How People Are Using Jev

September 25, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
September 25, 2026

MAIN STORY

How People Are Actually Using Jev

Jev is one of the buzziest models in a long time precisely because it isn't another LLM competing with GPT-6 or Opus 5.5. It's something fundamentally different and complementary, which is why it wasn't obvious at launch what it would be best used for. Ten days in, a clear set of real-world use cases has emerged beyond the viral game and visual demos, and the question worth answering is whether Jev belongs in the stack of people who aren't game designers or developers. The answer, increasingly, is yes.

What Jev is
A System 1 model built for snap judgments
TypeSafe calls Jev a System 1 model, after Daniel Kahneman's Thinking, Fast and Slow: fast, instinctive pattern matching rather than writing. It asks three kinds of questions. Choice picks the best fit from up to 255 options (which team handles this ticket?), score places something on a scale of two to ten levels described in words (how frustrated is this customer?), and Bernoulli returns a zero-to-one probability that a statement is true (is this person asking for a refund?). The simplest test for a Jev job: something repeatedly gets read, a small judgment gets made, and a predictable next step follows. The hype has already turned into money, with TypeSafe reportedly in talks to raise as much as $1 billion at a valuation of $10 billion or more, up from a $40 million seed at $200 million.

Why speed changes everything
Fast and cheap enough to judge everything, not just samples
Stefan's emoji demo shows the mechanics: type "things you can wear" and matching emojis race to the top; change it to "things you can wear in winter" and some fall back down. Every emoji gets the same yes-or-no question at the same moment, and nobody labeled or tagged anything, as traditional machine learning would require. Swap the emojis for leads, tickets or documents and the use case becomes obvious. Questions run in parallel, so a ninth or tenth question about the same item costs tokens but no extra time; TypeSafe found 13 questions in one call were 12.2x cheaper and 10x faster than asking one at a time, with identical answers. A million input tokens costs 4.2 cents, output is free, and calls take 70 to 500 milliseconds. The trade-off is that Jev won't write code, draft contracts or make nuanced multi-factor decisions. It's for small judgments at volume.

1. Analyzing what you already have
Analysis nobody could justify by hand, now in seconds for cents
Ask the same questions about every item in a collection, then count the answers. Matthew Berman broke down 724 live ads from 37 brands, asking 12 questions per ad (hook archetype, format, offer, CTA intent, awareness stage) at 173 milliseconds per ad: 40 seconds and nine cents for the lot. Two days later he had Jev scroll 723 ads as 30 buyer archetypes, from gym owner to toddler mom, generating 21,690 stop-or-scroll decisions for 22 cents. That's a hypothesis generator, not data from real buyers, but it's hard not to see marketing adopting this to pressure-test angles before paying for real tests. Ian Nuttall ran eight questions across nearly 3,300 past X posts for about 13 cents to compare against engagement. Any pile works: a year of customer emails, CRM notes, a folder of meeting transcripts.

2. Searching by meaning
Finding what you mean even when the words don't match
Justine Moore at a16z scanned thousands of Zillow listings and classified them by things no filter supports, like architectural style, renovation status and proximity to freeways. Burhan Usman clipped a 90-minute video by a described theme, such as predictions for when AI will automate AI research, in under two seconds for under two cents. Borja ran Jev over all 586 pages of a website to rebuild its internal link map in 45.1 seconds for 21 cents, while Opus 5 got through 21 pages and spent $1.43. His framing: internal linking isn't writing, it's 8,790 yes-or-no calls, and frontier prices have been paid to do it one page at a time. The same idea extends to filtering what you read, from Robin Bilgil's real-time AI slop detector for X to Elvis's run of 384 morning news stories against 15 brands in 24.9 seconds for 19 cents, in a window where Opus 5 read four.

3. Triaging what comes in
What is this, and where does it go?
Jonathan asked why every inbox is reverse chronological rather than live-prioritized by importance; Jev rated 100 emails in 453 milliseconds for about a tenth of a cent and matched his own ratings on every one. That's a use case worth wanting right now. Marcel Pociot built a macOS app that spots invoices in his Downloads folder and files them with the right name, with no other LLM calls. Dev Ed moderates livestream chat as messages arrive, and Steven Tey fed Jev 10,000 malicious domains caught on Dub's free link shortener, solving a problem they'd wrestled with since day one in two hours. As one marketer's list of use cases put it, most workflows eventually hit the same question: what should happen next? Box has experimented with incident triage for severity, customer impact and escalation, and Odoo CRM has a proposed module that scores every lead on priority, buying readiness and spam. For any company with multiple people touching the same leads or customers, this kind of triage and routing is where Jev becomes integral.

4. Checking work against your rules
Turn "please review this" into specific questions
LangChain used Jev to grade an agent's work the same way every time, with Harrison Chase calling it great for evals, especially online evals that grade lots of traces. As more companies put support bots and research agents into production, cheap, consistent evaluation could easily become core infrastructure. Every planted mistakes in 12 passages of writing: Jev caught six of seven against Fable 5.1's seven, but did it in 0.35 seconds versus 8.83, at roughly 580x lower cost. The practical move is translating a style guide or banned-phrase list into yes-or-no questions and running them on every paragraph. That points to the bigger rewiring Jev requires: checking every sentence for AI-isms is a difference in kind from running one generic LLM pass over a whole document.

5. Speeding up AI agents
Micro-judgments routing agents to the right intelligence, context and skills
Which model can handle this? How much reasoning does this step need? Which skill fits? Vechen had Jev change GPT-6's reasoning effort inside Codex mid-task, more when stuck and less for routine steps, cutting Astra costs by 50% with faster runs. Dani Avila built a Jev skill suggester for Claude Code that injects only the matching skill into context, for an 88% drop in tokens and cost, which matters for anyone who has built a big skills library. Expect judgment models built natively into harnesses too: AJ Asver's harness learns a job as it runs and moves steps from LLM calls to code, taking compliance alerts from about $2.95 each to 25 cents by alert 1,000.

6. Responding instantly
"What if copy/paste was smart?"
Marcus Lowe's version: copy a resume, paste it into an application, and the fields fill themselves. Nomandsign did something similar, splitting pasted text into pieces, working out what each field is, checking against the original and pasting only confident matches. So much of work is moving details from emails, PDFs and meeting notes into CRMs, intake forms and templates that this category feels immediately relevant.

Where to be careful
A judgment model shouldn't be left to judge alone
Hiring, money and security warrant caution. Ranking a flood of applications with Jev returns a number with no reasoning attached, which argues for multi-model systems where an LLM reviews borderline candidates. TypeSafe itself flags weak spots: accuracy drops on multi-step questions, it can extract facts but shouldn't do math or dates, and consistency and intent-reading have limits. Four criteria mark a good Jev task: answers can be written down in advance, there's a pile or stream of volume, a wrong answer is cheap or easy to catch, and the evidence fits as text under 32,000 tokens. Then write good questions: one judgment per question (not "is this a good lead?" but industry fit, company size and buying intent separately), describe every level of a scale in words, and test before trusting it in production, such as by hand-labeling 50 emails and comparing.