By Pavan Kumar T V · 24 September 2026

BTW, the AI world moves very fast. Jev is already last week's news. This week it's Contrastive Language Models, an open-source model that answers the same three kinds of question.

Jev, TypeSafe AI's decision model, blew up this week.

I've been playing with it, looking for two kinds of use. Stuff I could use at work, and fun UI stuff just for me.

I've seen this pattern a few times now. Something at the bottom of the stack stops being slow or expensive, and people start finding uses nobody planned for. With Jev, the thing that got cheap is asking. It doesn't generate anything, it just decides. So you can afford to ask it something every time someone pauses while typing, instead of waiting for them to hit Enter.

And the place I wanted that most was the text box. Chat and search boxes are notoriously hard. You give people one empty box and they type whatever comes into their head. Most products just sit there until you press Enter, then try to work out what you meant.

Then I came across Shapeshift by Anish, a text box that turns into a calendar card, a checklist or a poll while you type. Loved it, borrowed the idea, kept going.

Three questions

I like a good mental model, and this is a nice one. Jev only answers three kinds of question:

client.systemOne({
  state: { text: 'remind me to pay rent tomorrow urgent' },
  questions: {
    card: choice('What is this', { reminder: '...', todo: '...', event: '...' }),
    urgency: score('How urgent is it', ['Not urgent', 'Somewhat', 'Urgent']),
    repeats: noul('It repeats on a schedule'),
  },
})

• Pick from a list. Which card is this? You can build the list on the fly.
• Put it on a scale. How urgent, how done, how annoyed. You write the scale yourself.
• Yes or no. Is this a correction? Same topic as before? Are they done talking?

None of them write anything. Jev picks, and plain old code fills in the details. A date parser works out what "tomorrow" means. Normal arithmetic splits the bill. So nothing you see on screen was invented by a model.

Try it

Type something in the box. Or switch to Talk and describe a process out loud. (Heads up: in Chrome, speech gets transcribed on Google's servers. If you'd rather not, just type.)

Things I tried

The coffee

This one used to drive me nuts. I type "buy milk, eggs, bread", get a nice checklist, save it. Then I remember coffee. Most smart boxes have stopped listening by then, so I'm typing the whole list again.

Here, "and coffee" gets a yes or no for every open card: do I belong to you? The shopping list says yes, and you see "Adds to" before you've hit Enter.

My first go asked it as a single pick instead: which card does this go on, or is it something new? Jev never got confident. Once I split it into a separate yes or no per card, it did.

The diagram that draws itself

Talk through an expense approval the way people actually do, which is badly. "If it's over 500, finance approves it. Otherwise the manager can. Um, so yeah, basically."

Every sentence gets one question, picked from the boxes already on the diagram: which step does this follow? A UML activity diagram builds up next to you, with a lane for each person and a diamond for "It's over 500". The "um" doesn't turn into anything, thankfully.

Actually no

People change their minds halfway through. "Actually no, finance should see it before the manager." A yes or no question catches that it's a correction. Two quick picks work out which box it's fixing and how. Then the new step slides in above the manager's box, instead of getting stuck on the end.

That's the bit that made me trust it. If corrections pile up at the bottom, you end up redrawing the thing yourself anyway.

The shrug

"minecraft diamond" is a colour, but only if you've played Minecraft. Jev wouldn't commit. I reworded the colour option and it leaned towards colour. Just barely.

Honestly, I prefer that. A model that can only pick gives you a shrug instead of a confident wrong answer, and you can design around a shrug. When it gets something wrong, you go fix the options on the list. No giant prompt to fiddle with.

At work

Same three questions, pointed at actual work:

• A sales call where the right answers show up while the customer is still explaining the problem.
• A shared inbox that sorts itself. Every email gets a team picked from a list, an urgency on a scale, and a yes or no on "does a human need to see this?" It's routed before anyone opens it.
• A support form that fills itself in while someone explains what broke.
• A meeting where the action items are already a list before anyone says "so, next steps".
• An onboarding call where the SOP is drawn by the time the new hire stops nodding.

For fun

A snake game that plays itself. Every tick, the code lists the moves that won't kill the snake, and Jev picks one. Then you sit back and binge-watch it doing a great job.

Chat boxes wait for Enter because asking used to be expensive.

It isn't anymore. So why are we still waiting?