What Jev is: a plain-language look at the System One decision model
Using Kahneman's fast-and-slow thinking analogy, it explains Jev: you send program state and typed questions, and it returns constrained options with probabilities, suited to routing, scoring, checks and extraction.
Jev in simple words: Jev is a "System One" model from @typesafeai, derived from "System 1": Daniel Kahneman's concept of fast, intuitive thinking. A "System 2" is slow, deliberate reasoning. Jev is used for quick decisions a program needs to make countless times a day. Basically, you send two things: 1/ Program state: text, JSON, array describing what's happening 2/ Typed questions: a typed question is something where you tell it in advance what kind of answers are acceptable. The response of Jev is a set of typed answers with probabilities. For example, you might ask: "Should this ticket go to sales, support, or spam?" and Jev might say: "support, with 92% confidence". Jev is built for software automation and not human conversation. It works well when we know the answer space beforehand, like routing, classification, scoring, guardrail checks, etc. Consider the following few examples: 1/ Which team should handle this incident? Which tool should an AI agent use next? (routing) 2/ How urgent is this alert? Should we act now or escalate? (risk scoring) 3/ Does this model output pass our safety checks? (verification) 4/ Pull specific facts out of messy text and return them in a clean format. (extraction) An LLM generates text, one token at a time. If you want a structured output, you can only hope the model writes valid JSON; then parse and validate it yourself. It is error-prone and slow. Jev doesn't do text generation. It evaluates all the questions in parallel against the same state. This doesn't involve any JSON parsing, schema coercion, or cribbing about "model wrote a sentence, while I asked for a number" problems. The answers are limited to the options we declared, so it is easy to handle them in code. This also means Jev doesn't hallucinate. Internally, Jev is built on a "parallel sampling architecture". It means it processes the provided state and typed questions and produces every required answer in a single parallel pass. Usually, we give it three question primitives: 1/ Choice: pick one option from a set (e.g., "sales", "support", or "spam"). 2/ Score: rate something from an ordered scale (e.g., urgency from 1 to 5). 3/ Boolean: return a yes/no probability between 0 and 1. Since every question is evaluated against the same shared state, batching a dozen questions into one costs almost the same as asking just one question. Jev can't write text, hold a conversation, draft documents, or write code. It has no generative capability at all. It is only used for cases where the answer space is known in advance, and the output must be machine-parseable. This is its narrow scope, and TypeSafe AI argues that it is best suited for tasks that don't need creativity. If your software needs to make a fast, cheap, reliable judgment call, Jev is designed to do that one thing extremely well
This post quotes another post: x.com/jamesm/status/2101226027523412072