Jev 与 LLM 的分工：决定 vs 生成

dylan ツ @demian_ai原推发布于 2026.09.23

在 X 上打开原推

Been trying to make sense of what Jev actually is. When to use Jev, when to use an LLM, and why? Why people keep calling it a paradigm shift, and whether that’s earned or just a new wrapper around classification. something really clicked for me: we built models that write, then we spent years forcing those models to decide Jev is the first mainstream model that just decides, it's that simple. An LLM’s job is “what should we say next.” Token after token. A paragraph, some JSON, a plan, a reply. Even when you only wanted a yes or no, it still has to write its way there. That’s why it feels powerful and also why it feels heavy. You’re pretty much asking a writer to be a switch. Jev’s job is “what is this.” You give it some text like a ticket, an email, a snapshot of app state, and a list of questions you already defined. It does not write back. It returns answers your code can use: - yes/no, as a probability - pick one from this list - where does this sit on this scale That’s the whole interface. State in. Typed judgments out. So if a customer says I upgraded yesterday and now I can’t access what I paid for, you don’t ask Jev to draft the support reply. You ask: Is this billing? Which team? How urgent? Does a human need to see it first? It reads the message once, answers all of those at once, and you’re looking at numbers a few hundred milliseconds later. Your code does the if statement. Jev never sends the email The paradigm shift, if there is one, is not smarter than GPT, it’s the split Writing models are for output a person will read, or for work that has to be invented: copy, code, summaries, plans, explanations. Decision models are for the moment inside software where the next line is just a branch. Route this. Block that. Score this. Escalate or don’t. Most production AI is secretly the second thing wearing the first thing’s clothes. That’s why everyone is prompting a chat model for JSON, parsing it, retrying when the schema breaks. Jev removes that translation layer. It can’t invent a department you didn’t list. It also can’t explain itself. Those two facts are the same design choice. Does it understand words? yes, in the same way an LLM does. It takes text. It has the language circuitry to hear tone, intent, everything’s fine but. Officially it’s built to understand natural language input, it just doesn’t talk. Literate, mute. If you need the essay, take Jev’s decision and hand it to a writer model. One judges, one narrates. Why it’s fast and cheap is almost boring once you see it. A chat model answers by writing. Each word waits for the last one. You pay for that writing. Latency grows with the answer. Jev doesn’t generate an answer in language. It fills in a scorecard in one pass. All the questions in parallel, no output tokens. That’s why people are quoting 70-500ms and input-only pricing. keep in mind it can still be wrong. Can’t hallucinate only means it can’t wander off the form. It can still tick the wrong box on the form for example. The probability is interesting tho: high: let the program act Middle: ask a person Low: you don’t know yet. the version I’ll actually remember: An LLM answers what should we say Jev answers what is this Use the writer when someone has to read the result. Use the judge when your code has to move. Most systems that work will use both. Jev routes the ticket, the LLM writes the reply. Jev can even check the reply after: rude, off policy, promising a refund you can’t make. That’s the shift, we finally stopped asking LLMs to be if statements @typesafeai @CompleteSkeptic