Jev introduces a new shape of LLM (simonwillison.net)

TypeSafe AI has launched Jev, a pioneering "System One model" designed for fast and cost-effective decision-making in artificial intelligence. Unlike traditional large language models (LLMs), which return text-based outputs, Jev processes text input and produces floating point numbers that represent probabilistic decisions such as yes/no questions, ratings, and confidence scores. It leverages an efficient pricing model—charging only for input at $0.042 per million tokens, making it cheaper than alternatives like OpenAI’s GPT-5 Nano. Jev can process a variety of input formats and allows users to ask parallel questions, making it particularly adept for tasks such as classification, spam detection, and ranking relevance in search results. However, the introduction of Jev also raises concerns about model transparency and bias. As a black box, it provides minimal insights into its decision-making process, which could potentially obscure issues such as algorithmic bias when applied in sensitive contexts like hiring. The community has already begun exploring creative use cases for Jev, demonstrating its flexibility in various applications, from chat models to game simulations. As interest in Jev grows, the emphasis on evaluation and experimentation will be crucial to understanding and mitigating its limitations in the evolving landscape of AI/ML.