Typesafe launches Jev, an ultra-fast decision-making AI

Diogo Almeida, former OpenAI employee and co-creator of ChatGPT, launches Typesafe and its model Jev. Not intended for human dialogue, Jev produces structured decisions for programs, with response times announced between 70 and 500 ms and a cost of $0.042 per million tokens. Access is available via a waitlist, and demonstrations, including a segment of Doom, are already circulating.

Latency, cost, and access announced before public launch

Typesafe states that Jev provides responses in 70 to 500 milliseconds. The costs are also highlighted: while large language models range from $0.20 to $10 per million tokens, Jev is offered at $0.042 for the same volume. This model is presented as less expensive than text-based AIs. For now, developers must sign up on a waitlist to access Jev. Several demonstrations are already available online, including one where Jev plays Doom, showcasing the speed of its decision-making.

Positioning against text generated by LLMs

Diogo Almeida, who has worked on large language models, considers these systems useful for following instructions or conversing with humans, but not suitable for automation. He describes them as often overly confident and inconsistent. Typesafe believes there is a mismatch when models produce text for judgments used by code, and reminds that LLMs can hallucinate. Jev, classified in the System One category, does not generate text but produces structured decisions that can be directly utilized by programs.

How Jev works and intended uses in pipelines

Jev provides structured outputs, such as a choice from a list, a confidence score, or a binary response, and allows for breaking down complex questions into sub-questions. The model processes unstructured data, such as text, and then emits values in a predefined format that is machine-readable. Among the cited applications are routing other AI models and data sorting. Typesafe indicates that it has developed a new architecture and new training methods, aiming to provide structured responses that cannot hallucinate. Jev is announced as responding in real-time, while competing approaches require several seconds, and its outputs include a confidence score. This model aims to meet the needs for reliability, latency, and cost in automation.