A former OpenAI researcher has released Jev, a radically different kind of AI model that abandons text generation altogether. Instead of writing answers token by token like ChatGPT, Jev returns structured decisions with confidence scores in a single query. The model represents a bet that much of today's AI traffic is wasted on generating prose just to extract a simple yes-or-no answer buried inside it.

Diogo Almeida, who co-invented both RLHF (Reinforcement Learning from Human Feedback) and ChatGPT at OpenAI, launched Jev through his company TypeSafe AI on September 15. The company has raised $40 million. The model's name comes from Daniel Kahneman's concept of System 1 thinking, the fast and intuitive mode of decision-making, as opposed to System 2's slow, deliberate reasoning.

What Makes Jev Different From Chat Models?

Traditional large language models (LLMs) work by predicting one word at a time, generating complete sentences or paragraphs that humans then parse for the actual decision they need. Jev skips that entire step. You give it the state of the world and a set of questions, and it returns typed answers with calibrated probabilities, no parsing required.

The training method behind Jev is called RLCD, or Reinforcement Learning for Calibrated Decisions. TypeSafe positions it alongside RLHF, which optimizes for human preference, and RLVR, which optimizes for verifiable rewards. RLCD targets "calibrated decisions: answers with epistemically honest probabilities." Calibrated has a concrete meaning: when the model says 0.8, it should be right about 80 percent of the time.

The second key innovation is a "parallel sampler" that generates all outputs in a single query instead of going token by token. This dramatically reduces latency and cost. TypeSafe has not yet published the algorithm or architecture behind either piece.

How to Use Jev for Real-World Decision-Making?

Jev works with three primitive question types that developers can compose in code:

• Choice: Pick one option from a list of 1 to 255 possibilities, with a confidence score attached.
• Score: Place input on 2 to 10 ordered levels and return a fractional score, useful for ranking severity or priority.
• Boolean: Answer a yes-or-no question with the probability that the answer is true.

Consider a support ticket routing scenario. A customer writes in with a problem. You want to know which team should handle it, how blocked the customer is, and whether they are asking for a refund. With a chat model, you would send a prompt, receive text, parse it, validate it, and retry when validation fails. With Jev, you send the ticket text plus typed questions and get back a distribution over your options with confidence scores. The model might return {"billing": 0.08, "technical": 0.85, "sales": 0.07} with a confidence of 0.82. Your code then decides what 0.85 is worth.

Jev is now available on Vercel's AI Gateway as of September 16, with the AI SDK 7 as the client. Direct TypeSafe API keys are still in early access. The model integrates with TypeScript through the @ai-sdk/typesafe-ai provider.

What Does "Zero Hallucinations" Actually Mean Here?

TypeSafe's homepage claims a 0 percent hallucination rate, but the company adds an important caveat: "Our number is not empirical. Schema matching is guaranteed." That second sentence is the entire guarantee. Ask for one of billing, technical, account, or other, and TypeSafe guarantees one of those four comes back. No malformed output, no invented fifth team.

However, a schema guarantee stops malformed output and does nothing to stop the model from picking the wrong option. The model could still misclassify a billing ticket as technical support. What Jev adds, by TypeSafe's account, is a calibrated probability next to the answer, at far lower latency and cost than existing structured-output modes on major LLM APIs.

Why Does This Matter for AI in Production?

Most AI calls in production today are not generating creative content or writing essays. They are making routing decisions, classifying inputs, or scoring options. Almeida's observation is that a large share of AI traffic is essentially an if statement waiting to happen. You ask for one of four categories, and most of the time you get a usable answer, but occasionally you get something that requires a retry, a parser, a schema check, and human review.

Jev's approach eliminates that friction. The probability is what makes this different from a simple boolean flag. A billing ticket with a refund probability of 0.95 goes straight to the refund desk. One at 0.05 stays in the billing queue. Anything in between lands with a person, because that middle band is where an automated refund decision costs you most. Developers set these thresholds from a few hundred tickets their team has already labeled.

The philosophy behind Jev is stated on TypeSafe's documentation in one line: break judgments down into "atomic questions, composed in code." Every question is one of the three primitives, and several can be evaluated in parallel in a single request. This represents a fundamental shift in how AI is deployed in production systems, moving away from text generation as the default interface and toward structured decision-making as the primary use case.