Arcturus Labs Argues OpenAI Can Fold Jev-Style Classification Into Its Models With a <prediction> Tag
Arcturus Labs / Hacker News (295pts, 209 comments)·low signal
Published September 21, the post argues that OpenAI has used single-token tool-call decisions as implicit classifiers since 2024. On that view, TypeSafe's Jev has a shallow moat unless its training data is hard to replicate. OpenAI could expose classification inside the model through special tags and keep the decision on-GPU mid-reasoning, which a standalone Jev API cannot do. It is analysis, not a leak, but it drew 295 points and 209 comments.