What If the Best AI Doesn’t Need to Talk?

For the last few years, we have started to treat intelligence and language almost as the same thing.

Need to classify a support ticket? Use an LLM.

Need to decide which tool an agent should call? Use an LLM.

Need to review a transaction, route a request or check whether something looks risky?

Again, use an LLM.

It makes sense. Large language models are flexible, powerful and increasingly good at handling tasks that used to require separate systems.

But there is an interesting question hiding underneath all of this:

Does every intelligent decision really need language?

Decisions, not paragraphs

TypeSafe AI recently introduced Jev, a model designed around a different idea.

Instead of generating text, Jev returns structured decisions. A choice. A score. A probability.

You can give it a support request and ask which team should handle it. You can ask whether an action looks risky. You can ask how confident it is about a particular outcome.

The answer is not a paragraph.

It might simply be:

Billing — 87%

That sounds like a small difference.

I think it is a bigger one.

Most software does not need an AI to sound intelligent. It needs the AI to make a useful decision that another system can act on.

TypeSafe calls this a “System One” model: fast, structured and built for decisions rather than conversation. The company says Jev is trained to return calibrated probabilities and typed outputs, with the possible answers defined in advance.

That means the system cannot suddenly invent a new output format or respond with an unexpected paragraph.

It has to choose from the world you gave it.

Maybe AI will become less visible

Our current idea of AI is strongly connected to conversation.

We type something.

AI answers.

We read the answer.

But the AI systems that become most common may eventually look very different.

A security system may decide whether an action should be blocked.

A customer request may be routed before anyone sees it.

An agent may choose which model or tool should handle the next step.

A transaction may be sent for review because its risk score crossed a threshold.

No chatbot. No conversation. Maybe no visible AI interface at all.

The most important AI systems might eventually be the ones we never talk to.

This is already starting to appear in real systems. TechCrunch reported that Vercel tested Jev as a safety classifier for commands and said it returned results five to eighteen times faster than the LLM-based system it replaced, while another developer found it significantly cheaper than a generative model for email classification.

Those are early tests, not universal benchmarks. But the direction is interesting.

Maybe some AI tasks do not need a smaller chatbot.

Maybe they need a different kind of intelligence.

Removing language does not remove uncertainty

There is an important catch.

TypeSafe describes Jev as unable to hallucinate because its outputs are restricted to predefined types.

That is useful, but we should be careful with the word.

If a system can only choose between approve, review and reject, it cannot suddenly invent a fourth answer.

But it can still choose the wrong one.

Removing generation does not remove uncertainty.

It changes where the uncertainty appears.

With a generative model, we often ask:

What will the model say?

With a decision model, the more important questions become:

How confident is the decision? What threshold should trigger human review? What happens when the model is wrong?

That may actually be a healthier place to have the conversation.

A system that says “87%” gives us something we can design around. We can create thresholds, escalation rules and different levels of authority.

The uncertainty does not disappear.

It becomes visible.

One model does not need to be everything

There is another idea here that I find even more interesting.

The future AI stack may not be one giant model handling every problem.

It may be a collection of different kinds of intelligence.

A decision model handles fast routing and scoring.

A language model handles explanation and complex reasoning.

A tool performs the actual action.

A human steps in when judgment or accountability matters.

Each part does what it is good at.

That feels different from the current instinct to place a larger language model in the middle of every workflow.

And maybe that instinct is worth questioning.

Language is powerful because humans think, coordinate and communicate through it.

But software does not always need a conversation.

Sometimes it needs a probability.

Sometimes it needs a category.

Sometimes it needs a simple yes or no.

Choosing the right kind of intelligence

So the lesson I take from Jev is not that language models are the wrong direction.

They are clearly useful, and many problems genuinely need language and reasoning.

The lesson is simpler:

We should stop asking one model to be every kind of intelligence.

A thoughtful AI system should not use the most impressive model available just because it can.

It should use the form of intelligence that fits the decision.

That can mean a frontier LLM when the problem is open-ended and complex.

It can also mean something much smaller, faster and more constrained when the decision is narrow.

Responsible AI is often discussed as making models safer.

Maybe there is another part of it:

choosing the smallest form of intelligence that can responsibly do the job.

The best AI does not always need to explain itself in a beautiful paragraph.

Sometimes, the best answer is simply the right decision.

See you in a Thoughtful Future.

References

TypeSafe AI (2026). Introducing System One Models & Jev.
https://typesafe.ai/blog/introducing-system-one-models-and-jev

Fernholz, T. (2026). A new kind of AI model from a ChatGPT inventor is thrilling developers. TechCrunch.
https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/