Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
TypeSafe CEO Diogo Almeida explains Jev, a System One model trained with RLCD for calibrated production decisions.
Latent Space interviews Diogo Almeida, TypeSafe AI CEO and an InstructGPT coauthor, about Jev, a model built for reliable software decisions rather than chat. He describes Reinforcement Learning for Calibrated Decisions, which optimizes epistemically honest probabilities on System One tasks instead of RLHF preference ratings or rubric-based RLVR. Almeida argues refusals, sycophancy, and weak calibration make chat-tuned frontier models a poor fit inside software dependencies. The discussion covers coding agents, computer use, and treating Jev as a programmable primitive while optimizing intelligence per dollar.
- Jev targets fast, calibrated decisions for software rather than open-ended chat.
- RLCD optimizes honest probabilities instead of RLHF or RLVR rewards.
- Almeida, an InstructGPT coauthor, says chat alignment hurts composable automation.
- TypeSafe rejects public benchmarks and prioritizes intelligence per dollar.
- Organizations
- TypeSafe AIOpenAI
- AI models
- JevInstructGPT
From the original article
14 min read at latent.space
Tickets for AIE NYC now open, and apply for the invite-only AIE CODE . Join us !
latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.