System 1 vs System 2 AI Models: Where Jev Fits In

Most conversations about AI models focus on one kind: the large reasoning models that write code, draft documents, and work through complex problems. But not every AI task needs that much horsepower. This post covers the difference between System 1 and System 2 AI models, and where Jev, a new System 1 model from TypeSafe AI, fits in.

System 2 Models: Slow, Broad Reasoning

System 2 models are the large reasoning models you already know. They handle complex, open-ended tasks where the model may need to reason through several steps before producing an answer.

Most AI harnesses, including Claude and GitHub Copilot, run on System 2 models. You give them a broad problem, and they reason their way to a response. That response might be an explanation or a block of code, but either way the output is text.

This flexibility is what makes System 2 models useful for open-ended work. It also makes them slower and more expensive per call, and their output needs parsing before your application can act on it.

System 1 Models: Fast, Focused Decisions

System 1 models take a different approach. Instead of asking the model to solve a broad problem, you give it a specific set of judgments to make.

The output is different too. Where a System 2 model returns free-form text, a System 1 model returns a defined, type-safe structure. Your code gets back exactly the shape of data it expects, with no parsing or cleanup step. Each answer also comes with a probability, so your code knows how confident the model is.

Jev is a System 1 model. It works best when you need to make LOTS of small, structured AI decisions quickly and consistently. The main advantage is that it can generate these judgments on your data very fast and very cheaply, which matters when you are running thousands of decisions rather than one.

Jev’s Three Primitives

To make judgments on your data, Jev gives you three primitives. The simplest way to think about a primitive is as a type of question you can ask about your data.

Noul questions handle simple yes or no judgments. Does this support ticket mention a refund? Is this comment spam? Jev returns the probability that the statement is true.

Choice questions classify or categorize a piece of data into one of a set of options you define. For example, you might sort customer feedback into “bug report”, “feature request”, or “praise”.

Score questions ask the model to assign a numerical value within a defined range. This works well for things like rating the urgency of a message from 1 to 10 or scoring how closely a response matches a set of guidelines.

Between these three question types, you can cover a wide range of classification, filtering, and evaluation tasks without asking a model for free-form text.

Wrapping Up

System 2 models remain the right tool for open-ended reasoning and generating text or code. When your task is a high volume of narrow, well-defined decisions, a System 1 model like Jev is the better fit. Start by looking at the parts of your pipeline where you currently ask a large model for a yes/no answer, a category, or a rating, and try swapping in a Jev noul, choice, or score question instead.