TypeSafe AI has launched Jev, a new class of "System One" models designed for production environments. The company's CEO and co-author of the InstructGPT paper, Diogo Almeida, argues that current frontier models have prioritized alignment and refusals over reliability, leading to issues like hallucinations and sycophancy.
- Jev utilizes a novel training technique called Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for epistemically honest probabilities rather than human-rated feedback or programmatically verifiable outputs.
- The model aims to "disappear into the background" of software, functioning as reliably as regex in tasks like computer use, voice control, and coding agents.
- TypeSafe AI plans to release future models such as ReasoningJev, focusing on composable, programmable AI for automation rather than generic chat assistants.
This approach addresses the calibration problem inherent in previous training methods, enabling Jev to serve as a robust foundation for automated workflows and smart software applications.