Jev, explained.
A new kind of AI model from TypeSafe AI, in nine drawings. Dotted words explain themselves. Scroll, or let it roll.
AI learned to talk. Code does not speak essay.
Ask a chat model a yes-or-no question and you get a paragraph. A program cannot branch on a paragraph. It needs a value.
Jev is an if-statement with judgment.
It is a model from TypeSafe AI that does not write. Messy text goes in. A decision comes out, in a form your code can use as it is.
You write the questions, and the only answers allowed.
Yes or no. One of three. A number from 0 to 1. Jev cannot answer outside those shapes, so your code never has to guess what it meant. That is the type-safe part.
A chat model writes word by word. Jev answers all at once.
Text comes out one token at a time. Decisions do not have to. In TypeSafe's own test the same questions took 8.6 seconds one way and 0.11 the other. Fast enough to sit inside a loop.
Every answer says how sure it is.
The number is calibrated. You draw the line. Above it your software acts alone. Below it a person gets asked. Try three messages.
A message arrivesUse it where you would have written a fragile if.
Route a ticket. Sort a document. Score a lead. Check another AI's answer. Millions of small decisions, in a tenth of a second each, for $42 per billion tokens read.
Do not ask it to write.
No emails, no chat, no code, no open questions. If you cannot list the possible answers in advance, use a language model and keep a person near. And it is early: the speed and cost figures are TypeSafe's own.
Let models write. Let Jev decide. Let people take the rest.
Jev is in early access.
Nine drawings. One decision at a time.
Drawn by Bolted. Independent, not sponsored.
A reel that teaches.
The fourth reel, and the first made to explain somebody else's technology to a person one or two years into the work. Two additions to the format came out of that job.
| New · the pen | The drawing builds itself as the light arrives. Nothing is on the sheet before it has been explained, so the eye is never ahead of the lesson. The same feature lets the light draw: at the end it traces TypeSafe AI's mark in one stroke, in their colour. |
|---|---|
| New · the gloss | Words that explain themselves. Dotted words in the titles open one plain sentence on hover, focus or tap. Press G for study mode, with every one open. |
| The lesson | The problem, what it is, how it works in three steps (shapes, all at once, how sure), when to use it, when not to, and whose it is. The instrument from the last reel returns: three messages, one line, and a switch between acting alone and asking a person. |
| The facts | Everything about Jev comes from typesafe.ai and its launch post of 15 September 2026, including the 8.6 s and 0.11 s, which are TypeSafe's own test and which they describe as the high end. The mark was measured from their icon. Notes in JEV.md. |
| The controls | Roll plays it in one minute. The chapter bar, 1 to 9 and the arrow keys jump. Space rolls, N opens the director's notes, G opens every gloss. |