Episode 4 · about 7 min
Jev, its clones and your inbox
What’s likely inside Jev, how it compares with a chatbot, a classifier you train yourself and the new look-alikes, shown on a made-up email to an insurance agency.
Audio coming soon. The script and show notes are below.
With The Decision Layer desk
Show notes
- 00:00 Tom’s Hardware on the rush of Jev look-alikes
- 01:00 The problem: one service inbox, many kinds of request
- 02:00 What’s inside Jev, as far as anyone outside TypeSafe can tell
- 03:30 Jev next to a chatbot, a trained classifier and the clones
- 05:00 A made-up email with two requests in it
- 06:00 Where it breaks, and why your test set is the part you keep
Script
Cold open. Tom’s Hardware counted the rivals: Microsoft, OpenAI, Cloudflare, Perplexity, Amazon and more, all within weeks of Jev. If you run an agency, the question isn’t who copied whom. It’s which one you can trust with your inbox.
The problem. A service inbox mixes car changes, proof-of-insurance requests, billing questions, cancellations and new claims. The urgent ones wait behind the easy ones, and an email asking for two things often gets one done.
How it works. TypeSafe hasn’t published what’s inside Jev. Its primer says it adapts a pretrained language model, and one engineer’s outside probing points the same way, though he calls his own analysis speculative. So it reads like a chatbot but doesn’t write. Unlike a classifier you train yourself, it needs no labeled examples. Adding a request type is one more plain-English question.
The example. Dana bought a new car and needs proof of insurance for the lender by Friday. One yes/no question per request type catches both requests, and an urgency question catches the deadline.
Close. Models can be swapped. Your questions and a set of past emails with known answers are what you keep, so test every model on them. A licensed person still makes every policy change. We haven’t run this yet; when we do, we’ll publish everything.