Should AI models answer every question like a chatbot?

Last week, TypeSafe AI announced its first model: Jev.

When I first heard about it, I almost shrugged it off as “just another LLM.” But it was nothing of the sort; I was impressed. It's a decision model (with open-source alternatives too, like Laya). And it's very good.

Jev doesn't generate text. You give it a situation (plain text or JSON) plus a few questions, and it returns a decision directly. That decision can be one of several options, a score, or a yes/no probability. It also attaches a confidence score to every answer.

One developer described it as a “smart if” — and I think that's a spot-on analogy.

The name comes from Kahneman's “Thinking, Fast and Slow”: “System 1,” the fast, intuitive kind of thinking. The model name is also a nod to the Jevons paradox: the cheaper something gets, the more it's used. If making a decision becomes cheap enough, you'll make ten decisions a second.

The numbers are boldly ambitious: $0.042 per million input tokens, output free. Response times between 70 and 500 milliseconds. No more handing classification tasks to LLMs and burning tons of tokens!

The real value: the little decisions users never see

What interests me most is this: the little decisions inside your software that the user never sees. Is this content appropriate? Which flow should this user go to? Is this comment spam? None of these need a paragraph. A single answer plus a “how sure am I” is enough. If confidence is low, hand off to the big model; if it's high, carry on. Don't send everything to an LLM — in fact, let Jev decide what to send.

A note of caution, of course. The “no hallucinations” claim means this: it can't return an answer outside the schema. It doesn't mean it never gives a wrong answer. The company itself acknowledges that the performance figures it shared may be the best achievable in the real world.

The direction is clear to me: the era of dumping every task on a single giant model is slowly ending.

Fast work, a fast model; deep work, a deep model.

So how are you solving these small decisions in your product right now? Are your AI consultants still saying “an LLM for everything”?