I. How LLMs Work

A Large Language Model outputs text. It’s continually guessing the next word until it reaches a stopping point, sequentially building an answer to the prompt at hand.

When building LLMs, these answers are evaluated by humans. When the evaluator likes a response, the model is tuned to produce more of them. When they don’t like the response, it’s tuned in the opposite direction.

As is our nature, humans respond positively to long, thought-out, confidently sounding, and accurate answers. Through countless rounds of iteration, this review process produces what we’ve seen to be very intelligent chatbots that understand language and logic.

They also produce a lot of text. Alongside what you see as Claude or ChatGPT’s answer, there’s a long chain of thought where the model wrote out how it was reasoning. Sometimes you’ll see this in it’s “Thinking” state, but lot’s of it can be hidden in the final output.

Let’s say you ask, “How do you make soup?”. The LLM might show its work by adding the following text to what’s called the context window: “The user is asking how to make soup. I’m not sure what soup they want to make, so I should ask for clarification.” To which you’ll see, “What soup would you like to know how to make?”

All that text produced by the model eats up time, energy, and money. Additionally, the output is highly variable. You could get a response that’s 10 pages long or one that’s 2 sentences.

This variability is great when you don’t know what you’re looking for. We turn to LLMs with vague, ambiguous problems and let them decide what kind of output to give us.

But sometimes we know what output we’re looking for. “Is soup usually hot? True (1) or false (0)?” “Rate my essay on The Great Gatsby. Give a score from 1 to 10.”

In these scenarios, an answer containing text is inefficient when all you’re looking for is a number. If only a model could be as smart as an LLM, but be restricted to a numeric output.

II. System One Models

Such was the task a research lab called TypeSafe AI undertook when developing what they’re calling System One Models. On Sep 15, 2026, TypeSafe AI released to the public a version of this model called “Jev.”¹

Built on a similar structure that LLMs use, Jev has a similar innate ability to reason from text input. It differs, however, by outputting numbers in a format you set ahead of time.

Unlike LLMs, you can’t ask Jev a question without providing what the output should look like.

You can ask Jev three types of questions:

1. Choose an option from a list
2. Give something a score
3. Make a true-or-false judgment

First, you need to specify which of these questions you’re asking.

If you’re asking...

Choose an option from a list, you’ll need to specify the items in the list. The model will output a probability distribution across the list of options.

This is very similar to how an LLM works. Given a list of words in a long vocabulary, which word should come next? Only here, it stops after one choice, and you can be more creative in what it’s choosing from. Maybe it’s a list of ingredients, and the prompt is “What goes best with lamb ragù?”.

Next. If you’re asking...

Give something a score, you’ll need to specify a ranking of objects. The model chooses a number along that ranking, with the ability to choose in between rankings as well.

For example, “Score my essay: (1) great, (2) ok, (3) bad”.

Instead of only being able to choose (1), (2), or (3), it can output a number like 2.5, indicating it’s halfway between “ok” and “bad.”

Last, but not least, if you’re asking...

Make a true-or-false judgment; you don’t need to specify much, the output is already set. Given a prompt, it’ll respond with a number from 0 to 1, indicating the probability that whatever you’re asking is “true.”

TypeSafe AI calls these three types of questions “primitives.” It’s named them “Choice,” “Score,” and “Noul,”² respectively.

Since the underlying infrastructure behind System One Models does not output long stretches of text in the way an LLM does, it costs significantly less per prompt and is much faster. As AI becomes increasingly useful across countless domains, cost-effectiveness and speed are two very important metrics.

As a software engineer using LLMs to write code on a daily basis, these metrics are more important to me than most — optimizing token cost is now a big part of my job, and I suspect it will be for other industries moving forward.

III. Potential Limitations

Part of what you're paying for with LLMs is observability. Even if the answer is wrong, you can look through its chain of thought to see how it reasoned through the problem. System One Models omit this kind of language in their output.

This is a common tradeoff in Machine Learning: performance vs observability. Do you pick a model that's more accurate but you can't always tell why? Or a model that's less accurate but you can easily trace its steps?

I err on the side of more visibility, so I can see the "why" behind these models. Funny enough, LLMs' chain of thought is not what statisticans mean when talking about observability. That type of output is still just the output, not an entirely accurate represenation of the models inner workings.

So a model simmilar to an LLM with even less observaibilty makes me wonder how it'll hold up to peoples' bar for trust.

I’ve gotten early access to try Jev, so I'll be playing around with it. I’m excited to see what it can do.

Footnotes

1. Named after William Stanley Jevons, a 19th-century economist who argued that making resources cheaper and more efficient would increase their demand. ↩

2. Short for Bernoulli, an 18th-century mathematician whose work helped describe probability events with two possible outcomes. ↩