A Clean-Room 'Jevlike' Scoring Model Hits 451 Stars in a Day by Guessing at TypeSafe's Unpublished Architecture
Created September 16, jevlike implements one-pass option scoring: every option becomes a query vector, queries attend over context tokens, dot products score option-context pairs, and a softmax returns probabilities, so all options are ranked in a single forward pass instead of decoded token by token. The author is explicit that TypeSafe has not published Jev's design and this is an independent starter with the same input and output shape, reporting 98% on synthetic menus, 26% on Wikispeedia next-click data with a frozen Qwen2.5-0.5B encoder, and roughly 100x faster than small decoders on eight-option tasks. The speed of the replication attempt is the story: a closed architecture announced two days earlier already has a public approximation with checkpoints.
↳ Follow the thread