System 1 / System 2 (Dual-Process Thinking)
System 1 / System 2 is the cognitive framework psychologist Daniel Kahneman laid out in Thinking, Fast and Slow (2011):
- System 1: fast, intuitive, automatic, parallel, low-energy. Recognizing a face, driving, answering "2 + 2 = ?"
- System 2: slow, deliberate, serial, high-energy. Solving equations, chess positions, reasoning about "what if…"
Humans spend ~95% of time in System 1 and ~5% in System 2; but System 2 is lower-error, so we deliberately switch into it at critical moments.
The AI analogue
In 2024-2025, AI labs began explicitly borrowing this framing to classify models:
| Paradigm | Example AI models | Characteristics |
|---|---|---|
| System One | TypeSafe Jev, Mistral 7B / Mixtral, etc. | fast, parallel, structured output, single-task |
| System Two | OpenAI o1 / o3, DeepSeek R1, Claude thinking mode, QwQ | slow, serial, long chain-of-thought, complex reasoning |
TypeSafe AI's September 2026 white paper explicitly named its first model a System One Model — "designed for the small decisions inside software, not for conversations," end-to-end 70-500 ms.
OpenAI took the opposite route with o1, doubling down on System Two: the model thinks at length before answering, end-to-end 3-329 s.
Limits of the analogy
Kahneman himself warned about several mismatches:
- Human System 2 energy: the brain consumes ~20% of body energy
- AI System 2 energy: reasoning takes 100% of one GPU; more tokens = more cost
- Human System 1 "intuition": essentially System 2 internalized and automated
- AI System 1: more likely a different training objective than "internalized System 2" — Jev was designed System One from scratch, not pruned out of a big LLM
Why the analogy matters
- Splits "AI model" from a single category into multiple
- Stop stuffing "fast judgments" and "slow reasoning" into one model
- Pick the right "system" for the application
- Re-prices everything: System One models (fast, cheap) do high-volume classification / routing / guardrails; System Two models (slow, expensive) do critical decisions / math / proofs
- Re-divides labor: future software will have an "AI conductor" — System One schedules System Two (Jev decides "this one needs o1"), System Two solves the hard stuff
Future directions
- A System One model matrix: TypeSafe has shipped Jev (classify / route) and plans more modalities (image, video)
- System Two models go open: DeepSeek R1, QwQ etc. let small teams use them
- System Three?: not named yet, but "model-coordinating-models" meta-controllers are already emerging