System 2

System 2 is the deliberate half of dual-process theory: slow, effortful, sequential reasoning that can follow rules and check its own work. Kahneman contrasts it with the fast, intuitive System 1.

In AI, System 2 usually means a model that spends variable computation per input: chain-of-thought prompting, reasoning models that generate hidden deliberation, tool-using agents, search or self-consistency sampling. These methods improve accuracy on hard, multi-step problems, at the cost of latency (seconds rather than milliseconds), token spend and less predictable output.

System 2 alongside Laya

Laya is not a System 2 model and does not try to be. It answers every question in one pass with a probability distribution over the options you supply. The useful design is to combine them:

1. Every event goes through Laya first: triage, routing, guardrail checks, moderation.
2. Decisions with high confidence are acted on directly.
3. Low-confidence decisions, or ones where the cost of an error is high regardless of confidence, are escalated to a System 2 path: a reasoning LLM with full context, or a human reviewer.

Because most real traffic is routine, this pattern keeps the expensive path for the cases that need it. Laya's own router preset goes one step further and uses a score question for request difficulty to decide which LLM to call.

One caution from the model card: the action.act_probability field is not yet a usable escalation signal (issue #185). Gate on confidence and your own thresholds. See System 1 vs System 2 AI and decision models for AI agents.