Sep 21, 2026 · 2h 21m
TypeSafe bets machine-native AI can make software reliable
Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
The episode tests whether specialized decision models can turn AI from a conversational interface into dependable, economical infrastructure for software.
- 1Jev treats machine-readable decisions, not human conversation, as the core interface for useful AI.
- 2TypeSafe rejects benchmark optimization in favor of reliability, task-specific data, and intelligence per dollar or second.
- 3The company frames Jev as an early model category, not a finished answer, with broader machine-native systems still ahead.
Don't miss
Diogo explains how Jev’s choices, scores, and Newells translate machine-native intelligence into verifiable software primitives.
The brief
Diogo Almeida, TypeSafe AI’s CEO and a former OpenAI researcher, presents Jev as a programmable System One model built for software rather than conversation.
His critique targets RLHF, benchmark culture, and refusals: systems trained to produce expected text can lose behavioral diversity and make poor decisions.
TypeSafe’s alternative emphasizes task-specific data, structured decomposition, and RLCD, with intelligence measured by useful decisions per dollar or second.
Jev’s choices, scores, and Newells are designed as semantic primitives for switches, thresholds, sorting, and other verifiable software operations.
The conversation’s sharpest tension is between impressive demos and dependable automation, especially in computer use, coding agents, and enterprise dark data.