Isaac Flath 用 Jev 做的六件小事
作者列出 60 天后仍会用的 Jev 用法：脚本事实核查、新闻排序、PDF 找文本、查引用、归类评审笔记、用 trace 评估 agent 失败原因。
I've been using Jev by @typesafeai Here's the six things i've tried and am confident I'll still use Jev for 60 days from now. There's many more experiments, ideas, and things I think I will use it for. It's a big deal (more on why in next post). But I am only sharing things that I am 99% sure will lead to stuff I will still be using Jev for in 60 days. That means I started with small, boring, but useful, stuff. - Fact-checking my scripts - Ranking my news feed - Finding the right text in PDFs - Checking citations - Grouping my review notes - Figuring out why agents fail (eval over traces) https://isaacflath.com/writing/six-things-i-tried-with-jev
I think the new Jev model by @typesafeai is a very big deal. I think of it as : LLMs are made for chat, and we have workarounds to use them for software logic and workflows. Basically we ask this general-purpose model to write JSON for us. All that can be malformed, and it often is. It often isn't consistent, so we have a lot of workarounds for that. And we need to use more expensive and slower models than feels like should be necessary, and all that adds cost and latency. Jev has the general-purpose intelligence like LLMs, but is scoped down to classification tasks. What that means is it can be a very, very fast general-purpose classifier. There are a ton of use cases where training and fine-tuning a classifier was worth it. There's a whole bunch where training a classifier would be very useful, but creating the dataset and then evaluating it and all that was just much higher lift than was worth it for an experiment. With Jev, you just do that everywhere. You don't have to fine-tune a model, and it's still pretty fast. I think this is a huge deal because most software is a series of if statements or switch statements where we're trying to decide which path to go down. So much so that people have been able to use Jev to drive browsers, play video games, write text via heavy classification and to do all sorts of things. Those are demos today, and TBD on how it all shakes out for which use cases. But the point is that having a strong classifier that you don't have to fine-tune can do A LOT.