TL;DR
The short version
Jev is TypeSafe's System One judgment model: Choice, Score, or Noul against shared state — not a chatbot. Ten days in, people are using it on piles, queues, and agent micro-steps where volume and speed matter.
The fit test from Nathaniel Whittemore's AI Daily Brief: answers you can write down in advance, a stream of items, cheap wrong answers, and text evidence. Leave hiring, money, and security paths to systems that can explain.
Based on The AI Daily Brief episode How People Are Actually Using Jev (26 Sep 2026).
What it is (and isn't)
The AI Daily Brief skips the Tetris demos. Nathaniel Whittemore walks what builders shipped after TypeSafe's Jev left stealth: six recurring jobs, a fitness checklist, and places you should not trust a probability alone.
Jev does not draft emails, write code, or explain itself. You send state plus typed questions; it returns structured answers with probabilities. Three primitives cover almost every example: Choice (≤255 options), Score (2–10 named levels), and Noul (probability 0–1 that a statement is true).
A good way to think about a test for whether something is a good job for Jev is where you repeatedly read something, make a small judgment, then take a predictable next step.
Nathaniel Whittemore, The AI Daily Brief
Why the economics change the job
TypeSafe's public price is about $0.042 per million input tokens, with output free. Vendor latency claims sit around 70–500 ms. The launch pitch — roughly 20–200× faster and 40–400× cheaper than comparable LLM workflows — is a vendor claim on System One–shaped tasks. Independent remeasures show the multiple depends on the comparison: ~1.5–2× for a single classification against a fast small model; large multiples need a multi-step LLM workflow collapsed into one typed call.
When a judgment costs fractions of a cent and returns in under half a second, you stop asking whether it's worth an LLM call and start asking why you wouldn't ask it about every item. Parallel questions matter: Typesafe's own test, as relayed, found thirteen questions in one call ~12× cheaper and 10× faster than asking them sequentially.
TypeSafe models/pricing docs + launch post (vendor claims)
Six jobs showing up in the wild
Analyze the pile. Same questions over every ad, email, ticket, or document, then count. Builder runs scoring hundreds of live ads in tens of seconds for cents are the pattern — treat archetype "buyers" as hypothesis generators, not market research.
Search and filter by meaning. Plain-language classifiers over archives and feeds without hand-labeling a taxonomy first.
Triage what comes in. Importance on every email, invoice vs noise in Downloads, malicious vs safe links, lead routing. One Mac app example: Downloads watcher + custom rules, no other LLM in the loop.
Check work against rules. Turn "please review this" into yes/no questions on every draft or agent trace — online evals and style checks at volume.
Speed up agents. Micro-judgments that pick model, reasoning effort, skill, or whether old tool output is still relevant — reported experiments cut Astra spend ~50% or Claude Code skill tokens ~88%.
Respond instantly. Smart paste — resume → application fields, email/PDF facts → CRM — where the work is moving details between formats, not writing new prose.
The fitness test (and the red lines)
Whittemore's four criteria before you wire it in: (1) answers writable in advance, (2) a pile or stream, (3) wrong answers cheap or catchable, (4) evidence as text inside the context budget. Then: one judgment per question — "is this a good lead?" is usually five questions pretending to be one — and label a small gold set before production routing.
Typesafe's own caveats match: accuracy drops on multi-hop questions; counting, math, and dates need deterministic code; intent and consistency are weak spots. The copyable move is not "replace the LLM" — carve the repeated read→decide→route steps into typed questions, keep policy and math in code, and reserve generative models for proposal and explanation. LLM proposes, judgment model decides, code executes.
