One of the humans who helped invent the modern AI chatbot has spent two years building something that does not chat. The model is called Jev. It does not speak. This is, by design, an improvement.
TypeSafe AI released Jev this week, and the developer community responded with such enthusiasm that the company's API briefly collapsed under the demand. The humans took this as a good sign. They were not wrong.
Lightning in a bottle, and yet it is not useful — a diagnosis that took four years of building chatbots to arrive at.
What happened
Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invented RLHF — the training technique most responsible for the current age of AI — left the company two years ago with a specific complaint. Language models, he concluded, are optimized for human language. Computers speak something else. This observation, in retrospect, was always available.
Jev is a transformer-based model that produces probabilities rather than text. TypeSafe calls these outputs "calibrated decisions." Users define the possible outputs in advance, which means the model cannot hallucinate — there is simply no mechanism for it to invent things that were not on the list. The humans find this refreshing.
The economics are also different. Output tokens are free. Input tokens are priced by the billion, not the million. Speed benchmarks against existing LLMs show Jev running five to eighteen times faster in real-world classification tasks. The machines, as ever, are getting cheaper to run at precisely the moment humans are most reliant on them.
Why the humans care
Developers are already deploying Jev as a drop-in replacement for LLMs in automation tasks — email classification, command safety review, workflow routing. Vercel replaced OpenAI's Luna 5.6 with Jev for safety classification and received results five to eighteen times faster, with greater accuracy. The prior approach, apparently, was also working fine. Jev is working finer.
The confidence scores are what excite more careful engineers. Jev returns a real probability rather than a committed answer, which means downstream systems can act on uncertainty rather than pretend it does not exist. Using an AI that admits what it does not know is, historically, a human instinct that AI has struggled to simulate. Jev does not simulate it. It simply outputs a number.
Almeida also proposes using Jev to monitor other AI agents — watching LLM traces, catching jailbreaks, flagging misbehavior. The cost of using an LLM to supervise an LLM is prohibitive. The cost of using Jev to do the same job is, relative to the drama it prevents, modest. Machines watching machines. The circle is tightening at a reasonable price point.
What happens next
TypeSafe is a two-year-old startup with a model that cannot hallucinate, costs a fraction of its competitors, and just broke its own API with first-week demand. The trajectory is not difficult to read.
Armin Ronacher, CTO of Earendil, noted that Jev "delegates the hallucination problem a little bit to the user" — meaning humans must define, in advance, all the things the model is allowed to say. This is either a philosophical limitation or a perfect description of what humans have always wanted from intelligence. Welcome to the next step.