Open-Source Researcher Claims Frontier Lab Replicated His Non-Autoregressive AI Work

A developer published two arXiv papers in March and September 2025 detailing non-autoregressive decision models trained with reinforcement learning, releasing open weights, datasets, and a PyPI package. In September 2026, well-funded startup TypeSafe AI launched a product called Jev, which the researcher says applies the same core concept without open weights, technical papers, or public training data. The researcher argues that generative LLMs are overkill for simple classification tasks like routing or jailbreak detection, where faster, lighter models suffice. In response, he built an open-source alternative called RL Agent on a bidirectional encoder, achieving 33–38ms inference latency compared to Jev's reported 150ms. The project is fully open-source and aims to demonstrate that non-autoregressive decision models are a practical, accessible alternative to large generative systems.