I Built Non-Autoregressive Decision Models a Year Ago. Then a Frontier Lab Called It a Breakthrough
Summary
The post describes a 421M-parameter non-autoregressive decision model built with RLCD to produce calibrated probabilities for decision tasks, avoiding text generation. It compares to a frontier lab's approach, claims significantly lower latency and open-source availability, and presents benchmarks across routing, moderation, and phishing detection.