OpenAI is about to eat Jev's lunch – Arcturus Labs

Jev repurposes the token-level probability outputs of a conventional large language model as a general-purpose classifier: at a single token position the model’s logits for “true”/“false” or for multiple-choice labels are read and renormalized into calibrated probabilities. That pattern builds on practices already visible in ChatML tool invocation, where single-token predictions act as micro-classifiers to decide whether to call a tool, which tool to call, and when a response ends. Jev’s rapid adoption demonstrates demand for a packaged, general classification primitive, and the core technical idea - using LLM next-token probabilities as classifiers - is straightforward enough that a well-resourced provider can fast-follow.

TypeSafe’s durable advantage is not architecture but its training and reinforcement-learning pipeline: large, synthetic, outcome-labeled datasets that teach calibration across domains (examples include routed support tickets, hiring outcomes, product ratings, moderation verdicts, and resolved prediction markets). With that calibration, a classifier built into a model lets the model “ask and answer” discrete questions inline without external tool handoffs, improving model selection, reasoning efficiency, security guardrails, and cost. The decisive test is accuracy and generality; current evidence shows failures in some domains, so long-term competitiveness depends on whether calibration scales across varied, high-stakes use cases.