The author of Laya says TypeSafe's Jev is his March 2025 architecture relaunched as a breakthrough, and publishes latency numbers to back it
Nandakishor Mukkunnoth of ConvAI Innovations posted a priority claim that hit 1,231 points on Hacker News: he shipped non-autoregressive RL decision models in March 2025 (arXiv 2503.23303) with Apache 2.0 weights, and says TypeSafe AI's Jev is the same idea shipped in September 2026 behind a metered API with no weights or datasets. His numbers: Laya checkpoints at ModernBERT-large 421M and mmBERT-base 322M, 32.8ms on a single GPU against Jev's 236-276ms, 0.950 on AG News, 0.766 on typed-decisions, and calibration error 0.081 versus Jev's 0.246. This is the first substantive attack on Jev's novelty rather than another reimplementation of it, and the calibration gap is the number worth checking if you are routing with a discriminative model.
↳ Follow the thread