How People Are Actually Using Jev

Context

Jev is a new AI model from Typesafe that got buzzy for game and visual demos, then attracted talk of a $10 billion-plus funding round. NLW looks past the demos at how people are actually using it a week and a half after launch. It matters to PMs because Jev is not a chatbot: it is a fast, cheap "judgment model" that can be wired into products and workflows to make small decisions at massive scale, and understanding where that fits changes what a PM should consider building or buying next.

The Big Idea

Jev is not a smaller LLM. It is a different primitive built only for fast, cheap, repeated judgment calls, and its real value shows up not from doing one judgment well but from doing thousands of them at a speed and cost that turns a task from "not worth doing" into "trivial to do everywhere."

Across every example in the episode, the pattern is the same: a task nobody would justify doing by hand, or even with a frontier LLM, becomes cheap enough to run on every item in a pile. That shift from "can't afford to check" to "check everything" is the actual unlock.

Key Insights

Jev is a System 1 model

Typesafe built Jev around Daniel Kahneman's System 1 concept: fast, instinctive pattern matching rather than deliberate reasoning. It cannot write code, draft contracts, or weigh many unquantifiable factors. It is built for a narrow test: you repeatedly read something, make a small judgment, then take a predictable next step, like classifying a downloaded file as an invoice and routing it.

Three question types cover everything

Jev answers exactly three kinds of questions:

• Choice: which of these fits, from up to 255 options (example: which team should handle this ticket, billing, support, sales, or other).
• Score: where something falls on a scale you describe in words, from two to ten levels (example: how frustrated is this customer, from calm to very angry).
• Nul (Bernoulli): a yes or no probability that a statement is true (example: is this person explicitly asking for a refund).

Every answer comes back with a confidence or probability, not just a label.

Speed and cost unlock volume

Typesafe claims Jev runs 20 to 200 times faster and 40 to 400 times cheaper than comparable LLM processes. A million input tokens cost 4.2 cents and output tokens are free, with each call taking 70 to 500 milliseconds. Questions about the same item run in parallel, so asking a tenth question costs more tokens but not more time. Typesafe found 13 questions handled in one call was 12.2 times cheaper and 10 times faster than asking them one at a time, with identical answers.

Jev crushes frontier LLMs on scale

Several head-to-head comparisons in the episode make the gap concrete:

• Borja had Jev read all 586 pages of a website to rebuild its internal link map in 45.1 seconds for 21 cents. Claude Opus 5 only got through 21 of those pages and spent $1.43 in the same task.
• Elvis Sananex had Jev scan 384 morning news stories to flag which of 15 brands should care, done in 24.9 seconds for 19 cents. Opus 5 got through 4 articles in that same time window, leaving 380 unread.
• Matthew Berman ran 12 questions across 724 live ads (hook, format, offer, CTA intent, and more) in 173 milliseconds for one request, then had Jev role-play 30 buyer personas reacting to those same 723 ads, producing 21,690 stop-scrolling decisions for 22 cents.

Cheap enough to check everything

When a check gets cheap enough to run on every single sentence instead of a document as a whole, NLW argues it stops being the same task done faster and becomes a different kind of task. The example given: checking every sentence of a document for AI-isms is qualitatively different from running one generic LLM pass over the whole thing, because the first one can actually catch what the second one would average away.

Four-part checklist for a good task

NLW distills a checklist from the examples in the episode:

• Answerable in advance: the response is a category, a yes/no, or a scale, not a sentence or a calculation.
• Volume: there is a pile or stream of items (hundreds of tickets, emails, documents), not a one-off.
• Low stakes or easy to catch: a wrong answer is cheap, or an error is easy to spot and fix, since Jev returns a number with no reasoning attached.
• Fits in context: the evidence needed can be given as text under 32,000 tokens.

Where Jev breaks down

Typesafe itself flags weak spots: accuracy drops with each hop on multi-step questions, Jev is weak at counting, math, and dates (it can extract the facts but shouldn't do the arithmetic), and it has consistency and intent-reading issues. NLW adds his own caution for hiring, money, and security decisions, since Jev returns a score with no attached reasoning, so a bad candidate ranking is invisible until someone checks it.

Mental Models & Frameworks

Six real use case categories

NLW groups the examples people are actually shipping into six buckets:

• Analyzing what you already have: run the same questions across an entire archive (ads, tweets, emails) to get data that was never worth compiling by hand before.
• Searching by meaning: describe what you want in plain words and let Jev check every candidate, finding matches even when the words don't match, like filtering Zillow listings by architectural style or clipping a video by theme.
• Triaging what comes in: decide what something is and where it goes, like rating email importance or classifying a downloaded file as an invoice.
• Checking work against rules: turn "does this meet the bar" into specific yes/no questions run against every draft, used for AI agent evals and catching planted writing mistakes.
• Speeding up AI agents: route a task to the right model, reasoning effort, or skill before or during an agent run.
• Responding instantly: fill in a form or field the moment something is pasted, by working out what each piece of text is and where it belongs.

One judgment per question

A tip repeated across the examples: a vague question like "is this a good lead" is not one judgment, it is several judgments stacked together (industry fit, company size, buying intent). Break it into its separate parts and ask each one on its own, since Jev is built to answer one narrow judgment reliably, not a bundle of them.

Practical Application

Run a pile-of-data audit

Pick one large archive you already have (a year of customer emails, CRM notes, meeting transcripts) and write 3 to 5 yes/no or scale questions you'd ask about every item in it. If you can't afford to answer those questions by hand across the whole pile today, that gap is the signal this is worth testing with a judgment model.

Build one recurring triage rule

Take one thing your team classifies by hand every day (which queue a ticket goes to, how urgent an email is, whether a lead is spam) and write it as a single yes/no or choice question. Jonathan Unikowski's live-prioritized inbox test rated 100 emails in 453 milliseconds for about a tenth of a cent and matched his own judgment on every one, which is the bar to test against.

Turn style rules into yes/no checks

Convert your team's writing rules or banned phrases into individual yes/no questions, then plan to run them against every paragraph of your next long document. The every.to team planted mistakes in 12 writing passages and found this kind of check caught most of them in a fraction of a second and at roughly 580 times lower cost than using a frontier model for the same review.

Test against your own judgment first

Before relying on any judgment-model classifier for something that matters, label 50 examples yourself first (50 emails, 50 leads, 50 tickets) and compare its answers to yours. Treat mismatches as the signal to rewrite the question, not as noise to ignore.

Keep high-stakes decisions human-reviewed

If you use a judgment model to rank job applicants or similar high-stakes items, do not let the number alone decide. Build in a second pass, for example having a separate LLM re-review borderline candidates, since the score comes back with no reasoning attached and a wrong ranking is easy to miss.

Bottom Line

Jev's value is not that it's smart, it's that it's fast and cheap enough to run the same simple judgment across everything, turning checks and classifications that were never worth doing by hand into something you can run on every ticket, email, or page. The practical move for a PM is to find one recurring judgment your team already makes by hand at volume, and test whether a model like this can make it at a speed and cost that changes what's possible.

Tools & Products

People to Follow

Matthew Berman

Ran the 724-live-ads analysis and the 30-buyer-persona scroll-stopping test cited in the episode as examples of Jev's analysis-at-scale use case.

Harrison Chase

From LangChain; pointed out that Jev is well suited to evals, especially online evals that need to grade large numbers of agent traces.

Notable Quotes

"Internal linking is the perfect Jev job. It's not writing. It is 8,790 yes or no calls." (Borja)

"Every email app shows your inbox in reverse chronological order. What if it was live prioritized by importance instead?" (Jonathan Unikowski)