TypeSafe AI, a startup backed by $40 million in funding, has launched Jev, a new type of AI model designed to communicate with software and other machines rather than with people. Unlike chatbots such as ChatGPT or Claude, which produce natural language text, Jev returns typed, probabilistic decisions that other programs can use directly without needing to parse or validate freeform text. The company positions this as a foundation for "machine-native AI", arguing that as automation spreads, most AI outputs will be consumed by other software rather than by human readers.

Jev works by taking a state input, such as a JSON object or a simple description like a customer complaint, and returning structured responses (categorised as Choice, Score or Noul) complete with confidence probabilities, useful for tasks like routing customer service queries or, as a demonstration, playing Doom. Built on a "System One" architecture using Reinforcement Learning for Calibrated Decisions, it processes outputs in parallel rather than sequentially like conventional large language models, giving response times of 70-500 milliseconds versus minutes for typical LLMs, and costing far less per token than rivals such as OpenAI's GPT-5.6 Terra. TypeSafe AI's CEO, Diogo Almeida, a former OpenAI researcher and co-inventor of RLHF, said the firm aims to build AI aimed at machines rather than solely at people.