My Jev pill · three AI models · 300 use cases
Ever since Jev was released, I’ve been fascinated by where all it can be used. Any use case comes in and my first thought is: hey, why don’t we use Jev over there? So this is me sharing my Jev pill with all of you. Hopefully you’ll see what I see in this model.
Here’s what I did. I shared the exact same questions with Grok, Claude and Codex, and asked each one for 100 places Jev could be used, with an example for every one. Below is what each of them kept telling me.
Deep and specific
Broad and careful
Compact and scannable
Start here
Most AI models write. Jev, from TypeSafe AI, decides. You give it some text and a question with a fixed set of possible answers. It picks one and tells you how sure it is. That’s the whole job.
Because it never writes, it is fast (around a fifth of a second in published tests) and cheap enough to run on every message, claim or output instead of a sample. The catch: it won’t explain itself, so your code, and sometimes a person, decides what to do with the answer.
Want the deeper take? Sebastian Raschka argues it’s more than a classifier, and that how well it generalizes across tasks is the real story: It’s Easy to Dismiss Jev as Just a Classifier →
The building blocks
Jev’s makers call them Noul, Score and Choice. In plain words: a yes/no, a rating, or a pick from a list. You can ask several in one call, and each is answered on its own.
“Is this answer supported by the source?”
Use it for: policy checks, evidence checks, “needs a human?” gates.
“How frustrated is this customer?”
Use it for: severity, urgency, rubric grading, quality scores.
“Which team owns this ticket?”
Use it for: routing, categorizing, picking A vs. B.
Why it matters
Today most teams check a small sample of AI outputs, claims or tickets, because a large model reading every one costs too much and takes too long. Jev flips that. It grades all of them, settles the clear-cut cases on the spot, and hands only the uncertain ones up the chain.
That’s the pattern all three guides keep coming back to: Jev for the high-volume, closed questions; an LLM when someone needs the reasoning; a person when the decision really matters.
Honest caveat. Jev is in early access. Speed and cost figures come from vendor and partner tests, and TypeSafe doesn’t claim its probabilities are calibrated, so test thresholds on your own data before trusting them.
The three versions
The three guides weren’t written in isolation. Each model worked from what came before, which is why some parts look alike and others don’t.
Claude Fable wrote the first draft: 100 one-line ideas across 17 industries. It then expanded every card with a “how it works” paragraph and a worked example.
Codex Astra kept the first draft’s look (its stylesheet is the same) but wrote a new set of 100: five per industry across 20 industries, each marked as fictional.
Grok took Codex Astra’s intro, industry list and first three cases word for word, then wrote 97 shorter cases of its own in a lighter, JavaScript-built page.
See them whole
Open the three full guides in columns and scroll each one. On a phone, switch between them with the tabs at the top.