Latent Space: Diogo Almeida on Jev, System One Models, and Reinforcement Learning for Calibrated Decisions
In this episode, TypeSafe AI CEO Diogo Almeida, a coauthor of the InstructGPT paper from OpenAI's early post-training team, explores why autoregressive chat models are misaligned with software integration. Almeida introduces Jev, which is built around Reinforcement Learning for Calibrated Decisions (RLCD) rather than human feedback or verifiable rubrics. He argues that standard RLHF promotes hallucinations, sycophancy, and mode collapse, whereas RLCD targets epistemically honest probabilities for System One tasks. The discussion covers why TypeSafe views itself as a data lab rather than a model lab, prioritizing tasks and data over massive compute. Almeida also examines key architectural topics including the KV cache, structured state over massive prompts, coding agents, computer use, and how AI should cleanly integrate into software control flow.
Original sources
Podcast Latent Space
Episode:Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI ↗
Newsletter Latent Space
Issue:Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.