Why OpenAI Could Clone Jev's Classifier Trick Overnight

TypeSafe’s Jev has become the AI world’s fastest-adopted model by reframing a large language model as a general-purpose classifier: rather than generating free text, it reads the probability distribution over the next token and reshapes it into structured answers — a true/false split for boolean questions, or the relative weights of candidate tokens for multiple-choice. The author argues this is essentially the logprobs-based classification trick people have experimented with for years, now productized and, crucially, calibrated through training.

The central claim is that OpenAI is uniquely positioned to copy it fast. OpenAI has quietly relied on single-token micro-classifiers since it introduced tool calling: at each step inside its ChatML format, the model is effectively deciding whether to call a tool, which tool to pick, and when a message is finished — each a tiny classification encoded in one token’s probability. The only real gap between those narrow specialists and Jev is generality, and since an LLM is already an enormously general next-token classifier, the author sees that gap as small. Folding a Jev-style capability into existing models could improve routing, guardrails, and efficiency in ways a standalone product cannot.

That leaves training data as TypeSafe’s likely moat. The author points to cofounder Diogo Almeida’s suggestion that the value lies not in raw data but in assembling vast, cross-domain examples with known outcomes — routed support tickets, hiring decisions, star ratings, moderation verdicts, resolved prediction markets — each paired with a question whose answer is already established, teaching the model to generalize calibrated judgments. Whatever reinforcement learning sits on top of that, the author suspects, is where any durable secret sauce would be.