Built by one of the people behind the research that became ChatGPT
"At OpenAI, I helped build the methods that made language models useful at following instructions and talking with people. That work ended up as the research behind ChatGPT."— Diogo Almeida, founder, TypeSafe AI · launch essay · Sep 15, 2026
§2The numbers he put on it
Up to 200 times faster. Up to 400 times cheaper.
numbers: TypeSafe AI launch essay + docs§3Then somebody plugged it into Claude Code
A Claude session went from nearly one million tokens to 86,000
§4So what actually is this thing?
Why is everyone building with it inside a week?
Tweet · the reaction
“A genuine phenomenon”
Obie Fernandez wrote the book on putting AI into real software. He says ordinary programmers are learning to plug fast, cheap intelligence into normal code, and he expects every big lab to ship a Jev-style model within months. Shopify's founder replied “totally agreed”.
§5What Jev is not
It is not a chatbot. It does not write. At all.
THINKING IT?"If it can't write anything, what use is it to me?"
Most of what slows an agent down is not the writing. It's the pile of tiny decisions between the writing.
That pile is the part Jev takes over.
§6How you use it
You hand it the situation and a question. It picks — and tells you how sure it is.
§7System One
The fast part of your brain: you see a red light, you just stop
§8The problem
A big model, a full history, seconds of thinking — for a one-word answer
Seconds go by. Tokens get spent. And the actual answer was one word.
§9Three kinds of question
Choice, Score and Noul cover almost everything
real Jev run · my machine · replayed at reading pace
What you're watching: one real lead email goes in with three questions — a Choice, a Score and a Noul — and all three answers come back from a single request, with the real probabilities, latency and cost.
§101 · Choice
Pick one from a list — with a probability for every option
real answer from my research-agent test§112 · Score
Rate it against levels you define — and it can land between them
§123 · Noul
A yes or no — returned as the probability of yes
§13The part most people skip
It tells you how sure it is — so you can set a line
Tweet · a line in the wild
“75%+ → show me. Everything else gets ignored.”
Knowix built a content radar. Every hour it pulls posts and news, then Jev judges each topic against five questions. Jev writes nothing. Anything it's 75% sure about lands on a shortlist, and the rest is logged and ignored. That's the set-a-line idea running in real life.
§14Why that changes everything
The problem was never the AI being wrong. It's not knowing when it's wrong.
It's not the AI being wrong. It's not knowing when it's wrong.
§15Lots of questions at once
Five questions instead of one barely changes the speed or the cost
real Jev run · my machine · replayed at reading pace
What you're watching: one research situation, five tiny questions in a single request — all five answers come back together. Then my timing test: one question took 377 ms, five questions took 322 ms (median of six runs each).
§16What people built in the first few days
The results are specific, and they're all public
Tweet · the playful one
50 games at once, for less than a cent
Max Blade pointed Jev at Subway Surfers. It plays at superhuman speed and runs fifty games at the same time. His own caveat is the important part: Jev does not replace the big models, it opens up a new kind of job.
§17Build 1 · research sorting
1,018 research papers sorted for eight cents
§18Build 2 · email sorting
500 emails sorted in seconds for three and a half cents
real Jev run · my machine · replayed at reading pace
What you're watching: I ran the same idea on a 60-email test inbox I wrote. 2.53s wall clock, $0.0010 total. The small number on each email is Jev's confidence.
§19Build 3 · lead scoring
700 leads scored in 40 seconds — and the mismatches flagged
real Jev run · my machine · replayed at reading pace
What you're watching: my 40-lead test list with 8 deliberately mismatched messages hidden inside. Jev flagged 8 as “wrong person” — and all 8 were the planted ones. 1.52s, $0.0007.
§20Build 4 · browsers
Browser Use found flights in seven seconds, for under half a cent
honest note: it finds flights, it doesn't book themthe 7-second clock starts after the first page loadtheir code checks the result separately after Jev says “done”
Tweet · the step-by-step
“Give those decisions their own model”
Codila's long article walks the whole setup: find the part of your agent Jev can take over, try one decision in the Playground, then run Browser Use's agent — the one that found flights in 7 seconds for $0.0039. The clearest line in it: your writing agent keeps writing, and the small judgments between become a separate part you can inspect and price.
