A JSON classifier got 12M views: Niels Rogge visualizes how Jev skips autoregressive generation

rickasaurus · x · 2026-09-17

Hugging Face engineer NielsRogge created a visual explanation (made with Claude) of how Jev works, based on the Qwen2.5-RLCD model. Instead of autoregressive LLM generation, a single Transformer decoder pass processes the context plus JSON schema once, caching the keys and values. For each JSON schema field, the field suffix tokens go through the decoder again reusing the KV-cache; the final hidden state is passed through the LM head, logits are restricted to the tokens valid for that field, a softmax normalizes them, and the highest-probability token wins. The thread hit 12 million views, prompting the quip: "12 million views for a JSON classifier? Yeah, we're in a bubble."

Original post →

More from Fun

Fun channel →