Faster AI decisions with Jev open the way for more automation

The AI model Jev writes no text. It takes in a piece of material and questions whose possible answers are fixed in advance. Jev then replies with one of the options and a number for how sure it is. A program can then use the answer directly, and for bounded decisions of that kind Jev is built to be considerably faster and cheaper than using an ordinary language model.

At the cloud company Vercel, where developers connect their programs to AI models from different providers, nearly 13 percent of paying customer teams had used Jev after one day, the company writes. No earlier model has spread as widely there during its first day, and no other recently launched model has even reached 7 percent in the same time.

Jev was released on 15 September by the company TypeSafe AI. The material can be a customer email, for example, and the questions which department should handle the case, how urgent it is and whether a person needs to read it. Jev answers all the questions at once. After that, ordinary program code decides what happens, for example that everything where the model is less than 95 percent sure goes to a person.

According to TypeSafe's own measurements, Jev responds in between 70 and 500 milliseconds, and the company charges only for what is sent in, not for the answers. TechCrunch has spoken to two developers who have tried the model. At Vercel they replaced a model from OpenAI that judged whether commands were safe to run, and got answers 5 to 18 times faster and with higher accuracy. A developer at another company had Jev and Google's Gemini sort business emails. Gemini was slightly more accurate but 10 to 20 times more expensive.

TypeSafe writes that Jev cannot hallucinate. By that the company means that the answer can never fall outside the options the developer has set up. TypeSafe's own documentation says that Jev reads questions literally, neither counts nor compares dates reliably and can be swayed by text written to mislead it. The developer Armin Ronacher tells TechCrunch that the model hands part of the hallucination problem over to the user, who has to decide how high a probability is needed to trust an answer.

The answers from Jev go to the program, not to the person using the service. The model makes small decisions in the background. According to TypeSafe this can mean going through large amounts of data, making decisions in real time inside a program or checking what another AI model has answered. How sure an answer has to be to count is decided by whoever builds the service. It remains to be seen whether Jev's initial popularity lasts.