OpenAI Alumnus Unveils Jev AI for Direct Decisions, 200x Faster
Former OpenAI researcher Diogo Almeida has officially introduced Jev, a novel artificial intelligence model developed by his newly emerged startup, TypeSafe AI, on September 15. Departing from the prevailing large language model paradigm that relies on token-by-token text generation, Jev is engineered to bypass natural language output entirely. Instead, it directly produces structured decisions, probabilities, and calibrated confidence scores, targeting high-throughput automation workflows. The announcement marks TypeSafe AI's formal market entry alongside a 40 million dollar seed funding round led by DCVC, which values the company at approximately 200 million dollars. Almeida, previously recognized for his foundational contributions to InstructGPT and reinforcement learning from human feedback at OpenAI, argues that current generative AI architectures are inefficient for backend software operations. Traditional models force programs to parse verbose textual responses to extract simple values, creating unnecessary latency. Jev addresses this by accepting unstructured business data and outputting predefined probabilistic parameters through a parallel sampling mechanism. To support this architecture, TypeSafe has developed a new training methodology called RLCD, or Reinforcement Learning for Calibrated Decisions. Unlike standard alignment techniques that optimize for human preference, RLCD trains models to ensure their stated confidence levels accurately reflect actual outcome accuracy, enabling software systems to automatically route low-confidence predictions for human review. The company reports that Jev can operate up to 200 times faster and 400 times cheaper than frontier generative models within specialized structured workflows. These performance gains stem from the model's ability to output multiple decision parameters simultaneously, eliminating auto-regressive generation overhead. TypeSafe also claims a zero percent error rate on schema validation tests, a metric derived from structural constraint enforcement rather than semantic accuracy. The startup acknowledges that current benchmarks were conducted on simplified, high-density tasks and may reflect optimized conditions. Independent verification of the calibration metrics and cross-domain generalization remains pending. Jev is positioned as a System One model designed for backend automation, including classification, data routing, risk scoring, and agent guardrails. By functioning as an intelligent decision layer within code pipelines, it aims to reduce reliance on expensive general-purpose LLMs for routine logical operations. The launch reflects a broader industry recognition that separating open-ended text generation from deterministic decision-making may improve software efficiency. As TypeSafe prepares to deploy Jev across enterprise automation, the model's long-term viability will depend on external validation and its capacity to maintain calibrated reliability across diverse, real-world data distributions.
