# System 1 and System 2 Thinking System 1 and System 2 are the two modes of thinking from [[Dual Process Theory]]: System 1 is fast, automatic and intuitive; System 2 is slow, deliberate and effortful. The psychology lives in that note. This one is about how AI borrowed the vocabulary, because the metaphor now shapes how models get designed and sold, and how we combine them. Quick recap of the origin: the terms were coined by Keith Stanovich and Richard West in 2000, picked up by [[Daniel Kahneman]] and Shane Frederick in 2002, and made famous by Kahneman's [[Thinking Fast and Slow (book)|Thinking, Fast and Slow]] in 2011. Kahneman (Nobel in economics, 2002) died on 27 March 2024, right as "System 2" was becoming AI jargon. ## How AI picked up the metaphor 1. **Deep learning is System 1 (2019).** At NeurIPS 2019 in Vancouver, [[Yoshua Bengio]] gave a keynote titled "From System 1 Deep Learning to System 2 Deep Learning". His argument: neural networks are excellent at fast, intuitive pattern recognition (System 1), but lack what System 2 does: attention over a few concepts, composition, reasoning out of distribution, and causality. The related paper is *The Consciousness Prior* (2017). A few months later, at AAAI 2020, Kahneman himself sat on a panel with Bengio, Yann LeCun and Geoffrey Hinton 2. **Reasoning models become "System 2" (2024-2025).** OpenAI's o1 (September 2024) and DeepSeek-R1 (January 2025) spend inference-time compute on a long chain of thought before answering. Everybody started calling that System 2 thinking, and "System 1" ended up as shorthand for "a plain model that answers immediately". See [[AI Reasoning Models]] 3. **Someone builds System 1 on purpose (2026).** [[TypeSafe AI]] went the other way and named its category after the fast system: [[System One Models]]. [[Jev]] answers in a single pass (~100 ms), never "thinks out loud", and returns a calibrated probability instead of text. Almeida has hinted at a "ReasoningJev" later, i.e. a System 2 companion ## The mapping | | Human System 1 | AI "System 1" (e.g., Jev) | Human System 2 | AI "System 2" (reasoning models) | |---|---|---|---|---| | Speed | Instant | 70-500 ms | Slow | Seconds to minutes | | Cost | Effortless | $0.042 per million input tokens | Tiring | Thousands of thinking tokens per answer | | Good at | Recognition, snap judgments | Bounded questions: which option, yes/no, what level | Multi-step reasoning, rules, math | Math, code, multi-step problems | | Weak at | Statistics, logic, unfamiliar situations | Math, dates, indirection, [[Jagged Intelligence\|jagged]] edge cases | Speed, attention budget | Latency, cost, overthinking easy questions | ## The irony I can't stop noticing Most of *Thinking, Fast and Slow* is a catalog of System 1's failures: [[Cognitive biases|biases]], [[Heuristics|heuristics]] gone wrong, and above all *overconfidence*. Human intuition is famously badly calibrated. TypeSafe's pitch is a System 1 that's fast *and* calibrated, which is exactly the combination humans don't have. Whether they've actually achieved it is debated: Archer Hume measured an expected calibration error of 0.031 on MMLU, Rajesh Beri measured 0.107 on synthetic support tickets (details in [[Jev]] and [[AI Model Calibration]]). Kahneman's own work gives a good test for when fast judgment is trustworthy. In a 2009 paper with Gary Klein, they agreed that intuition can be trusted in environments that are regular enough to learn and that give quick feedback. That maps well onto AI System 1: bounded answer spaces, questions a knowledgeable person decides in about a second, and outcomes you can check. Outside those conditions (novel situations, multi-hop reasoning, arithmetic), send the question to System 2. ## How to apply it when building - **Route by difficulty.** Let a fast model triage and only escalate the uncertain or hard cases to a reasoning model or a human. That's [[Confidence-Gated Routing]] and [[AI Model Cascades]] - **Decompose System 2 work into System 1 questions.** A lot of apparent "reasoning" is really several snap judgments plus some arithmetic. Ask the snap judgments in parallel and do the arithmetic in code ([[Atomic Question Decomposition]]) - **Don't make System 2 do System 1's job.** Running a reasoning model to decide whether an email is spam is paying for deliberation you don't need ## Limits of the metaphor The brain analogy is loose. A reasoning trace isn't necessarily deliberate thought, and a single forward pass of a big transformer can do things that look very System 2. Kahneman himself presents the two systems as fictitious characters in a story that helps you think; he never claimed they were two separate modules in the brain. Use the metaphor for architecture decisions (fast and cheap vs slow and expensive), not as a claim about how models "think". ## References - [From System 1 Deep Learning to System 2 Deep Learning (Yoshua Bengio, NeurIPS 2019)](https://neurips.cc/virtual/2019/invited-talk/15488) - [The Consciousness Prior (Bengio, arXiv:1709.08568)](https://arxiv.org/abs/1709.08568) - [Conditions for intuitive expertise: a failure to disagree (Kahneman & Klein, 2009)](https://doi.org/10.1037/a0016755) - [System One (TypeSafe docs)](https://docs.typesafe.ai/concepts/system-one) - [Jev: System One models for Prod, not God, with Diogo Almeida (Latent Space)](https://www.latent.space/p/jev) - [Jev's Architecture Unmasked (Archer Hume)](https://archerhume.com/posts/jevs-architecture-unmasked/) ## Related - [[Dual Process Theory]] - [[Daniel Kahneman]] - [[Thinking Fast and Slow (book)]] - [[Yoshua Bengio]] - [[AI Reasoning Models]] - [[System One Models]] - [[Jev]] - [[Decision Models (DMs)]] - [[Confidence-Gated Routing]] - [[AI Model Cascades]] - [[Intuition]]