News - TypeSafe AI launches Jev, a “System One” model built for decisions instead of text
Daniel Kahneman in his book "AMAZON - Thinking fast and slow, ARCHIVE.ORG - Thinking fast and slow" described human thinking as a combination of a fast, intuitive System 1 and a slower, deliberate System 2. Modern AI systems may evolve in a similar direction—not because models “think like humans,” but because the architecture is economically sensible. Use fast specialized models for routine decisions, and invoke expensive reasoning models only when the problem actually requires deliberation.
LLMs are optimized to generate strings for people. Jev is optimized to make typed decisions for software. Instead of returning text, Jev accepts some application state plus bounded questions and returns structured decisions such as:
- Choice — select one option from a defined set
- Score — place something on an ordered scale
- Noul — return a yes/no probability
The output includes probabilities/confidence so application code can decide whether to act automatically or escalate.
- Traditional LLMs = state -> LLM -> text / JSON / explanation -> parse / validate / retry -> software action
- Jev / System One = state + typed question -> System One model -> typed decision + probability -> software action
TypeSafe AI says Jev uses a new architecture, a parallel sampler and a training approach it calls "Reinforcement Learning for Calibrated Decisions (RLCD)". It's stated goal is not fluency but calibrated, machine-consumable judgment. Jev can answer System One-style tasks in roughly 70–500 ms and currently charges $0.042 per million input tokens, with no output-token charge. It claims around two orders of magnitude improvement in speed/efficiency for its targeted workloads compared with general LLMs, but this is currently not yet verified by independent measurements.
Important: TypeSafe AI uses phrases such as “can't hallucinate” or “zero hallucinations”. Jev can be type-safe (expected: HIGH | MEDIUM | LOW = it can't return "probably medium-ish...") so it may eliminate output-shape hallucinations, but it cannot eliminate semantic errors. This distinction is very important, so for example "Fraud probability = 0.97" is perfectly typed, but it can be still wrong.
Summary1: Today's agent stacks often use expensive generative models for trivial decisions such as routing, scoring and gating. Jev proposes a different architecture: use extremely fast specialized models for routine semantic decisions and reserve frontier LLMs for difficult reasoning
Summary2: The future agent stack may not be one giant LLM. It may be a hierarchy of intelligence: deterministic code → System One models → frontier reasoning models.