System One Models: What the Term Means
Last checked · Independent guide, not affiliated with TypeSafe AI
A System One model is TypeSafe AI's name for a model that makes fast, structured judgments that software can use directly, instead of writing text. The name borrows Daniel Kahneman's 'System 1', the fast and intuitive mode of thinking. Jev is the first and, as of September 2026, only System One model.
“System One” is not an industry standard term. It is the category name TypeSafe AI chose for the kind of model it builds, and Jev is the first model in it.
Where the name comes from
Section titled “Where the name comes from”Psychologist Daniel Kahneman popularized the idea of two modes of thinking in Thinking, Fast and Slow. System 1 is fast, automatic and intuitive: recognizing a face, sensing that a sentence is angry. System 2 is slow and deliberate: working through a proof, planning a trip. TypeSafe’s documentation says it borrowed the term to stress fast, focused judgments.
The analogy maps onto current AI like this:
| Aspect | System One model (Jev) | Reasoning model |
|---|---|---|
| Kahneman’s mode | Fast, intuitive | Slow, deliberate |
| Output | A typed judgment with probabilities | A chain of reasoning and a text answer |
| Time per call | Hundreds of milliseconds | Seconds to minutes |
| Good at | Classification, routing, scoring, yes/no checks | Multi-step problems, math, planning |
What makes a model “System One”
Section titled “What makes a model “System One””According to TypeSafe, a System One model:
- Evaluates a state. You pass it content: a message, a record, a chunk of application state. See State.
- Answers typed questions. Each question is a Noul (yes/no), a Choice (one of a list) or a Score (a level on a scale). The possible answers are fixed in the request.
- Returns probabilities, not prose. Every answer comes with a probability, and Choice and Score answers add a confidence value.
- Is trained for calibration. Its probabilities are meant to match how often it is right across many answers. TypeSafe’s training method for this is RLCD.
It does not write replies, generate code or explain its reasoning.
How a System One model fits into software
Section titled “How a System One model fits into software”The intended pattern is that your code stays in charge and asks the model narrow questions. For a refund request, for example:
- Your code gathers the customer’s message, the order’s charges and the refund policy into one state.
- It asks several independent questions in one call: Is a refund being requested? Do the charges show a duplicate? Does the policy cover this case?
- It combines the answers with ordinary checks (Is the amount under the auto-approve limit? Is the account in good standing?) and either acts, or routes the case to a person when confidence is low.
Each judgment is small enough to test, log and adjust. TypeSafe calls this approach building “prod, not God”: useful automation from many narrow decisions rather than one model that tries to do everything.
When a task is not System One
Section titled “When a task is not System One”TypeSafe’s own notes on Jev 1.13 list the kinds of work that go badly: questions with several hops of indirection, arithmetic and counting, comparing dates, and very long states padded with irrelevant detail. Those are System 2 problems. The practical rule is to do the reasoning, math and filtering in code, and give the model only the judgment that needs language understanding. See Jev limitations.
Is Jev the only System One model?
Section titled “Is Jev the only System One model?”As of September 19, 2026, yes. TypeSafe describes Jev as its flagship and the first System One model; its model list contains only Jev 1.13 under two aliases. Community projects have built open-source imitations with the same question format, but they are separate efforts; see Jev alternatives.