Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
Latent Space interviews TypeSafe CEO Diogo Almeida about models designed for code, calibrated decisions, workflow decomposition, and why TypeSafe optimizes Jev for intelligence per dollar rather than chat.
Original by Latent SpaceEvaluationPodcast transcriptOriginal published Source reviewed
Before you dive in
What you’ll find in the original
System One models are designed for code to consume directly, unlike pretrained autocomplete or chat-tuned models whose primary product is text.
RLCD aims to optimize probability-bearing decisions for software, contrasting its objective with human-preference RLHF and rubric-verifiable RLVR.
Developers should decompose workflows into small measurable decisions, keep state structured, and use different models for generation or deeper reasoning.
Worth knowing
The interview presents TypeSafe’s creator perspective and unpublished training claims; it is not an independent technical validation.