TypeSafe AI's Jev offers an alternative to LLMs that claims to be 193x faster and 445x cheaper — System One type model is bespoke for probabilistic decision-making

TypeSafe AI has launched its first System One model called Jev, which promises to be 193 times faster and 445 times cheaper than existing large language models (LLMs) such as GPT-6 Astra. Unlike conventional LLMs, Jev is specifically designed for statement evaluation and decision-making, making it particularly suitable for programming purposes.

Developed by ex-OpenAI engineer Diogo Almeida, Jev is geared towards producing structured answers and can process individual questions in parallel, rather than generating continuous text like LLMs.

Jev operates by receiving a state, which includes a given situation and its associated data, and then answering specific questions with a confidence factor. This design allows Jev to be easily integrated into code, where developers can use the confidence factor to make decisions based on the answers provided. For instance, a company could show Jev a list of a customer's credit card transactions and ask whether the customer is requesting a refund, with the confidence rating determining whether to proceed with the refund or request further clarification.

Jev's output and input are in predefined data format, typically plain JSON, which simplifies integration into code. Unlike LLMs, Jev does not require prior context or retain a global knowledge database. It relies solely on the provided state for its responses and does not hallucinate. However, Jev is not immune to misclassification, adversarial attacks, or answering literal wording instead of meaning.

As a result, TypeSafe AI expects that Jev's most common use cases will involve wrapping logic workflows around questions and integrating it into decision-making processes, particularly when a confidence factor is available.