AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Diogo Almeida on TypeSafe's Jev, System One Models, and Moving Beyond Chat

· Latent Space

In this interview on Latent Space, TypeSafe co-founder and CEO Diogo Almeida explains the philosophy behind Jev and machine-native System One models engineered for programmatic reliability, code, and intelligence per dollar. Drawing on his experience helping build instruction-following models at OpenAI, Almeida analyzes the failure modes of RLHF—such as mode collapse, poor calibration, and problematic refusals when models are embedded as software dependencies—contrasting them with RLCD and RLVR. He outlines why TypeSafe operates as a data lab rather than a model lab, arguing that the right task and data matter more than simply scaling compute, and why he would decline to pre-train a model from scratch even with a billion dollars. Almeida also details how developers can replace massive prompts and system messages with structured state and small decomposed decisions, and shares his view on how Jev reshapes coding agents and multi-agent software architectures beyond the KV cache.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.