Community roundup: four real-world uses of Jev
A roundup of community replies outlining four ways people use Jev: adding classification to existing workflows, asking new questions of old data, deciding when to call an LLM, and judging each element individually.
The Jev thread! Save this one to feed to your agent. You have build patterns, real examples, and the TLDR on why it's such a big deal. So yesterday I asked what everyone is building with Jev. I’ve been reading the dozens of replies alongside other projects people are sharing, and I'm seeing a range that helps lay out where Jev is going to fit in our AI architectures going forward. Four patterns I’m seeing: 1. Add interpretation to an existing workflow: @nbarreto is categorizing 70,000 support tickets for an insurance broker, with the workflow already running on new tickets. We get multiple OOM cheaper, faster classification across an existing problem set. 2. Ask new questions of old information. @iannuttall applied eight questions to 3,282 posts to examine topic, hook, tone, and performance. Lots of variant versions around technical analysis of marketing copy, etc. Jev enables extremely fast and efficient inference across complex data. 3. Decide where deeper reasoning is worth spending. @joshest3p is building a monitor that checks changes in screen text, uses a classifier to identify potential attention needs, then asks an LLM whether an alert is warranted. The classifier and the LLM have different responsibilities within the same system. Think of it as Jev tells you where to apply intelligence, and also Jev potentially unlocking new use-cases for frontier intelligence by digging through data you'd never pay to analyze before. Cheaper means more! 4. Give individual elements their own judgment. @DanRWilloughby built a prose linter that asks ten yes/no questions per paragraph. A paragraph, a spreadsheet row, or an agent checkpoint can each have a small interpretive function attached to it, and it's so cheap it's actually worth doing. I think stepping back all of these are examples of what I’ve been calling semi-determinism: we can define the possible outcomes, but choosing between them requires deeply understanding the situation. Jev is the shim between complex inputs and a defined set of outputs. It's a new computing primitive. I’ve used classifiers before LLMs made AI sexy. What feels like a moment here is the way Jev built on an LLM pattern and attacked a different problem space both 2 OOM faster and 2 OOM cheaper. Step changes in cost and speed unlock entire new-use cases for compute and that's why I keep trying to explain to non-devs how big a deal this is. It's going to change computing for all of us. So get entirely new ways to arrange software work: • LLMs develop questions, reason through exceptions, and write. • Classifiers apply bounded judgments. • Code calculates, executes, and enforces permissions. How much useful interpretation have we left out of software because it was too expensive or cumbersome to include? These builds are starting to show us where it could go. Jev TLDR Think of Jev as being like an LLM that can only talk in multiple choice. You give it context and questions. It returns choices, scores, or yes/no probabilities. It can interpret language, but it doesn’t generate sentences. TypeSafe built it for fast, inexpensive decisions inside software: which category, how urgent, which action, does this match? The possible answers are constrained, and it's surprisingly correct over a wide range of use-cases. Look how much software looks like messy in > neat out! It's a HUGE amount of problem space. Welcome to the future, we just got a third compute primitive :)
This post quotes another post: x.com/natebjones/status/2101152547742568498