TypeSafe’s Jev, a large language model designed for classification tasks, has seen rapid adoption according to Vercel, which reported it was “adopted faster than any other model in AI Gateway history.” However, one analyst at Arcturus Labs argues OpenAI is well positioned to fast-follow and potentially replicate the technology.
The core mechanism underlying Jev’s approach, according to the analysis, involves using an LLM to generate probability distributions over possible tokens, then extracting relevant probabilities to answer classification questions. For boolean questions, the model examines probabilities for “true” and “false” tokens. For multiple-choice questions, it compares probabilities across provided options. For scoring tasks, similar patterns apply.
The analyst contends OpenAI has long used LLMs as implicit classifiers through features like tool calling, where the model predicts whether a tool should be invoked, which tool to use, and whether a response is complete—all through single-token probability decisions. “OpenAI has been using large language models implicitly as specialized classifiers since at least the introduction of tool calling,” the analysis states. This existing expertise suggests the company could adapt Jev’s general classification approach relatively quickly.
According to the analysis, OpenAI’s potential advantage extends beyond replication. The company could integrate classification capabilities into existing models and agents to improve “quick model selection, more efficient thinking, better security guardrails, and generally smarter, faster, and cheaper models.”
Regarding TypeSafe’s defensibility, the analyst identifies training data and training processes as the primary moat. The approach requires large datasets of known outcomes paired with classification questions—such as support tickets with their actual routing decisions or resumes paired with hiring outcomes. Additionally, reinforcement learning techniques applied during training could represent “secret sauce” protecting the company’s competitive position.
“Architecture-wise, I don’t think there’s much of a moat,” the analyst states, suggesting conventional LLMs are suitable for general classification work. However, the analyst expresses hope for TypeSafe’s success, acknowledging that execution and proprietary training methods could provide meaningful differentiation.
Key facts
- TypeSafe’s Jev was adopted faster than any other model in AI Gateway history, according to Vercel
- Jev uses LLM probability distributions over tokens to perform classification tasks
- OpenAI has used LLMs as implicit classifiers since at least the introduction of tool calling in early 2024
- The analyst identifies training data and training processes as TypeSafe’s primary competitive moat