OpenAI is well positioned to fast-follow Jev
Points and comments are a snapshot, not live.
OpenAI can likely replicate Jev's classification model using its existing architecture and data.
Jev, TypeSafe's LLM for instant calibrated decisions, was adopted faster than any model in Vercel's AI Gateway history. The article argues Jev uses conventional LLMs to generate single-token probability distributions for classification. OpenAI has used similar micro-classifiers for years, e.g., tool-calling tokens. The author posits OpenAI could fast-follow by replicating Jev's training, then fold classification into its models for self-questioning, reasoning checks, safety guardrails, or model routing. TypeSafe's moat may be its synthetic training data and reinforcement learning process. The author found domains where Jev's probabilities don't hold up, questioning its generality and accuracy.
What commenters are saying
Commenters are skeptical Jev's technical approach is novel or its accuracy claims hold up. Several note Jev's outputs change with choice order and that deterministic formatting is achievable with grammars on any LLM. One argues the article's explanation is likely wrong, suggesting an encoder model instead. There is a camp predicting OpenAI or Anthropic will simply acquire TypeSafe for its team and PR. Another observes that replicating calibrated probabilities requires specific training data, which could be a genuine moat. Some argue Jev is best suited for narrow tasks like sentiment analysis where cheaper models already excel.