§21Build 5 · the one closest to what I run every day
586 pages re-linked in 45 seconds. Claude Opus 5 got through 21.
real Jev run · real agentos.guide pages
What you're watching: I ran it on this website. 257 real agentos.guide pages, 1,542 yes/no decisions, 5.8 seconds, $0.011. It placed 356 links and left 92 pages alone because nothing honestly fit.
THINKING IT?"Internal linking tools always force a link onto every page."
That's exactly what goes wrong. “Nothing fits here” is a scoring job, not a writing job.
On my own site Jev left 92 pages alone. A tool told to “add links” would have stuffed all 257.
“Nothing fits here” is a scoring job, not a writing job.
§22Why this one matters to me
Hundreds of small yes-or-no decisions, every single week
THINKING IT?"Doesn't running an agent operating system burn a fortune in tokens?"
No. The everyday work runs on free local models on your own machine, free APIs slot in for more, and the heavy work drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and the Agent OS plugs straight into it.
A decision model like Jev pushes that even lower, and the Boardroom has full token-efficiency tutorials.
§24Feature 1 · routing between models
“Choose the cheapest model that can finish this job”
real Jev run · my machine · replayed at reading pace
What you're watching: six real requests, one plain-English instruction. The timezone question and the rename go to the small model, and the strategy plan and the race-condition hunt go to the frontier model.
LangChain shipped this as an official Jev integration§25Feature 2 · safety checks on tool use
The guard rail that used to be locked inside someone else's product is now yours
real Jev run · my machine · replayed at reading pace
What you're watching: eight tool calls an agent might try while building a report. ls and git status pass, emailing all 4,000 members stops to ask me first, and rm -rf, DROP TABLE, a force-push and a piped install script get blocked.
§26The context one
Score every tool call in the history. Drop the ones that don't matter anymore.
real Jev run · a small test history I built
What you're watching: my own small version. A 14-step agent history, one Jev request scoring every step against the current goal. 132,000 tokens became 10,900 in 362 ms.
§27The honest pushback
Theo: cleaning up history is not the same as filtering it
§28Where I land
Right about the risk. And the direction is correct.
§29What Jev can't do · 1
It can't write anything. Jev only picks.
§30What Jev can't do · 2
Confident is not the same as correct
§31What Jev can't do · 3
It can be tricked
real Jev run · my machine · replayed at reading pace
What you're watching: I tested it. A clean refund complaint gets flagged for the owner with certainty. Paste a hostile note into the same email and “archive it” jumps from 0% to 25%. One hardening line in the question pulls it back to 97%.
§32A detail that's easy to miss
Jev doesn't see the name you give your question
real Jev run · my machine · replayed at reading pace
What you're watching: the same draft, containing a password and a client's contract value, asked three ways. A tidy field name with an empty question gets a shrug (0.41). A real question catches it (0.99). The same real question under a junk name gives the identical answer.
§33Same goes for what you feed it
The quality of the decision follows the quality of the situation
real Jev run · my machine · replayed at reading pace
What you're watching: same question, three situations. Tell it only “the researcher finished” and it says “write the draft” at 87% — a guess. List the sources and what's missing, and it says “keep researching” at 99%. Close the gaps and it says “write” again, this time for a reason.
§34Cost
You pay for what you send in. What comes back is free.
§35The number to actually watch
Cost per finished task, not cost per decision
§36What's changed · one
Writing and deciding used to be the same job. Now they're separate.
§37What's changed · two
Decisions used to queue up. Now a whole batch comes back together.
§38What's changed · three
Confidence became a number you can act on
§39Why this matters more to a business owner
The thing holding people back is trust, not capability
“I'm 62 percent sure — you should look at this one.”
§40The honest state of things
Days old. Marked experimental. Smart people are still arguing.
§41“This sounds like developer territory”
Look at what these jobs actually were
Wrong: “This is developer territory. It's not for a business like mine.”
Right: Every job on this page was an admin task — sorting, scoring, linking, handing out work.
Wrong: “I don't have time to set any of this up.”
Right: The tools got cheap and fast enough this week that “no time” stopped being the reason.
Wrong: “I'd need to be technical to use it.”
Right: Every result here came from someone describing a decision in plain English and giving the model a list of options. That's the whole skill.
Don't take my word for it
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries.
Get the AI agent system and learn how to get more customers and leads.
The Agent OS zip, the decision-layer walkthroughs and the 30-day roadmap live inside the AI Profit Boardroom, with four coaching calls a week to set it up with you.
