Diogo Almeida
Founder of TypeSafe, the AI lab behind Jev and the System One Model class. Before TypeSafe he worked at OpenAI, where by his own account he helped build the methods that made language models useful at following instructions and conversing — work he describes as ending up as the research behind ChatGPT.
Position
The question he frames as driving four years of his work: models have been superhuman at chat for years, so where is all the automation? His answer, and TypeSafe’s founding premise, is that chat-shaped models are the wrong interface for software — strings have to be parsed and validated, and a model free to emit anything can always go off the rails. TypeSafe spent roughly two years in stealth building a stack around the opposite bet: a fixed output space, parallel sampling, and training aimed at calibrated confidence rather than human preference (Reinforcement Learning for Calibrated Decisions).
TypeSafe announced Jev on 2026-09-15 with a reported $40M round led by DCVC.
A note on attribution: third-party coverage (How to Use Jev - A Practical Guide) describes him as having co-invented RLHF and InstructGPT. His own wording in the launch post is more measured — that he helped build the methods behind instruction-following and conversation, and that the work ended up as the research behind ChatGPT. This note follows his own framing.
Relevance to Search
Indirect but real. Nothing he has published is about retrieval, but the model class he describes supplies the property that made Jev competitive as a reranker in Hev meets Jev — a Calibrated Relevance Probability on every judgment, which no purpose-built reranker returns.
Articles in This Vault
- Introducing System One Models & Jev (2026-09-15) — the launch announcement for the model class and for Jev, including the workflow-eval methodology and the company’s own list of caveats on its headline figures
Related Concepts
- System One Model · Reinforcement Learning for Calibrated Decisions · Calibrated Relevance Probability