Jev 入门实测：用分类决策替代 LLM 做模型路由
作者体验 Jev 后总结：它只做 yes/no、单选、打分这类分类任务，跳过文本生成，比普通 LLM 快 20-200 倍、便宜 40-400 倍，适合模型路由和欺诈检测等场景。
Tried catching up with AI this week and digged into Jev, a new model from an engineer who helped build ChatGPT and is now trying to help reduce AI costs by this model. Most LLMs today are "System 2" thinkers. What that means is they reason step by step and write out an answer word by word. Great for it's usecase but a total overkill when all you actually need is yes/no, pick-one, or a score out of 10 kind of a classification response. Enters Jev. It only does those classification tasks. And because it skips text generation entirely, it is 20-200x faster than a normal LLM, and costs 40-400x less. Also, you are billed for what you feed, output is always free. This isn't a replacement or competitor to LLMs we are used to. This is something which can be used along with them to save cost and time. For example, it can be used for Model Routing. Have Jev decide how hard a task is, then send it to a cheap model or an expensive one based on that. It can do classification at scale. Usecases like fraud detection, categorisation of leads can be done by Jev for a fraction of cost and time. If you're burning tokens on tasks that don't need real thinking, and is a classification kind of usecase, it is really worth a look to use Jev for it. Let me know what you think 👇