Jev primer: using Choice, Score and Noul to make software decisions
An introduction to TypeSafe AI's Jev, which turns a state into probabilities and scores through three question types so code can route tickets, rate risk, or decide whether an agent should continue.
Jev just dropped — and it’s not another chatbot ⚡🧠 You’ve never heard of Jev. That’s fine. It only launched this week. Jev is a new kind of AI from TypeSafe AI, a startup founded by Diogo Almeida — one of the OpenAI researchers behind the RLHF work that helped make ChatGPT possible. After two years in stealth (and a reported $40M seed), they shipped something that does not write essays, emails, or code. 🚫📝 It makes decisions. You give Jev two things: 1️⃣ a state — a support ticket, an email, a log, an agent trace, a JSON blob 2️⃣ a set of questions you define in advance It answers all of them at once and hands your software numbers it can actually use. 🎯 Three question types: 🔘 Choice — pick one option from a list you gave it 📊 Score — rate something on a rubric ✅ Noul — yes/no as a probability (0 to 1) Example: “Is this customer angry?” doesn’t come back as a paragraph. It comes back as 0.91 — 91% likely yes — plus a confidence score. Your code can then refund, escalate, or ignore. No parsing. No “please output valid JSON.” No invented fields. 🧩 They call this a System One model — Kahneman’s fast-judgment brain, not the slow “write me a strategy memo” brain of ChatGPT / Claude / Grok. Jev is meant to sit inside software: routing tickets, scoring risk, deciding whether an agent should continue, retry, or ask a human, checking another model’s work. 🛠️ Why people are shouting about it: ⚡ Speed — typically 70–500 milliseconds 💰 Price — about $0.042 per million input tokens. Output is free. Vendor numbers claim tens to hundreds of times cheaper than asking a frontier LLM the same classification questions 🧱 No free-text hallucinations — it can’t wander off-schema because it never generates text. It can still pick the wrong option inside the box you defined. That’s a different failure mode, not “zero errors.” Name check: Jev is named after economist William Stanley Jevons (Jevons paradox — when something gets cheaper, you use more of it). The bet is that cheap, reliable judgments get used everywhere, not just in chat windows. 📉➡️📈 What it is not: ❌ A replacement for Claude, GPT, or Grok when you need writing, planning, or open-ended reasoning ❌ Magic ❌ Fully proven on your data yet — English is strongest; treat the 200× / 400× headlines as vendor-best-case until you test Think of it as a frontier-intelligence function call: messy state in, typed probabilities out. The hype is that agents and production pipelines have been forcing chat models to do a job they were never designed for. Jev is built for that job. 🤖➡️⚙️ Official: http://typesafe.ai + http://docs.typesafe.ai Also showing up on Vercel AI Gateway and Cloudflare. Not another “GPT-but-cheaper.” A different tool. That’s why it’s loud this week. 🔥