What If AI Didn’t Write Anything?

Imagine an AI that doesn’t chat. Doesn’t write essays. Doesn’t generate code. It just looks at a situation, weighs the options, and picks one. Fast. Cheap. No drama.

That is JEV.

It sounds almost backwards in a world where every AI tool wants to be your conversational partner. But that is exactly the point. JEV was built because the founder of TypeSafe AI, Diogo Almeida, spent years at OpenAI helping build the chat models everyone uses now. And he came to a realization that sounds almost heretical. We have been optimizing AI for human language. But computers do not speak human language. They speak structured decisions.

So he built something different.

JEV Is Not a Language Model

Let me say that again because it matters. JEV does not generate text. Not a single word.

It is a transformer-based model, yes. But it is not a large language model. It takes unstructured input, which they call a state, and returns typed probabilistic decisions. It answers questions. It does not write answers.

Think of it like a form. You hand JEV a piece of text. A customer email. A conversation transcript. A page of DOM. Then you ask it a question with predefined options. Which category does this belong to. Is this urgent. Should this go to billing or support. JEV ticks a box. It also tells you how confident it is.

Three primitives power all of this. Choice picks one option from a set. Noul returns a yes no probability between zero and one. Score rates something on an ordered scale. That is it. That is the whole toolbox.

No free text. No JSON parsing. No schema coercion. Your code just branches on the answer because the answer is already exactly what you defined.

Why This Matters for Product Designers

Here is where things get interesting for product designers.

Modern interfaces are becoming adaptive. They change based on who you are, what you are doing, and what you need right now. Should this dashboard show a chart or a table. Should this chat assistant ask a follow up question or call a tool. Should this page display a warning or stay quiet.

Every one of those decisions currently requires a call to a language model. That call takes seconds. It costs money. And it often returns something you have to parse and hope you can use.

JEV cuts that down to about 100 milliseconds. A real world example from an Angular developer showed a flight assistant dashboard where JEV selected and arranged UI building blocks in 1076 milliseconds for the entire round trip. The processing inside TypeSafe itself took 162 milliseconds. The rest was network.

The simplest way to put it is this. An LLM creates. JEV chooses. When you already have a design system full of approved components, JEV is the thing that decides which component belongs in which moment.

That is why designers need to pay attention. JEV points toward a future where design systems are not just reusable libraries. They become the infrastructure that keeps generative UI useful, accessible, and on brand. The AI handles the selection. The design system defines the options.

Other Places JEV Fits

The UI space is compelling. But JEV is already finding homes everywhere software needs fast, reliable decisions.

• Software automation and routing. Companies are using JEV to classify incoming emails, route support tickets, and check whether AI generated output is safe before it reaches users. One Vercel engineer replaced an OpenAI classifier with JEV and got results five to eighteen times faster with better accuracy.
• Guardrails for other AI. JEV can watch what a chat model is about to say and decide whether it is acceptable. Using a language model to monitor another language model gets expensive fast. JEV makes it cheap enough to actually do.
• Model routing. Not every task needs a massive frontier model. JEV can look at a request and decide whether it should go to a fast cheap model or a slower more capable one. That kind of real time sorting was previously too expensive to attempt.
• Data labeling. Feeding JEV large amounts of text and getting back structured categories is one of the simplest and most powerful use cases. No parsing. No cleanup. Just labels with confidence scores.
• Security and monitoring. JEV can scan network traffic, judge whether an action is suspicious, and make a decision before anything bad happens. Same pattern. State in, decision out.

The Bigger Picture

JEV is named after William Stanley Jevons, the nineteenth century economist who noticed that when something becomes cheaper, people use more of it. The same principle applies here. When decisions become cheap enough to automate, software will start making a lot more of them.

That is the real shift. Not a new chatbot. Not a better assistant. A new layer in software that handles the moment to moment judgments we currently either skip or hand off to slow expensive language models.

For designers, that means the interface is no longer just a static layout. It becomes a living system where decisions about what to show and what to hide happen in real time.

It is early. JEV launched in September 2026 and is already seeing adoption across developer tools and platforms. But the direction is clear. The next wave of AI is not about talking more. It is about deciding better.

And sometimes the smartest thing a model can do is stay quiet and pick the right answer.

What If AI Didn’t Write Anything? was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.