System 1

System 1 comes from dual-process theory in psychology, popularised by Daniel Kahneman's Thinking, Fast and Slow (2011). System 1 is fast, automatic and effortless: recognising a face, reading a word, sensing that a sentence is hostile. System 2 is slow, deliberate and effortful: multi-step arithmetic, weighing a contract clause.

In AI the terms are used by analogy. A System 1 model maps input to decision in a single, bounded computation. A System 2 model spends variable compute, typically by generating intermediate reasoning, before answering. Most production pipelines need both: cheap, fast calls on every event, and expensive deliberation on the minority that deserves it.

Laya as a System 1 model

Laya describes itself as a System 1 decision model, and the API endpoint is literally named /v1/systemone after the Jev protocol. The properties that make it System 1:

• One forward pass for every question in the call, with no generated tokens. Single-question latency is 32.8–39.5 ms on a T4 GPU.
• A fixed answer space you define per request: a label, a level or a yes/no.
• Calibrated uncertainty as the signal for when System 1 is not enough.

That last property is the bridge. When confidence is low, the right move is to hand off to System 2: an LLM, a rules engine with more context, or a human. See escalation and act/escalate routing.

The analogy also sets honest expectations. System 1 is prone to confident errors on unfamiliar input (Laya's English checkpoint on non-Latin scripts is a stark example), so it needs guardrails like language routing. See System 1 vs System 2 AI.