System One Models

Canonical version: System One Models.

A System One model is TypeSafe AI's name for a general-purpose decision model. It takes context and a set of typed questions and returns probabilities: which option, yes or no, what score. It doesn't generate text. The name comes from Daniel Kahneman's two systems in Thinking, Fast and Slow: System 1 is fast and intuitive, System 2 is slow and deliberate. Chat Large Language Models (LLMs) and reasoning models play the System 2 role; decision models take the System 1 one. (How AI borrowed Kahneman's vocabulary, from Bengio's 2019 keynote to reasoning models, is in System 1 and System 2 Thinking.)

TypeSafe AI coined the term when it launched Jev in September 2026, but the idea is older. Laya shipped a very similar model in 2025, and classifiers have routed intents in production for decades. What's new is the generality: one model that answers arbitrary questions over arbitrary options at inference time, with no task-specific training.

What makes them different

• One pass, no decoding. The state is read once and the answer comes straight from a readout layer. No token-by-token generation, so latency drops to tens or hundreds of milliseconds
• Shared state, parallel questions. Ten questions about one document cost roughly one document's worth of compute
• Typed output. A decision head can't return malformed JSON; retries and validators go away
• Calibrated confidence. The probability is meant to be true, so you can set thresholds ("escalate below 0.7") and actually trust them. With LLM-written confidence ("I'm 90% sure"), you can't

How you talk to one

In practice, a System One model exposes a tiny interface: a state plus typed questions, each one a Choice (which option), a Score (which level) or a Noul (yes/no probability). See System One Primitives. The design method that goes with it is to split every broad judgment into narrow questions and let code combine the answers (Atomic Question Decomposition), and TypeSafe wraps the whole philosophy under the name Machine Native Intelligence.

Where they fit

I like Nate B. Jones' framing: software now has three compute building blocks.

• Code calculates, executes and enforces rules
• System One models apply bounded judgment: messy input in, a defined set of outputs out
• LLMs reason through exceptions, write, and figure out which questions to ask

Plenty of LLM calls in today's agents are really System One questions in disguise (is this relevant, which tool, is the agent stuck, how urgent is this). Moving them to a decision model makes them cheap enough to ask everywhere.

Watch out for

• No explanation. You get a number, not a reason. Bias is harder to spot, so evals matter more, not less
• Bounded answer space only. If you can't list the options, it's the wrong tool
• Calibration drifts. Calibrated on one distribution doesn't mean calibrated on yours. Measure it

References

• Introducing System One Models & Jev (TypeSafe AI)
• Jev notes (Simon Willison)
• Nate B. Jones on Jev build patterns

Related

• Decision Models (DMs)
• Jev
• TypeSafe AI
• Laya
• Kev
• SemIf
• Jevlike
• LLM Structured Outputs
• AI Reasoning Models
• System One Primitives
• System 1 and System 2 Thinking
• Machine Native Intelligence
• Atomic Question Decomposition
• Relative vs Absolute AI Judgments
• Self-Consistency
• Reinforcement Learning for Calibrated Decisions (RLCD)
• AI Model Calibration
• Dual Process Theory

About Sébastien

Ready to get to the next level?