Jev AI Just Killed the Most Wasteful Habit in Software (And Nobody’s Talking About It)

TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, emerged from stealth on September 15, 2026 with $40M in funding led by DCVC and released Jev, the first public model in its "System One" class, which skips autoregressive text generation and instead returns probabilistic decisions in a single parallel pass from application state plus structured, typed questions. Jev offers three question primitives (Choice, Score, Noul) and an RLCD calibration approach, and TypeSafe's self-reported benchmarks compare its speed, cost, and reliability against frontier models, though the company warns the evaluation is self-run and measures agreement rather than independently verified accuracy. The model is positioned for high-volume structured judgment calls such as triage, filtering, guardrailing, model routing, and agent control signals, while struggling with literal answering, arithmetic, world knowledge, and text generation.

Author s : Shobhit Agarwal Originally published on Towards AI. Why a model that can’t write a single sentence might be the smartest AI release of 2026 Jev is a new AI model from TypeSafe AI, a startup founded by Diogo Almeida a former OpenAI researcher who worked on RLHF/InstructGPT , which came out of stealth on September 15, 2026 with $40M in funding led by DCVC. It’s the first public release in TypeSafe’s “System One” model class, named after economist William Stanley Jevons. Image: Screenshot captured from https://typesafe.ai/The article explains that Jev differs from chat-style “frontier” LLMs by skipping autoregressive text generation entirely: it takes application state plus structured, typed questions and returns probabilistic decisions in a single parallel pass. It details Jev’s three question primitives Choice, Score, Noul , its calibration approach RLCD that makes confidence a real statistical signal, and how this design eliminates schema/output parsing issues while enabling fast, cheap fan-out of many decisions per request. It also compares Jev’s speed, cost, and reliability against frontier models using TypeSafe’s self-reported benchmarks—while warning that the evaluation is self-run and measures agreement rather than independently verified accuracy. Further sections cover practical integration typed API endpoint usage, SDK patterns, and decision workflows like fan-out, confidence-gated routing, cascades, and retrieval-then-judge , typical use cases triage/routing, filtering, guardrailing, model routing, agent control signals, and batch classification , where Jev struggles literal answering, no arithmetic/calculations, text-only dates, lack of world knowledge or generation, and non-adversarial default state handling , and why the broader takeaway is that products often need many small, structured judgment calls—not models that just talk more. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI