TypeSafe AI releases Jev, an AI model built for software decisions

In short: TypeSafe AI has released Jev, a new AI model that does not write text and is designed to help software make clearer yes or no style decisions.

What happened

TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, released a new model called Jev. Almeida previously worked on ChatGPT and a training method called RLHF, which is short for “reinforcement learning from human feedback” (teaching an AI using people’s preferences).

Unlike chatbots, Jev does not produce paragraphs of text. Instead, it returns probabilities, meaning it tells you how likely something is. Think of it like a dashboard light that says “90% chance this is safe” rather than a chatbot that explains its reasoning.

TypeSafe says this approach can reduce “hallucinations,” which is when an AI makes things up. Because users define the allowed outputs ahead of time, Jev is meant to stay within those options.

Developers have been testing it for software automation tasks, like classifying emails or checking whether a command is safe to run. TechCrunch reported that demand was briefly so high that TypeSafe’s API, which is the connection developers use to plug a tool into their software (like a power outlet for apps), could not serve all users. Vercel engineer Pranit Sharma said swapping an OpenAI model for Jev made one safety classification task 5 to 18 times faster, with better accuracy.

Why it matters

Many people experience AI through chatbots, but lots of business uses are simpler, like sorting messages, flagging risky actions, or routing a request to the right tool. If models like Jev are cheaper and faster for those jobs, companies may add more automated checks in everyday software, including tools people use at work.

Source: TechCrunch AI