TypeSafe AI debuts model for machines that plays Doom

TypeSafe AI, a startup funded with $40 million, has unveiled Jev, a novel AI model designed for machine-machine interaction. Unlike traditional AI models that generate natural language responses to text prompts, Jev produces probabilistic decisions suitable for software or other AI models. The startup emphasizes "type safety" in programming, aimed at preventing errors that occur when software processes unexpected data types.

By delivering structured, typed values, Jev eliminates the need for text parsing and validation typically required for processing responses from large language models (LLMs). This approach proves beneficial for scenarios requiring limited AI responses, such as playing games like Doom. Jev operates using a System One model based on Reinforcement Learning for Calibrated Decisions (RLCD) architecture.

Co-founder and CEO Diogo Almeida, an OpenAI researcher, explains that TypeSafe aims to pioneer machine-native AI to facilitate AI's role in transforming work processes. Jev demonstrates superior speed compared to traditional LLMs, delivering responses between 70ms and 500ms, approximately 40x to 200x faster. The cost advantage of Jev over top-tier models like GPT-5.6 Terra is also noteworthy, with a significantly lower price per token.

TypeSafe claims Jev to be hallucination-free, although it's essential to note that its output consists of structured responses with probabilities rather than natural language. The model is particularly suited for AI automation, real-time applications, data classification, input verification, and AI model harnesses.