Meet Jev: The AI Model That Doesn’t Chat, It Decides
A new kind of model from TypeSafe AI promises faster, cheaper decisions inside software.
Most of us think of AI as something we talk to. We type a question, and a large language model (LLM) writes back. But a lot of the AI work inside real products isn’t conversation at all. It’s small, repetitive decisions: Which team should get this support ticket? Does this action need a human to approve it? How urgent is this request?
A new model called Jev, from San Francisco startup TypeSafe AI, is built for exactly that kind of work.
What is Jev?
Jev launched in limited early access on 15 September 2026, alongside a US$40 million seed round led by DCVC. TypeSafe was founded in 2024 by Diogo Almeida, who previously worked at OpenAI on RLHF, InstructGPT, ChatGPT and GPT-4, together with Erik Gafni and Sasha Sheng.
Jev is the first of what the company calls “System One” models, a nod to the fast, intuitive System 1 thinking described by Daniel Kahneman. Unlike an LLM, it does not write text. It reads a situation and returns a structured answer, along with probabilities and confidence scores, that other software can use immediately.
How does it work?
- Give it the context.This could be a ticket, a message, a set of form fields or a JSON object.
- Ask typed questions. For example, pick one option from a list, give a score, or answer yes or no.
- Act on the result. Your code receives a clean, structured answer it can branch on, sort or threshold, with no text to parse.
One developer described it as a “smart if statement”. Normal code can check whether an order total is above a number. Jev lets code branch on judgments that need understanding, such as whether a message sounds urgent.
Why it matters
Where it could be used
- Routing: sending tickets to the right team, or choosing which model or tool should run next.
- Approvals and safety checks: flagging risky actions before an AI agent carries them out.
- Scoring: ranking urgency, priority or risk.
- Real-time control: early demos include a Minecraft bot, a driving simulation and a simulated drone.
Our take
Sources & References
-
TypeSafe AI — Introducing System One Models & Jev
Primary source for Jev, System One models, RLCD, and TypeSafe AI's design approach. Safe AI
-
DCVC — TypeSafe Emerges from Stealth with a New Way of Doing AI
Source for the September 15, 2026 announcement and $40 million Series Seed investment led by DCVC. DCVC
-
TypeSafe AI — Team
Information about the company's founders and their professional backgrounds. Safe AI
-
Tom's Hardware — TypeSafe AI's Jev Offers an Alternative to LLMs
Independent technical reporting on Jev's architecture, use cases, and claimed performance advantages. Hardware
-
TechCrunch — A New Kind of AI Model from a ChatGPT Inventor Is Thrilling Developers
Background on Diogo Almeida, TypeSafe AI, and the motivation behind Jev. TechCrunch
-
Financial Times — The Cheap New AI Model Taking Aim at OpenAI and Anthropic
Independent reporting on Jev's launch, market positioning, and early industry interest. Financial Times
-
Early Independent Research — Typed Decision Models: An Early Evidence Audit and Evaluation Checklist
Research discussing the current state of independent evaluation and limitations of early evidence. arXiv
Comments are closed