System One models

A model category from typesafe-ai, named for the fast intuitive judgments in Daniel Kahneman’s Thinking, Fast and Slow. System One answers without deliberation, the part of the brain that sees a shirt and knows it is blue. System Two deliberates. In TypeSafe’s framing, System One models like jev take the first kind of task and reasoning LLMs keep the second.

Properties as described in jev-is-incredible:

• Input is application state plus a schema. Output is typed JSON with a confidence value on every decision.
• The shape is guaranteed and the content is probabilistic. Theo’s split: format deterministic, answer reliable but not deterministic.
• Speed and price of a function call rather than a conversation: 70 to 500 ms classification at about 4 cents per million input tokens. Output is free.
• The name Jev comes from Jevons’ paradox, the finding that a more efficient steam engine increased coal consumption because cheaper power got used for more things. The company reads Jevons the same way: cheap judgment gets used everywhere, and that is the business case.

The category excludes reasoning tasks

Judging between reasoning-model outputs is out, and so is compacting a conversation. State does not carry between calls, so long-running jobs need something else. Theo’s 10-second test marks the line. Under 10 seconds after perceiving everything, System One fits. Past that, the task needs a reasoning model. See structured-outputs for the reliability problem the category solves, and human-ai-synergy for where fast machine judgment sits in a combined system.

The Graepel counterpoint

conflicts with graepel-2026-llms-dont-reason

The framing this page reports keeps deliberation on the reasoning-model side. System One judges fast, and the reasoning LLMs carry the slow work. thore-graepel rejects the second half of that division: chain of thought, on his account, is the same next-token process iterated longer, so no current model reaches System 2 in Kahneman’s sense, and scaling the networks does not produce one. See chain-of-thought-faithfulness for the evidence behind the charge and epistemic-state for the architecture he proposes in its place. The two positions share the daniel-kahneman frame and disagree about who gets to claim the deliberate half of it.