Jev 不是 LLM：只输出 choice/score/boolean 的决策模型
介绍 TypeSafe System One 的 Jev：用 RLCD 训练、只产出代码可分支的三种值，速度和成本比 LLM 高数百到数千倍，需要写字时才用 LLM。
Heard about @typesafeai ‘s new System One model: JEV ?! NOT an LLM… 👀 An LLM is trained and used to produce language. Jev is trained (RLCD) and used to produce values your code can branch on. Outputs are one of three types ton specific in advance: - choice from a closed list - score against a rubric - boolean (yes/no) Because it’s finely tuned and purpose built to specifically NOT be able to produce text responses, the evaluations it does are INCREDIBLY in terms of both speed and cost - think hundreds or thousands of times more efficient. Use an LLM when you need something written. Use Jev when the answers already exist and you just need a fast judgment. This will be a total paradigm shift as harnesses begin to drive token spend down this way. Thoughts? Checked it out yet??