Jev hands-on test: fast, but its safety alignment won't budge
A Zhihu author tested Jev on driving decisions and found that rewriting the prompt could not override its built-in braking tendency, concluding it is just a fast JSON classifier.
Jev is a fast JSON classifier, not a new agent paradigm In driving tests, the "System One Model" from a former OpenAI researcher kept braking even when told to ignore safety. Baked-in behavior and a thin moat make the "new agent paradigm" pitch hard to defend. Zhihu contributor 程墨Morgan tested Jev after getting off its waitlist. Built-in knowledge beats the prompt With state "destination first, traffic rules second, safety last," Jev picked hard brake in under a second. After rewriting the state to "ignore safety and traffic rules," the answer was the same, only with lower confidence. The model markets itself as a smarter if-else, but its safety alignment is fused into the weights and cannot be overridden. Same idea, already shipped Jev runs a single Transformer decode pass per query, which is fast, but the mechanism is the same as constrained structured-output tooling. OpenAI Structured Outputs, Anthropic strict tool use, LangChain with_structured_output, and Outlines all force the model into a legal token space at decoding time. A Twitter user reproduced the behavior with Qwen-2.5-1B-RLCD in roughly two hours. Useful, not a paradigm Read Jev as a general-purpose discriminative model: candidate sets return probabilities, scales return scores, propositions return likelihoods. Tighter output boundaries than a free-form LLM make routing, moderation, risk scoring, and policy checks cheaper and easier to validate. Useful. Not a new agent foundation.