do it at negligible cost, eg sub 0.1cents
Discussion (12 Comments)Read Original on HackerNews
But what Jev actually has done that's new and interesting is that it allows you to define a new semantic decision at runtime in natural language, without training a model for that task, and get a fast, cheap, bounded, probability-bearing result designed to be consumed directly by software.
In other words, you don't have to fine tune a model. You should if super high quality is really important to you -- but you should also do that with other things that people reach for LLMs for as well. But for the rest of us for whom good enough is, well, good enough, and who aren't interested in training or fine-tuning models? Jev unlocks a lot off-the-shelf.
You should be impressed. It's doing something more novel than the haters suggest.
In this case, its people who have built BERT classifiers for years, pointing out you can just fine-tune some model to classify. Of course that's true, and still have value, but its not exactly turnkey. And Jev is surprisingly high quality at what its built for.
1 - https://news.ycombinator.com/item?id=9224
If anybody has pointers on how to make Jev code in any language I'd like to know
You don’t need permission to feel how you feel.
As William Gibson said, the future is already here, it’s just not evenly distributed.
So people get excited about existing technologies because they are new to them.
Wrapping an existing technology in a new package is a common way to introduce it to new people…there are many wrappers around ffmpeg, html2pdf, etc. that are marketed as products.
And they seem like magic…hell cassette tapes feel like magic to me even though I had a cassette player in 1973.
Often but not always, the best tool is the one in your hand or the one you know of because then you can just get to the work you are trying to do.
First make it work, then make it good. Good luck.
BeRT, BART and FLAN-T5, for text classification. All of those models are ~half a decade old now, are based on the transformer model, and small enough to run locally. They're not perfectly SOTA, but perform quite well in embedded applications. Fine for low-stakes stuff, I think.
Image classifiers are a hugely competitive space. You've got BeiT, DeiT, MobileOne, ConvNext, FastViT and several more models that all support image classification pipelines.
So other than untestable claims of _how_ Jev generates responses, it's main feature right now is its API design?