A new kind of AI model from a ChatGPT inventor is thrilling developers
TechCrunch Tim Fernholz ● Covered by 11 sources
An ex-OpenAI researcher built Jev, a model that gives probabilities instead of text. Developers say it’s faster, cheaper, and can’t hallucinate because it never tries to write.
Based on reporting by TechCrunch, Tim Fernholz — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Diogo Almeida says ChatGPT left him with a bad feeling. He helped build it at OpenAI, worked on RLHF, and still came away thinking the real problem was obvious: models had gotten very good at human language, but computers don’t actually speak that language. For automation, that mismatch matters. So after leaving OpenAI two years ago, he started TypeSafe AI to attack the problem from the other side.
This week the startup shipped Jev, a transformer-based model that is not an LLM. It does not generate text. Instead, it returns probabilities, which TypeSafe calls calibrated decisions. That design strips out a lot of the usual baggage. The model is cheap and fast. Users define the outputs ahead of time, so there’s no room for hallucinations. Its output tokens are free, and input tokens are billed by the billion rather than the million.
Developers noticed. TypeSafe briefly couldn’t serve users from its API because demand surged. The clearest early wins are in software automation. At Vercel, software engineer Pranit Sharma said the company swapped OpenAI’s ChatGPT Luna 5.6 for Jev in a safety classifier for commands and saw results that were 5 to 18 times faster, with better accuracy. Nikhil Mudholkar, CTO of Bryo AI, ran Jev against Gemini on business-email classification and found Gemini a bit more accurate, but 10 to 20 times more expensive. What stood out to him was Jev’s confidence score: a real probability, not a shrug in machine form.
That confidence opens a few doors. Jev can replace LLMs in some tasks, but Almeida also wants it to sit beside them as a check on bad behavior, including agent traces and jailbreaks. Armin Ronacher, CTO of Earendil, said the model pushes some of the hallucination burden back onto the user: if the probability is 50%, treat it like a coin toss; if it’s 95%, act on it. It can also help route workloads to the right model without paying LLM prices for the privilege.
The name is a nod to William Stanley Jevons and the paradox that cheaper goods get used more. Almeida’s bet is that cheaper intelligence will spread the same way, as “smart software all over the place” rather than giant apps. He’s cagey about the architecture, though outsiders suspect an open-weight LLM under the hood. TypeSafe plans more versions in new modalities. And Almeida is not pretending this is a frontier-lab religion project. He’d rather sell intelligence than fear.
My take — AI-written commentary, not fact-checked reporting
This is the kind of model the AI industry keeps pretending it invented by accident. Probabilities beat polished chat when the job is classification, routing, or safety checks, and the hype machine hates that because it’s less theatrical. The bigger tell is that the useful thing here is not a chatbot at all, which is exactly why it feels so practical.
Read more about this at: TechCrunch